Next Article in Journal
The Demand for Gastronomic Tourism—Characterization and Tourists’ Profiles
Previous Article in Journal
Application of an Integrated DEMATEL-ISM-BN and Gray Clustering Model to Budget Quota Consumption Analysis in High-Standard Farmland Projects
Previous Article in Special Issue
Innovative Solutions for Combating Climate Change: Advancing Sustainable Energy and Consumption Practices for a Greener Future
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Research on the Impact of Artificial Intelligence on Urban Green Energy Efficiency: An Empirical Test Based on Neural Network Models

1
School of Economics, Shandong Normal University, Jinan 250300, China
2
School of Business, University of Aberdeen, Aberdeen AB24 3FX, UK
3
School of Information Science and Engineering, Shandong Normal University, Jinan 250300, China
*
Author to whom correspondence should be addressed.
These authors contributed equally to this work.
Sustainability 2025, 17(16), 7205; https://doi.org/10.3390/su17167205
Submission received: 17 June 2025 / Revised: 21 July 2025 / Accepted: 5 August 2025 / Published: 8 August 2025
(This article belongs to the Special Issue Sustainable Energy Economics: The Path to a Renewable Future)

Abstract

In recent years, the rapid progress of artificial intelligence (AI) technologies has significantly influenced urban green energy efficiency. Leveraging panel data from 271 cities in China spanning the period of 2010–2022, this paper conducts an empirical analysis of the impact of AI on urban green energy efficiency from multiple perspectives, including green finance, industrial chain resilience, and the intensity of environmental regulation. The key findings are as follows: ① AI has a substantial positive effect on urban green energy efficiency, a conclusion that is consistently confirmed through multiple robustness tests; ② Heterogeneity analysis shows that the influence of AI varies markedly across different regions, city sizes, and whether cities are central, coastal, or transportation hubs, yet it maintains an overall positive correlation. However, its impact is relatively weaker in the northeastern region and in megacities; ③ Mechanism tests reveal that AI enhances urban green energy efficiency by improving green finance, strengthening industrial chain resilience, and intensifying environmental regulation; ④ Spatial spillover analysis indicates that AI exerts a positive spatial spillover effect on local urban green energy efficiency. Based on these findings, this paper offers targeted policy recommendations to enhance urban green energy efficiency and advance sustainable development.

1. Introduction

In recent years, artificial intelligence (AI) technologies have swept across the globe, profoundly influencing the trajectory of economic and social development [1]. As the world’s largest developing country and a technological powerhouse deeply integrated into the global scientific and technological landscape, China has witnessed rapid progress in the research and application of AI technologies. Data indicates that among the 45,000 new generative AI patents disclosed globally in 2024, China contributed 27,000, accounting for 61.5% and ranking first in the world; by the end of 2024, the number of generative AI users reached approximately 250 million, equivalent to one out of every 5.6 Chinese people utilizing AI tools. The application of AI technologies has brought about positive structural transformations. In terms of production efficiency, industrial robots, with their exceptional stability, precision, and ability to work uninterruptedly, have significantly enhanced the level of automation and quality control standards in production processes, thereby strengthening industrial competitiveness [2]. Regarding the transformation of employment structures, AI technologies have given rise to a multitude of emerging high-value-added occupations, such as robot maintenance engineers and AI algorithm developers, driving the shift in labor from physically intensive to knowledge- and skill-intensive positions [3]. From the perspective of income distribution, industrial upgrading and increased technological content have enabled enterprises to create more economic added value, offering more competitive salaries and benefits to employees, especially in emerging technological fields where talent scarcity drives wage increases and optimizes the income returns for highly skilled individuals. Therefore, fully leveraging AI technologies and harnessing their powerful dynamic influence to promote economic development holds significant research value. Given that industrial production is often characterized by large scale, rapid technological iteration, and a high demand for advanced technologies, and that it has sufficient capital to introduce a large number of high-tech products, the level of AI in a region can be measured by the penetration rate of industrial robots in that region. The penetration rate of industrial robots reflects the popularity and penetration capability of industrial robots in manufacturing, which not only helps analyze the impact of robot technology on the labor market but also serves as an important indicator for studying multiple aspects of the economic system, such as corporate innovation, improved production efficiency, and industrial upgrading.
Since the dawn of the twenty-first century, anthropogenic pressures have accumulated to the point where they now relentlessly erode Earth’s ecological fabric, triggering cascading environmental upheavals: planetary fever, encroaching seas, and an accelerating extinction ledger. Instrumental records indicate that the mean surface temperature for 2023 exceeded the 1850–1900 pre-industrial baseline by 1.45 °C—an all-time high largely ascribed to carbon emissions generated by human enterprise [4]. Over the past twenty-four months alone, roughly two million species have seen their habitats contract or vanish, imperiled by a triad of contaminated waterways, biological invasions, and rapid climatic shifts [5]. This planetary turbulence has propelled the notion of green economic trajectories into mainstream policy discourse worldwide. Energy remains the lifeblood of modern civilization, sustaining every facet of production and daily life [6]. In 2024, China’s aggregate energy appetite reached 5.96 billion tons of standard coal equivalent, up 4.2% year-on-year, while national apparent natural-gas demand climbed to 426.05 billion m3, an 8% surge. Such figures underscore a stubborn reality: both China and the global economy remain deeply tethered to high-throughput energy systems. Yet conventional energy carriers—coal, oil, and natural gas—extract an environmental levy measured in acidified rainforests, methane plumes, and record-breaking atmospheric CO2. Cities crystallize this dilemma. Occupying barely 2% of terrestrial land, they devour three-quarters of primary energy and exhale 70% of all energy-related carbon dioxide. The hyper-concentration of people and production within urban boundaries creates “island-like” amplification of both energy intensity and pollution loads, magnifying ecological stress. Simultaneously, geopolitical tremors have tightened energy markets [7]. Between 2020 and 2023, upstream oil-and-gas investment contracted by half, while renewable additions lagged the widening supply–demand gap, driving the global energy-supply elasticity index down to 0.12 in 2023 and eroding resilience to shocks. Russia’s pipeline deliveries to Europe, for instance, plunged from 155 billion m3 in 2021 to merely 60 billion m3 in 2023 after the Ukraine conflict, leaving numerous nations scrambling for alternatives. Against this backdrop, enhancing energy efficiency and accelerating the diffusion of low-carbon energy carriers is not merely desirable—it is imperative. Urban agglomerations, as pivotal engines of national GDP, concentrate risk and opportunity alike. They must navigate climate hazards, structural economic transitions, and resource bottlenecks simultaneously [8]. Consequently, interrogating how cities can elevate the efficacy of green-energy utilization is both an urgent ecological necessity and a prerequisite for long-term socio-economic viability. Across China, municipalities are already experimenting with greener energy paradigms. Beijing and Tianjin have carved out coal-free districts and re-engineered their energy mixes; in 2024 these measures trimmed coal burn by more than 3%. Chongqing and Fuzhou, among others, have rolled out incentives for equipment retrofits and green makeovers, urging firms to conduct energy audits, benchmark efficiency, and deploy cutting-edge carbon-abatement technologies. As China’s economy races ahead and citizens demand cleaner, more harmonious living environments, the quest for ever-higher green-energy productivity will remain a central policy frontier.
The development of AI not only enhances economic production efficiency but also impacts the development of urban green energy utilization levels. On one hand, leveraging its strengths in big data analysis and predictive algorithms, AI can significantly improve energy extraction efficiency. Data shows that AI technologies can increase oil and gas drilling efficiency by 25 percentage points, and by optimizing operating parameters through AI, the utilization rate of wind turbines can be improved by 15 percentage points [9]. On the other hand, through its global data analysis capabilities, AI can effectively aggregate and control distributed energy dynamically using smart grid dispatching and virtual power plant technologies, optimize power distribution, and reduce energy transmission and distribution losses. Moreover, on the energy consumption side for end-users, AI algorithms, through the integration of “data + algorithms + control”, can help users adjust the usage timing of high-energy-consuming equipment to achieve load leveling, effectively reducing energy consumption in production and daily life [10]. Therefore, the application of AI technologies can not only improve social production technology levels but also, by fully leveraging its computational power advantages, contribute to the enhancement of urban green energy efficiency.
Yet, scholarly inquiry into the nexus between artificial intelligence (AI) and urban green-energy efficiency remains strikingly sparse. On one hand, a handful of contributions acknowledge that AI exerts some influence on the green-energy performance of cities [11], but they stop short of deploying rigorous empirical designs capable of gauging the magnitude or reliability of that influence. On the other hand, both Ren [12] and Tian et al. [13] endorse—at the conceptual level—a positive role for AI in advancing urban green-energy efficiency, yet they leave the precise channels and quantitative contours of that contribution largely uncharted. Moreover, cities diverge along axes such as geographic position and transport connectivity, heterogeneities that may condition the efficacy with which AI raises green-energy productivity. Consequently, a granular dissection of how AI technologies empower urban green-energy efficiency is not only theoretically consequential but also operationally vital for crafting green and sustainable urban policies. Motivated by these lacunae, this study sets out to unpack the mechanistic pathways through which AI affects urban green-energy efficiency. In light of data representativeness and the paper’s analytical objectives, the empirical focus is narrowed to industrial robots—the tangible embodiment of AI in the manufacturing sector—and leverages corresponding statistics. Drawing upon a balanced panel of 271 Chinese cities spanning 2010–2022, we subject the hypothesized effects to rigorous econometric scrutiny. Beyond the baseline relationship, the analysis probes three macro-level conduits—green finance, supply-chain resilience, and the stringency of environmental regulation—through which AI may shape green-energy outcomes. Attention is further devoted to spatial spillovers and to heterogeneous impacts across cities that differ in location and developmental stage. By redressing the current empirical deficit, the study aspires to furnish actionable guidance for cities intent on elevating energy productivity and accelerating the green transition, thereby contributing to the global pursuit of sustainable urban development.

2. Literature Review

Currently, scholars both domestically and internationally have conducted extensive research on artificial intelligence (AI) and urban green energy efficiency, focusing primarily on the following three dimensions:

2.1. The Definition, Social Impact, and Mechanisms of AI Technology

The concept of AI technology originates from the interdisciplinary fields of computer science, mathematics, and neuroscience, aiming to simulate functions related to human intelligence. It mainly encompasses core areas such as machine learning, natural language processing, computer vision, and robotics [14]. Regarding the social impact of AI, existing studies have highlighted its influence on economic growth and industrial upgrading. AI applications are believed to enhance production efficiency and quality, innovate products and services, and give rise to new industries and business models [15]. These advancements drive the transformation and upgrading of traditional industries, create new employment opportunities and economic growth points, and promote sustainable economic development. Additionally, research has focused on the role of AI in optimizing social services. AI can improve the quality of social services such as education and healthcare through specific applications like precision medicine and digital personalized education, thereby enhancing the accessibility and equity of social services. Moreover, Majrashi et al. pointed out that AI can create new job positions, such as data entry clerks and customer service representatives, leading to changes in the structure of labor market demand [16]. In terms of the mechanisms of AI’s impact, Hui et al. argued that AI can enhance production efficiency by increasing the level of automation and intelligence in production processes, thereby improving labor productivity, reducing costs, optimizing production procedures, and achieving optimal resource allocation [17]. Mariani’s research indicated that AI also has an innovation effect, suggesting that the large-scale application of AI can significantly stimulate innovative thinking and production models at the societal level. By developing new applications through intelligent algorithms, AI injects new momentum into the development of various fields, driving technological, business model, and social system innovations [18]. Furthermore, studies have also focused on the role of AI in education enhancement. In the context of widespread AI application, there is an urgent need to cultivate new types of talent that are more adaptable to the AI era [19]. This has led the education sector to adjust its talent cultivation goals and models, emphasizing the development of students’ innovative thinking, practical abilities, adaptability to change, and interdisciplinary knowledge integration skills, thereby improving the efficiency of talent cultivation in society.

2.2. The Definition, Measurement, and Optimization Paths of Urban Green Energy Efficiency

Zhou’s research defined green energy efficiency as the ability of an economy to maximize output while considering both the energy costs consumed in production activities and the environmental damage costs. This means achieving maximum expected output and minimum undesired output with a given energy input, or minimizing energy input and undesired output at a given expected output level [20]. Regarding the measurement of urban green energy efficiency, existing studies have mainly approached from two perspectives: single-dimensional efficiency and dynamic total factor efficiency. Lin et al. conducted a static assessment by comparing the relative efficiency of decision-making units, which achieved a single-dimensional energy efficiency calculation but failed to account for undesired output factors [21]. Building on this, Long introduced directional distance functions, such as the SBM (Slacks-Based Measure) and ML (Metafrontier) indices, to address the shortcomings of the previous models [22]. Wang et al. attempted to integrate geographic information technology with multi-source data, combining the SBM model to analyze the spatiotemporal evolution of carbon emission efficiency in depth [23]. Overall, the SBM model-based approach to measuring green energy efficiency is more scientific and comprehensive, providing a more refined method for assessing urban green energy efficiency in this study. Research on the optimization paths of urban green energy efficiency has focused on the impacts of technological innovation, industrial structure optimization, and legal regulation. Specifically, Liu’s study indicated that increasing research and utilization of renewable energy technologies can effectively improve the conversion and utilization efficiency of renewable energy, thereby reducing production costs [24]. Tan also argued that promoting the integration of green energy industries with other industries to form related industrial clusters and value chains can effectively facilitate the coordinated development of green energy industries with industries such as manufacturing, construction, and transportation. This enhances the market demand and application efficiency of green energy [25]. Du’s research further highlighted the role of policy and regulation, suggesting that reasonable legal constraints can significantly improve urban energy efficiency, reduce pollution emissions from energy use, and promote green and low-carbon urban development [26]. Additionally, strands of research are striving to push the frontiers of contemporary energy-storage technologies, seeking to wring higher utilization efficiencies from every kilowatt-hour generated [27]. In summary, research on urban green energy efficiency has preliminarily outlined the basic framework for exploring urban green energy utilization models.

2.3. The Impact and Mechanisms of AI Application on Urban Green Energy Efficiency

When considering the impact effect, researchers have primarily concentrated on how AI applications contribute to enhancing urban green energy efficiency. Wang et al., utilizing panel data models, demonstrated that the deployment and use of AI noticeably bolster urban energy efficiency [28]. Their findings indicated that AI applications can markedly improve a city’s capacity to enhance energy efficiency and elevate the share of renewable energy in consumption. These findings offer significant theoretical backing for grasping the role AI plays in bolstering urban green energy efficiency.
Exploring the mechanisms by which AI influences urban green energy efficiency, Zhou et al. scrutinized AI’s role in urban financial decision-making. They discovered that AI, characterized by big data analysis and intelligent algorithms, could refine investment decisions and mitigate the chances of improper capital allocation [28]. Consequently, this leads to a substantial enhancement in green finance levels and fosters the progression of green energy efficiency. Dong et al. further exposed that AI applications can effectively advance the intelligent management and upkeep of green energy machinery. For instance, AI technologies can identify issues in photovoltaic and wind power installations through smart monitoring systems, facilitating proactive maintenance, curbing equipment downtime, and ensuring efficient power production [29].
Additionally, scholars have turned their attention to how policy support and collaborative governance under AI’s purview impact urban energy efficiency. They suggested that governments could apply AI to dissect and design pertinent policies, like offering tax breaks and financial assistance to companies utilizing green energy [30]. This strategy stimulates businesses to boost their green energy consumption, which in turn aids in enhancing urban green energy efficiency. This assertion was further validated by Zhao et al. in subsequent research [31], providing crucial theoretical insights into the interplay between green infrastructure investment and sustainable economic growth.
In summary, existing research has achieved certain results in the application of urban AI and the improvement of urban green energy efficiency, but there are still the following shortcomings: (1) Insufficient data foundation research—existing studies are mainly concentrated in a few cities or regions, lacking a comprehensive analysis of different types of cities nationwide. They also fail to examine the differences in the relationship between AI application and urban green energy efficiency among cities with different conditions. (2) Insufficient research on mechanisms—existing studies provide limited analysis of the mechanisms of AI application and lack a systematic examination of how AI enhances urban green energy efficiency through pathways such as industrial chain resilience, green finance financing, and policy coordination. (3) Existing research has not yet paid attention to the spatial spillover effects of urban AI on urban green energy efficiency and lacks relevant analysis in this regard. Based on this, this paper conducts a series of empirical analyses using data from 271 cities in China over 13 years to scientifically analyze the impact of urban AI application on green energy efficiency, aiming to provide references for urban economic development in China.
Compared with existing research, the innovations of this paper are as follows:
First, a leap in data granularity. Confronted with the skeletal evidence base of prior work, we assemble—for the first time—a balanced panel covering 271 Chinese prefecture-level cities from 2010 to 2022. By fusing city-level penetration of industrial robots—precisely gauged via a Bartik instrumental-variable design—with energy consumption inferred from nocturnal luminosity, we plug the “sample fragmentation” hole that has long plagued the literature.
Second, an excavation of chain-mediating and synergy mechanisms. Departing from the shallow causal sketches common in earlier studies, we triangulate three conduits—urban supply-chain resilience, green finance, and environmental-regulation intensity—and deploy a chain-mediation architecture to trace how these levers interlock and amplify one another. This approach transcends the single-pathway lens endemic to previous analyses and uncovers policy-multiplier effects.
Third, a cross-jurisdictional quantification of spatial spillovers. Responding to the scant attention paid to the geographic diffusion of AI technologies, we craft an economic–geographic nested weight matrix and apply a spatial Durbin model to corroborate positive spillover impacts on green-energy efficiency. The exercise illuminates inter-city technology propagation and policy resonance, shattering the inward-looking perspective that has dominated the field.
Fourth, a stereoscopic yet micro-level dissection of heterogeneity. Moving beyond the coarse regional dummies traditionally used, we stratify cities along multiple axes—region, size, and location—then drill down to the individual metropolis. A neural-network routine extracts marginal-effect slopes for all 271 cities, spotlighting high-impact cases (Jinan, Shenzhen) and low-impact ones (Dongguan, Longnan), thereby furnishing micro-evidence for targeted policymaking.
Fifth, a methodological synthesis. By wedding neural-network architectures with conventional econometrics and rendering the resulting marginal effects through intuitive visual analytics, we redress the explanatory frailty of classical models when confronted with intricate causal fabrics.

3. Mechanism Analysis and Research Hypothesis

3.1. Industrial Structure Upgrading Effect

Firstly, AI applications achieve precise energy supply and demand forecasts and intelligent scheduling through accurate data analysis and intelligent dispatching. This is performed by extensively collecting and analyzing power generation data from urban energy devices such as solar panels and wind turbines, as well as energy consumption data from the urban energy usage side [32], thereby effectively enhancing the utilization efficiency of urban green energy. For instance, AI can predict changes in the power generation of solar and wind energy by analyzing historical weather data, sunlight duration, and wind speed and direction. This allows for the rational arrangement of energy storage and distribution strategies, improving the overall operational efficiency of the energy system. Secondly, electric power is the backbone of current industry. In the context of smart grid construction, AI-enabled smart grids utilize AI technologies to achieve real-time monitoring, fault diagnosis, and automatic repair of the power grid. This ensures the stable operation of the grid, facilitating the smooth integration and effective consumption of green energy, and further optimizing the integration and distribution of urban green energy. Thirdly, AI-driven smart home and smart building systems can automatically adjust the operating status of energy-consuming devices based on users’ energy usage habits and needs, achieving refined energy management and energy-saving optimization [33], thereby further improving the energy utilization efficiency of cities in the terminal application phase. Therefore, with the continuous development and application of AI technologies, their role in enhancing urban green energy utilization efficiency will become increasingly significant. Based on this, the following hypothesis is proposed for this study:
Hypothesis 1 (H1). 
The application of artificial intelligence significantly promotes the improvement of urban green energy utilization efficiency.

3.2. Industrial Agglomeration Effect

The main factors influencing urban green energy efficiency include the level of green finance, industrial chain resilience, and policy regulation. Investment in urban green infrastructure enhances the city’s green energy utilization capability and efficiency by increasing the city’s green finance financing capacity, industrial chain resilience, and intensity of environmental regulation [34]. This investment promotes the transformation of industrial economic development towards a high-speed, coordinated, and sustainable model, thereby facilitating the city’s green and low-carbon transition and improving its sustainable development capacity.
① The Effect of Enhanced Green Finance Level: First, artificial intelligence can rapidly process and analyze large amounts of data to accurately identify green projects and enterprises, addressing the issue of information asymmetry [35]. This reduces the cost and time for financial institutions to search for and verify green projects, thereby improving the efficiency of green financing project applications. Second, AI can assist financial credit institutions in building more accurate and dynamic risk assessment models, comprehensively evaluating the risks of green projects, and promptly adjusting risk response strategies. This helps financial institutions enhance their risk tolerance [36], encouraging credit institutions to actively participate in green finance financing. Third, AI-based fintech platforms exhibit network externalities [37]. As the number of users increases, marginal costs decrease, enabling green financial institutions to serve more clients at a lower cost. This expands the coverage of green finance, enhancing the accessibility and convenience of financing. Fourth, the application of AI can drive innovation in green financial products and services. For example, robo-advisors can provide personalized green investment portfolio recommendations based on investors’ risk preferences and environmental goals. Additionally, AI’s big data disclosure capability can promote transparency in green financial project information [38], enhancing investor confidence and attracting more funds to the green sector.
The level of green finance is an essential component for cities to improve their green economic development level. Its impact on urban green energy utilization efficiency mainly includes providing necessary financial support, promoting technological innovation, and guiding resource allocation. First, an enhanced green finance financing capacity can raise more funds for the investment and construction of urban green energy projects such as solar power stations, wind farms, and smart grids, meeting their financial needs for construction and operation [39]. This increases the supply capacity and utilization scale of urban green energy. Second, sufficient funding enables green energy enterprises to conduct research and development activities, introduce advanced technologies and equipment, and promote innovation and upgrading of green energy technologies. This improves energy conversion efficiency, reduces costs, enhances the competitiveness of green energy in the urban energy market, and increases the stability and efficiency of urban energy utilization [40], thereby significantly improving green energy utilization efficiency. Third, through market mechanisms, green finance can guide funds towards green energy projects and enterprises with high energy utilization efficiency and environmental benefits [41]. This encourages urban resources to tilt towards the green energy sector, achieving optimized resource allocation and improving the overall utilization efficiency of urban green energy. This is of great importance for enhancing the sustainable development capacity of cities.
② The Effect of Enhanced Industrial Chain Resilience: First, the application of AI can establish intelligent analysis models to deeply mine and analyze data, effectively assisting in the integration of production, sales, and inventory data from upstream and downstream enterprises in the industrial chain [42]. This enables real-time monitoring of the industrial chain’s operation, identifies potential risk points in advance, and reduces the impact of external shocks on the industrial chain, thereby enhancing its risk resistance. Second, AI technology can break down information barriers between different links in the industrial chain, enabling rapid and accurate information transfer and sharing. This helps enterprises arrange production plans and logistics delivery based on their production capacity and needs, improving the overall coordination efficiency of the industrial chain. Third, the widespread application of AI across industries increases enterprises’ demand for specialized talent [43]. To meet their development needs, enterprises will invest more in training and recruiting talent in the AI field. This attracts more universities and research institutions to gather in the city, forming a favorable environment for talent cultivation and innovation. The aggregation of talent not only provides technical and intellectual support for enterprises but also promotes the exchange and dissemination of knowledge and technology within the industrial chain [44], enhancing the innovation and risk resistance capabilities of the entire industrial chain and further strengthening its resilience.
As a crucial element of the urban industrial chain, the stability of the energy sector is augmented alongside the general enhancement of the urban industrial chain’s resilience [45]. This bolsters the capacity of energy companies to sustain steady energy production and supply in the face of natural disasters and market volatility. A robust industrial base and heightened risk mitigation capabilities can diminish energy waste and inefficiencies stemming from disruptions or instabilities in green energy projects [46], thus broadening the application of green energy. Moreover, bolstering the industrial chain’s resilience fosters the diversification of urban energy markets and boosts the demand for green energy [47]. As the industrial chain evolves and stabilizes, more businesses and consumers, appreciating the benefits of green energy, will ramp up their adoption and seek more green energy options. This surges societal investment in green energy infrastructure, motivates companies to manufacture more green energy products and services, and extends the application of green energy across cities, ultimately boosting urban green energy utilization effectiveness.
③ Regarding the impact of intensified environmental regulation: First, AI facilitates the integration and analysis of vast environmental datasets sourced from sensors and satellite imagery to monitor environmental metrics such as air and water quality precisely. This empowers regulatory bodies to swiftly pinpoint pollution origins and environmental concerns, offering concrete data for tailored environmental regulations across various regions. Second, AI’s profound analysis of environmental data enables the prediction of environmental trends and risks [48], which lays a scientific groundwork for developing environmental policies and aiding governments in selecting optimal regulatory strategies. Third, AI applications enhance environmental law enforcement by enabling intelligent oversight and breach detection. By monitoring and analyzing corporate production and emissions data in real-time, authorities can swiftly identify unlawful emissions and other misconduct [49], thereby increasing enforcement efficacy and precision, and invigorating environmental regulation stringency.
The intensification of environmental regulation contributes to urban green energy efficiency advancements in three principal ways. First, heightened regulation is typically paired with stricter environmental policies and laws [50], which limit the use of traditional energy sources that are high in pollution and energy consumption, incentivizing businesses to transition to cleaner energy alternatives. Second, stringent environmental rules can increase companies’ inclination to adopt green energy [51], compelling them to increase R&D investments in green and energy-saving technologies, fostering innovation and application, and enhancing green energy utilization efficiency. Third, higher environmental regulation standards can prompt governmental support for green energy adopters through tax benefits and subsidies, encouraging a shift from traditional to cleaner energy sources. Additionally, by instituting mechanisms like unified carbon markets, governments can significantly raise the cost of traditional energy use [52], making green energy more market competitive. This draws more businesses into the green energy sector, promoting the extensive use of green energy [53] and collectively augmenting urban green energy efficiency.
Based on the above analysis, this paper proposes the following hypothesis:
Hypothesis 2 (H2). 
The application of artificial intelligence promotes the improvement of urban green energy utilization efficiency through three mechanisms: the effect of enhanced green finance level, the effect of enhanced industrial chain resilience, and the effect of enhanced environmental regulation intensity.

3.3. Spatial Spillover Effects

This research delves deeper into how AI applications influence urban green energy efficiency through spatial spillover effects. Initially, leveraging agglomeration and selection effects [54], AI application encourages the tertiary sector to concentrate in economically prosperous and technologically progressive core cities, while also promoting the secondary sector to gradually relocate to adjacent areas. This process establishes a specialized labor division that speeds up the entire region’s shift towards more efficient and environmentally friendly industries [55], effectively restricting the growth of energy-intensive and heavily polluting enterprises while nurturing opportunities for green industry development. Furthermore, AI’s advancement boosts inter-city data sharing and connectivity. It enhances spatial efficiency in diffusing benefits to urban green energy utilization by leveraging integrated infrastructure and resource factor sharing across different sectors, significantly improving the spread of positive impacts [56]. Additionally, studies have verified that environmental regulations exhibit positive spatial spillover effects [57]. AI applications, through coordinated policies and standardized practices, create a cross-sector environmental governance network. Cities equipped with advanced AI can stimulate development in less advanced regions via technological and institutional spillovers, thereby spreading economic benefits across various domains, including urban green energy efficiency.
Hypothesis 3 (H3). 
AI’s deployment positively impacts urban green energy utilization efficiency through spatial spillover effects.
Overall, AI’s application not only directly enhances urban green energy efficiency but also supports green and low-carbon urban growth by increasing the city’s capacity for green financing, strengthening industrial chain resilience, and intensifying environmental regulations. This study also suggests that AI’s application positively influences urban green energy efficiency through spatial spillover mechanisms, indicating that benefits from AI implementation in one city can extend to neighboring cities. Figure 1 outlines the research framework of this study.

4. Research Design

4.1. Variable Explanation

(1) Independent Variable: The core explanatory variable, artificial intelligence (Robot), is proxied by the penetration intensity of industrial robots, following the recent calibration proposed by Liu Qing (2024) et al. [58]. To quantify city-level exposure to robotic adoption, this study draws on the strategy further refined by Wang Yongqin & Dong Wen (2023) [59]: 2004 is fixed as the benchmark year, and the Bartik instrumental-variable technique is employed to craft the robotic-penetration index for each city. Tier-one data are harvested from the International Federation of Robotics (IFR) World Robotics annual series (2010–2022), which reports the stock and flow of robots across 21 two-digit manufacturing industries. Tier-two mapping assigns these province-level IFR figures to 271 prefecture-level cities by anchoring the 2004 employment structure recorded in the China Economic Census Yearbook as baseline weights. Through the Bartik instrument, this yields a city-specific robot-penetration rate that is plausibly exogenous. A step-by-step exposition of the computation is relegated to Appendix A.
(2) Dependent Variable: The dependent variable in this paper is urban green economic efficiency. Drawing on Shi Dan’s framework [60], this paper gauges city-level total-factor green-energy efficiency by integrating an SBM model with undesirable outputs and the Malmquist–Luenberger index. Inputs encompass labor, capital and energy; desirable output is proxied by gross regional product, while undesirable the China Energy Statistical Yearbook.
In the calculation of green energy efficiency, this study draws on the classification used by Shi [60] for the input indicators, which include three major categories: labor input, capital input, and energy input. Specifically, labor input is measured by the number of employees at the end of the year in the urban district (in ten thousands); capital input is measured by the year-end capital stock of the city calculated using the perpetual inventory method (in ten thousand yuan); and energy input is measured by the city’s nighttime light data (in ten thousand tons of coal equivalent).
For the output indicators, the desirable output is measured by the actual GDP of each prefecture-level city (in ten thousand yuan), which is calculated based on the constant prices of 2006. As for the undesirable output indicators, this study refers to the research results of Du (2025) [61] and selects industrial SO2 emissions (in tons), industrial wastewater discharge (in ten thousand tons), and industrial smoke and dust emissions (in tons) as specific indicators. Following the research methods of Shao Shuai (2022) et al. [62], the entropy method is employed to integrate the above indicators into a comprehensive environmental pollution index. The specific accounting system is shown in Table 1, and the detailed accounting procedures are provided in Appendix B.
(3) To address potential biases stemming from omitted variable issues, this study refers to the work of Li Jia (2021) [63] and selects control variables in alignment with the research theme: ① Fiscal Investment Intensity (Fin): This metric indicates the degree of financial backing from the government for urban development, potentially impacting urban infrastructure and green energy project investments. Including this control helps to factor in government policy’s influence on energy efficiency and mitigate biases from varying government investment levels. ② Financial Development Level (Finance): Financial development’s extent influences a city’s capital allocation and financing capacities, thereby affecting green energy project execution. Accounting for financial development enables a more precise gauge of artificial intelligence’s impact on energy efficiency across diverse financial contexts. ③ Urban Industrial Structure (Structure): A city’s industrial composition dictates its energy consumption patterns and is intricately linked to energy efficiency. Controlling for industrial structure minimizes industrial variance interference in evaluating artificial intelligence’s effect on energy efficiency. ④ Urban Science and Technology Expenditure (Science): Expenditure on science and technology directly impacts a city’s innovation and technological advancement, notably in artificial intelligence application. Controlling this variable facilitates a more precise appraisal of artificial intelligence’s autonomous effect on green energy efficiency. ⑤ Population Mobility Status (Passenger): Quantified by total passenger volume (measured in ten million people), this metric reflects a region’s economic dynamism and population mobility. Population mobility influences labor market dynamics and income levels, with economically vibrant areas more inclined to embrace artificial intelligence. Controlling this variable assists in excluding these factors’ interference, enhancing the study’s reliability.
All control variables are harvested from the China City Statistical Yearbook spanning 2010–2022. The calculation of the control variables is shown in Table 2:
(4) Mechanism Variables: Based on the theoretical analysis and research hypotheses of this paper, the following three variables are selected as mechanism variables: industrial chain resilience, green finance, and the intensity of environmental regulation:
① Industrial Chain Resilience (ICR): Industrial chain resilience denotes the capacity of the industrial chain to swiftly adjust and return to normal functioning when confronted with external disruptions. The deployment of artificial intelligence (AI) technologies can elevate the industrial chain’s intelligence and automation, consequently bolstering its risk resilience against market volatility and resource scarcity. AI contributes to the construction of a more durable and sustainable industrial chain by refining supply chain management, enhancing production adaptability, and accelerating response times. Moreover, bolstered industrial chain resilience fosters the efficacious distribution and reuse of resources, thereby improving the operational efficiency of urban green energy. As such, industrial chain resilience acts as a pivotal intermediary in the dynamics of AI’s impact on urban green energy efficiency.
② Green Finance: Anchored in the capital’s profit-driven essence, the advent of AI empowers financial entities to more precisely evaluate the risks and rewards of green initiatives, thus channeling increased investment towards superior green energy and conservation sectors. AI also supports regulatory bodies in enhancing the enforcement and oversight of green finance policies, guaranteeing the optimal deployment of capital. Through these avenues, AI significantly promotes advancements in urban green energy efficiency.
③ Environmental Regulation Intensity (Regulation): As AI evolves, a growing cadre of governmental entities leverages AI to manage environmental concerns. AI facilitates more efficient execution of environmental policies by monitoring emissions in real time, forecasting ecological hazards, and streamlining resource management, consequently enhancing the precision and efficacy of environmental stewardship. Additionally, AI aids businesses in uncovering energy conservation and emission cut possibilities, stimulating eco-friendly technological advancements. Propelled by environmental regulations, AI’s application can amplify urban green energy efficiency, attaining a harmonious balance between economic progress and ecological conservation.
In evaluating industrial chain resilience, this study follows Sun Hongxue et al.’s [64] method, focusing on two aspects: resistance and recovery capacity, as well as transformation and innovation capacity. Resistance and recovery capacity are mainly measured by the industrial structure level, upgrading status, and degree of sophistication, using the Herfindahl-Hirschman Index as a benchmark. A lower index value indicates higher industrial diversification, which is beneficial for maintaining industrial chain stability in the face of external shocks. Transformation and innovation capacity are assessed by the number of invention patents granted in the city. More patents suggest a stronger ability to develop and upgrade new products, reflecting the industrial chain’s flexibility and growth potential. After calculating the resistance-recovery capacity index and the transformation-innovation capacity index separately, this study uses the entropy weighting method to compute a weighted average of these indices to form a comprehensive industrial chain resilience index. The underlying data span 2010–2022 and are drawn from the China City Statistical Yearbook.
For green finance assessment, this study builds on Yan Tianshun’s [65] research and others’, creating a green finance evaluation framework for China. It includes four areas: green credit, green securities, green insurance, and green investment. The entropy method, recognized for its objective weighting, is applied for this assessment. The variables employed to quantify green finance are sourced from the China Financial Yearbook for the period 2020–2022.
To gauge environmental-regulation stringency, we adopt the protocol of Shao Shuai et al. (2024) [66], measuring it as the share of characters appearing in sentences that contain environmental keywords relative to the total character count of each city’s Government Work Report. The required corpus covers the 2010–2022 reports released by 271 prefecture-level governments across China.
(5) Instrumental Variables:
① Artificial Intelligence Patent Application Quantity (AIPC): The volume of artificial-intelligence patent applications mirrors a region’s inventive capacity and technological stock, and is, therefore, closely tied to the local level of AI development. Following Chen Nan (2022) et al., patent filings can serve as a valid instrument for AI adoption intensity [67]. Because this variable chiefly captures technological potential rather than any direct pathway to green-energy efficiency, it satisfies both relevance and exogeneity requirements and mitigates endogeneity concerns. ② AI enterprise stock (AIDE): The number of active AI enterprises proxies the degree of industrial agglomeration and reflects the strength of the local ecosystem that facilitates AI diffusion. A larger enterprise pool signals superior infrastructure and resource endowments, thereby accelerating deployment. Since the count is driven by industrial foundations rather than being causally shaped by urban green-energy performance, reverse causality is avoided, making the indicator a suitable instrument. Data on AI patent applications are extracted from the Key Digital Technology Patent Classification System (2023). Using the classification codes assigned to “artificial intelligence”, we match patent records from the China National Intellectual Property Administration for each prefecture-level city between 2010 and 2022. AI enterprise counts are harvested from Tianyancha by filtering firms whose primary business is AI-related and retaining only those currently registered as active.
Owing to data availability, the observation window is capped at 2022. To preserve contemporaneity and align with China’s AI development trajectory, the study employs city-level panel data spanning 2010–2022. Before estimation, the raw dataset is sanitized in two steps: (1) cities with more than 15% missing values are dropped; (2) linear interpolation fills remaining gaps. The final balanced panel encompasses 3523 city-year observations across 271 prefecture.

4.2. Model Specification

Prior to model specification, this study orchestrates a two-stage vetting procedure to anchor the modeling choice. Stage one deploys a random-forest audit to gauge how well the core regressor and the battery of controls explain the variation in the outcome variable, mindful of the panel’s temporal spine. To forestall look-ahead bias and safeguard predictive integrity, the 2010–2022 city panel is split along a time-stratified seam: 2010–2019 serve as the training forest, while 2020–2022 form the hold-out grove. Ten-fold cross-validation is nested within the training segment. The resulting fit delivers an R2 of 0.988 (MSE = 0.012), with the test slice registering 0.976—an inconsequential gap that signals negligible omitted-variable distortion.
Stage two invokes gradient-boosted regression trees to rank the covariates by incremental explanatory clout. Hyper-parameters are tuned to a learning rate of 0.01, max_depth of 5, and 300 boosting rounds, under a squared-error loss. Importance is scored by the cumulative reduction in loss attributable to each variable, then normalized to relative weights. AI towers at 0.32, Science trails at 0.21, followed by Structure (0.18), Fin (0.12), Finance (0.10), and Passenger (0.07). The ordering survives intact in the test set, corroborating the prominence of the AI proxy.
Finally, a Hausman diagnostic (χ2 = 63.51, p = 0.000) decisively favors a two-way fixed-effects architecture—absorbing both city and year dimensions—over its random counterpart, thereby sharpening causal inference. To complement the classical estimator, a neural-network engine is trained on the same panel and its partial-dependence surface is visualized, offering an intuitive lens on the AI–green-efficiency nexus.
The benchmark regression model for the econometric analysis is constructed as follows:
Effkt = α0 + α1Robotkt + α2Controlskt + ζk + δt + εkt
In model (1), the dependent variable is denoted as Effkt, representing the level of Urban green energy efficiency for city k in year t; the independent variable is Robotkt indicating the level of AI for city k in year t. α0 represents the constant term of the model, and αi represents the coefficient of change in the corresponding variable in the linear model with respect to the explained variable. Controlskt represents a set of city-level control variables. ζk signifies regional fixed effects, δt denotes time fixed effects, and εkt is the stochastic error term of the equation.
To examine the mediating effects of industrial chain resilience, green finance, and environmental regulation, this paper constructs a mediation model based on the method proposed by Wen Zhonglin and Ye Baojuan (2014) [68], employing a stepwise regression approach. The mediation model is formulated as follows:
Effkt = η0 + η1 Robotkt + η2 Controlskt + ζk + δt + µkt
Mediatorkt = ϕ0 + ϕ1 Robotkt + ϕ2 Controlskt + ζk + δt + εkt
Effkt = γ0 + γ1 Robotkt + γ2 Mediatorkt + γ3 Controlskt + ζk + δt + θkt
Here, Mediator represents the mediating variable. In models (2), (3), and (4), η0, ϕ0, and γ0 represent the intercept terms of the model, and ηi, ɸi, and γi represent the variable coefficients. According to the stepwise regression method for mediating effect testing, if η1 is significant and ϕ1 is significant, while γ1 is not significant and γ2 is significant, it indicates a full mediating effect. If γ1 is significant, it indicates a partial mediating effect.
In addition, this paper also constructs a neural network model for benchmark regression. Its network hierarchical architecture is input layer (input features)-fully connected layer-Dropout layer-fully connected layer-fully connected layer-fully connected layer (output layer). The list of input features includes Robot (independent variable), Fin, Finance, Science, structure, Passenger, year, city (control variables), Location_conditions, City_centre, Coastal_inland, Transport_facilities, and City_size. Among them, one-hot encoding is used for the five geographical features: Location_conditions, City_centre, Coastal_inland, Transport_facilities, and City_size. Z-score normalization is applied to continuous variables (independent variables and control variables). There are 4 fully connected layers in total, with the number of output channels being 128, 64, 32, and 1 in sequence. The activation function of the first three layers all adopts Relu, which extracts deep features through multi-layer nonlinear transformation. The activation function of the last layer is the default Linear, which linearly outputs through a single channel to accurately predict urban green energy efficiency and achieve the target output of the regression task. The specific construction process and model adjustment process of the neural network are detailed in Appendix C.

5. Empirical Results Analysis

5.1. Benchmark Regression Analysis

To begin, the study deploys a two-way fixed-effects estimator to run the baseline regression on Model (1); the results are reported in Table 3, where N denotes sample size, R2-adj the adjusted goodness-of-fit, City and Year the fixed effects, and Constant the intercept (definitions retained henceforth). Absent any controls, the coefficient on artificial intelligence is positive and significant at the 1% level: a one-unit increase in Robot raises Eff by 0.021 units. After sequentially adding the full suite of covariates, the estimate edges up to 0.026 and remains statistically indistinguishable from zero at the 1% threshold, corroborating Hypothesis 1 that AI enhances urban green-energy efficiency. Tracing the entire energy chain—generation, distribution, and consumption—this uplift materializes as follows.
Upstream, AI refines hydrocarbon drilling parameters and sharpens forecasts for wind-solar output, squeezing more useful energy out of each extracted joule [68]. Mid-stream, smart-grid dispatchers and virtual power plants choreograph distributed resources, trimming transmission and distribution losses [69]. Downstream, behavioral analytics coupled with intelligent appliance control shave peaks, fill valleys, and optimize end-use demand [70]. These inter-locking, chain-wide interventions map neatly onto the positive and significant baseline coefficient, supplying a coherent narrative for AI’s efficiency dividend.
Secondly, this paper uses a neural network model to fit the relationship between artificial intelligence and urban green energy efficiency. Figure 2, Figure 3 and Figure 4 show the dynamic trends of the relationship between artificial intelligence variables and urban green energy efficiency in different cities over time from 2010 to 2022.
In the figure, the crimson trace is the ordinary-least-squares line, capturing the year-specific linear thrust of Robot on eff-SBM, with its slope noted parenthetically; the emerald dashed curve is a LOWESS smoother that sketches the underlying, possibly nonlinear contour; the azure dots are the raw city-year observations plotted without centering or scaling. Each frame (i.e., each annual snapshot) thereby exposes how that year’s AI endowment aligns with green-energy performance, allowing a frame-by-frame inspection of temporal drift. All graphics are drawn from the unstandardized raw values rather than from model predictions or pooled averages.
Neural network simulations depicted in Figure 2, Figure 3 and Figure 4 reveal a pronounced, almost tenacious linear nexus between AI and green-energy efficiency during 2010–2022. Year-to-year, the cloud of points shifts only gently; neither explosive dispersion nor abrupt contraction occurs. The spatial pattern, therefore, appears sticky: cities do not lurch into sudden “quantum leaps” of efficiency once AI crosses some threshold. Heterogeneity, however, remains stark. A handful of municipalities yank the upper tail upward, and in certain years a few high-Robot cities plant isolated high-efficiency outliers, tugging the LOWESS trace slightly skyward. This warrants granular heterogeneity checks beyond the pooled regressions.
Classic OLS underpins the econometric estimates in Table 3. Compared with those results, the neural-network analog yields a flatter linear gradient, yet because the network digests a denser information set, it exposes subtler year-to-year tilts in the AI-efficiency relationship. Both the neural and the econometric verdicts concur: AI exerts a positive and significant lift on urban green-energy efficiency, echoing the baseline finding and sustaining Hypothesis 1.

5.2. Robustness Tests

(1) Addressing Outliers: In statistical analysis, outliers can potentially distort the primary outcomes of the study. To counteract this effect, this paper uses three strategies to manage outliers in continuous variables: (1) applying Winsorization at both ends by 1%, (2) cutting off extreme values, and (3) eliminating outliers from 2020, the initial year of the COVID-19 pandemic, prior to the benchmark regression analysis. The outcomes of the regression analysis are detailed in columns (1) to (3) of Table 4.
The results from column (1) of Table 4 indicate that, following Winsorization, the Robot variable significantly increases urban green energy efficiency (Eff) at the 1% significance level, validating the benchmark regression’s reliability. The findings in column (2) of Table 4 reveal that after truncating the extreme values, Robot continues to significantly enhance urban green energy efficiency (Eff) at the 1% significance level, further substantiating the robustness of the benchmark regression outcomes. In column (3) of Table 4, post-removal of the outliers from 2020, Robot’s positive impact on Eff remains significant at the 1% significance level.
(2) Substitution of Core Indicators: In the baseline regression analysis, the metric for urban green energy efficiency was computed using the Super-Efficiency SBM model. Table 5, column (4), displays values derived from the DEA-CCR model, based on input and output indicators from Table 1. These values replace the SBM-derived results. Column (1) of Table 5 shows that even after substituting the dependent variable indicator, artificial intelligence (AI) maintains a positive impact on urban green energy efficiency at the 1% significance level. Specifically, a one-unit increase in Robot corresponds to an average 0.026-unit increase in Eff, aligning with the baseline regression conclusions.
The baseline regression used industrial robot stock to measure the dependent variable. For robustness testing, this study still uses the Bartik instrumental variable method to construct city-level industrial robot penetration indices. However, instead of stock data, national industrial robot installation increment data across all industries were used to calculate provincial-level increments. These were combined with each city’s manufacturing employee numbers to obtain city-level industrial robot increment data, forming the AI indicator.
After replacing the explanatory variable, regression results are shown in column (2) of Table 5. The results indicate that AI significantly and positively impacts Eff at the 1% significance level. Each one-unit increase in AI leads to an average 0.146-unit increase in Eff, further confirming AI’s potential to enhance urban green energy efficiency.
When both dependent and explanatory variables are replaced simultaneously, regression results in column (3) of Table 5 show that AI still significantly and positively impacts Eff at the 1% significance level. Each one-unit increase in AI results in an average 0.133-unit increase in Eff. In summary, replacing the core indicators does not affect the validity of the baseline regression results and conclusions.
(3) Excluding Policy Interference: The influence of AI on urban-rural income gaps might be tangled with other policy effects, like the Smart City Pilot Policy and the Key Control Experimental Zone for Air Pollution. To avoid this bias and strengthen the research conclusions’ reliability, this paper constructs specific policy variables. Using a DiD model, the interaction of binary dummy variables for policy implementation and implementation time creates a policy dummy variable, which is added to the benchmark regression model. The adjusted benchmark regression results, displayed in Table 6, show that even after incorporating these policy variables, the Robot variable remains significantly positive for Eff at the 1% significance level. This means AI’s positive impact on urban green energy efficiency is robust and not an artifact of other policies, further confirming the reliability of the initial regression results.

5.3. Endogeneity Analysis

(1) IV-2SLS: Given the potential endogeneity of the explanatory variable, which may bias the estimation results of Model (1), this paper employs the Instrumental Variable-Two Stage Least Squares (IV-2SLS) method to address endogeneity. The IV-2SLS is an econometric technique designed to handle endogeneity issues. In the first stage, the endogenous variable is regressed on the instrumental variables to obtain the fitted values of the endogenous variable. In the second stage, these fitted values are used as explanatory variables in the original model for ordinary least squares estimation, yielding consistent estimators. The aim is to resolve estimation biases caused by the correlation between independent variables and the error term, thereby enhancing the accuracy of model estimation and providing a more reliable basis for causal analysis [71].
In the process of identifying instrumental variables, this study meticulously evaluates their pertinence to the explanatory variable as well as their independence from the dependent variable. Two such variables are selected for application in a two-stage regression analysis: ① AIPC (logarithm of AI patent applications in each prefecture-level city): This metric indicates the extent of a city’s innovational investment and technological advancement potential within AI, closely to AI’s growth trajectory. Its limited susceptibility to the urban-rural income disparity fulfills the criteria for the relevance and exogeneity of instrumental variables. ② AIDE (logarithm of the number of AI enterprises in each prefecture-level city): This measure represents the breadth and depth of the local AI industry. Its changes are comparatively disconnected from the urban-rural income divide, qualifying it as an appropriate instrumental variable for the potentially endogenous explanatory variable of AI and assisting in the resolution of endogeneity concerns.
The outcomes of the two-stage least squares regression are presented in Table 7. Based on these results, regarding IV1 (logarithm of AI patents), the first-stage regression (column 1) yields a coefficient of 0.221 for AIPC on Robot, which is significant at the 1% level, affirming the connection between the instrumental variable and the explanatory variable. In the subsequent second-stage regression (column 2), Robot’s predicted influence on urban green energy efficiency (Eff) is notably positive at the 1% level. This finding suggests that even after accounting for the endogeneity of the original explanatory variable, the assertion that AI augments urban green energy efficiency persists, thus confirming Hypothesis 1. For IV2 (logarithm of AI enterprises), in the first-stage regression (column 3), the coefficient of AIDE on Robot is 0.095 and significantly positive at the 1% level, satisfying the relevance requirement. In the second-stage regression (column 4), the predicted effect of Robot on urban green energy efficiency is positively correlated, with statistical significance at the 1% level, further confirming Hypothesis 1.
Moreover, the results of the weak instrument variable test show that the Kleibergen-Paap rk Wald F-statistic significantly exceeds the critical value of 16.38 at the 1% significance level, passing the weak instrument variable test well. The p-values of the Chi-squared (1) statistics are all greater than 0.1, indicating that the Durbin–Wu–Hausman test cannot reject the null hypothesis that the instrumental variables are exogenous with respect to the dependent variable.
(2) Alternative Estimation Methods: Additionally, considering that the calculated urban green energy efficiency might be influenced by historical factors or economic inertia, and that there may be serial correlation among economic variables leading to serially correlated error terms, this paper addresses such endogeneity issues by employing alternative estimation methods. Compared to the static panel data model used earlier, the dynamic panel data model incorporates lagged terms of the dependent variable, which to some extent alleviates the endogeneity problem caused by serial correlation in economic variables.
The analysis employs a dynamic panel data model, segmented into Difference GMM and System GMM approaches. In the Difference GMM method, this study incorporates the first and second lags of the explanatory variable as endogenous factors, while considering the control variables as exogenous for re-estimating the panel data. For System GMM, guided by the AIC information criterion, this study includes the initial and second lags of both the dependent and explanatory variables as endogenous, with control variables remaining exogenous. The results of these models are detailed in columns (1) and (2) of Table 8, respectively. The Robot coefficients are 0.008 and 0.005, both significant at the 1% level, indicating that artificial intelligence continues to positively influence urban green energy efficiency after adjusting for data dynamics. The AR and Hansen test results indicate no issues with serial correlation or weak instruments, supporting the model’s specification. Results from the dynamic System GMM reinforce the study’s robust findings.
(3) Temporal Lag Analysis: Considering AI’s role as urban digital infrastructure and its potential delayed impact on urban green energy efficiency, this study performs a temporal lag analysis. It examines the dependent variable with one and two periods of lag. As per columns (3) and (4) of Table 8, even after applying one and two periods of lag to the AI variable, AI’s effect on urban green energy efficiency remains significantly positive at the 1% significance level. This suggests that the enhancement of urban green energy efficiency by AI is consistent, regardless of the temporal lag in implementation and application of AI.
(4) Principal Component Analysis: This paper employs the Bartik instrumental variable method to measure the level of artificial intelligence in cities and the super-efficiency SBM-GML index method to calculate urban green energy efficiency. These approaches may introduce confounding information into the data, thereby affecting the benchmark regression results. Principal component analysis (PCA), as an effective dimensionality reduction technique, can extract key information and reduce variable redundancy, thereby simplifying the model and lowering the risk of overfitting. Additionally, PCA can mitigate the impact of outliers in individual variables on model results, contributing to enhanced model robustness. Based on this, to avoid interference from potential multicollinearity issues in the original dataset, this paper conducts principal component analysis. Figure 5 illustrates the decay of variables after dimensionality reduction, and the empirical results are presented in Table 9. After performing the benchmark regression with the principal components extracted from each variable, artificial intelligence remains significantly positive for urban green energy efficiency at the 1% significance level. Moreover, the model exhibits high explanatory power, further validating the robustness of the original model and bolstering the credibility of the research conclusions.

6. Further Analysis

6.1. Heterogeneity Analysis

(1)
Location Heterogeneity
Considering the disparities among various regions regarding economic advancement, industrial frameworks, technological prowess, and additional facets, which may shape the influence of artificial intelligence on the urban-rural income gap [72], this paper undertakes an analysis of heterogeneity with respect to location-based attributes. To begin with, utilizing econometric techniques and segmenting the samples by geographical areas, followed by conducting benchmark regression analyses within each category, allows for a more precise examination of AI’s influence on urban green energy efficiency across diverse geographical settings. This methodology, thus, delivers more focused insights applicable for designing region-tailored policies. The heterogeneity regression outcomes regarding location are detailed in Table 10.
According to the Heterogeneity Analysis Results in Table 10, artificial intelligence exerts a significantly positive effect on urban green energy efficiency in the eastern, central, and western regions at the 1% significance level. However, in the northeastern region, a significantly negative effect is observed at the 1% significance level. This may be attributed to the relatively traditional industrial structure of the northeastern region, which has weaker adaptability and application capacity for new technologies. The introduction of artificial intelligence may lead to short-term resource misallocation or shocks to traditional industries, thereby inhibiting the improvement of green energy efficiency.
Furthermore, this paper introduces a subgroup analysis strategy to deepen the research. After one-hot encoding the geographical characteristic variables of regional locations, characteristic subgroups are divided. To ensure the robustness of the model, separate models are trained for subgroup categories with a sample size of ≥30, to explore the geographical heterogeneity of the impact of artificial intelligence investment. The subgroup models continue to use the neural network structure of the main model for heterogeneity regression. The visualization results of the regression are shown in Figure 6, which are the eastern, central, western, and northeastern regions in a clockwise order. In the subgroup regression, the red regression line is the linear fitting of the data by the neural network model. According to the regression results in Figure 6, AI enhances urban green-energy efficiency across the eastern, central, and western belts, yet registers a negative correlation in the northeast—a core pattern fully congruent with the econometric findings. When it comes to the hierarchy of effects, the classical regression ranks the AI-driven boost as east > central > west, whereas the neural network reorders the sequence to central > east > west. The divergence likely stems from the contrast between the regression’s “plain-vanilla average effect” and the network’s nonlinear, attention-weighted effect; each lens refracts the data differently, spawning a benign misalignment rather than a contradiction.
(2)
Heterogeneity by City Size
The impact of artificial intelligence on urban-rural income disparity may vary with differences in city size. Large cities, with their abundant resources, diversified industries, and broad markets, can more effectively leverage AI applications to facilitate green corporate transformations, thereby enhancing urban green energy efficiency. In contrast, smaller cities, with limited resources and weaker capacity to apply AI, may experience different effects [73]. Based on this, this paper refers to the research of Jiang Ren’ai (2023) [74] and others, and categorizes cities into four size groups: megacities, large cities, medium-sized cities, and small cities. Separate benchmark regressions are conducted for each group to examine the impact of AI on green energy efficiency across different city sizes. The econometric regression results are presented in Table 11.
According to the regression results in Table 11, for three types of cities, artificial intelligence has a positive effect on their green energy efficiency. Notably, for large cities (2), the impact coefficient of AI is 0.054, and it significantly promotes green energy efficiency at the 1% significance level. This aligns with the characteristics of large cities, which have abundant resources, diversified industries, and broad markets. It indicates that in these cities, the application of AI helps drive green transformation and improve energy efficiency.
Additionally, the regression results from the neural network model in Figure 7 show that artificial intelligence generally has a positive impact on urban green energy efficiency. The image, in a clockwise direction, represents megacities, large cities, medium and small cities, and small cities. Notably, the positive impact is more significant in large cities, medium and small cities, and small cities. In megacities, the model’s fitting results tend to show a negative correlation, but based on the image, this may be due to fitting errors caused by outliers. The overall results are largely consistent with those obtained from the econometric model.
(3)
Whether a City is a Central City
Central cities typically exhibit strong economic influence and clustering effects, and their development and application levels of artificial intelligence are often higher. In major cities, AI can drive industrial upgrading and coordinated development in surrounding areas, making its impact on urban green energy efficiency more complex. In contrast, the impact in non-major cities is relatively limited. Therefore, drawing on the research of Zhao Tao (2020) [75] and others, this paper divides the samples into two groups: major cities and non-major cities, and conducts separate benchmark regression analyses to reveal the heterogeneous impact of AI on urban green energy efficiency in different types of cities. The regression results are shown in columns (1) and (2) of Table 12.
Columns (1) and (2) of Table 12 show that, at the 1% level, AI exerts a positive and significant impact on green-energy efficiency in non-core cities, whereas the effect on core cities is significant only at the 10% threshold. Relative to their core counterparts, non-core municipalities experience a markedly larger boost; these cities typically lag in economic endowment and industrial sophistication. The diffusion of AI, thus, injects fresh momentum, more forcefully accelerating their green transition. The neural-network analog is plotted in Figure 8: the left-hand panel depicts non-core cities, the right-hand panel core cities. Within each frame, the AI–green-efficiency gradient slopes sharply upward, and the marginal-response curves for the two city types are virtually parallel. This visual alignment corroborates the econometric finding that the uplift is steeper—yet similarly shaped—for non-core cities.
(4)
Whether a City is a Coastal City
Given the close relationship between artificial intelligence and urban green energy efficiency, coastal cities, with their advantageous geographical locations, are generally more likely to attract industries and technological innovations in the field of artificial intelligence. This geographical advantage may enable coastal cities to achieve more significant and unique positive impacts on green energy efficiency through the application of AI technologies. In contrast, inland cities may be relatively less open to the outside world, with fewer trade opportunities and less convenient transportation, which could affect the role of AI technologies in enhancing green energy efficiency. Therefore, drawing on the research findings of Luo Fangyong (2023) and others [76], this paper divides the research samples into two groups: coastal cities and inland cities, and conducts separate benchmark regression analyses for each group. This approach aims to compare and analyze the differences in the impact of AI on green energy efficiency between the two types of cities, thereby providing a scientific basis for formulating more precise and effective policies. The econometric regression results are presented in columns (3) and (4) of Table 12.
According to the regression results in Table 12, it can be observed that AI has a significantly positive impact on the green energy efficiency of both coastal and inland cities, with statistical significance at the 1% level. However, coastal cities, which may have more developed economic environments, more convenient trade channels, and more frequent international exchanges, tend to have a wider and deeper application of AI technologies. Therefore, their role in enhancing green energy efficiency is more pronounced than that of inland cities. This indicates that the geographical and economic conditions of coastal cities may provide a more favorable environment for the implementation and effectiveness of AI technologies.
Figure 9 visualizes the neural-network heterogeneity tests: the left pane portrays inland cities, the right pane their coastal counterparts. The simulation reaffirms that artificial intelligence exerts a positive push on urban green-energy efficiency, broadly echoing the econometric verdict. Yet the magnitudes diverge: the classical regression assigns the larger uplift to coastal agglomerations, whereas the neural network tilts the scale toward inland cities, signaling that nonlinear interactions captured by the network re-weight the coastal–inland hierarchy.
(5)
Whether a City is a Transportation Hub
In contrast to non-transportation hub cities, transportation hub cities, benifit by their robust transportation networks, are better positioned to facilitate the spread and implementation of artificial intelligence technologies, which is especially important for boosting urban green energy efficiency. These cities can swiftly disseminate AI advancements to nearby areas, fostering coordinated regional development and amplifying the impact of AI on urban green energy efficiency to be more substantial and multifaceted. Consequently, leveraging the Medium and Long-term Railway Network Plan (2016) [77] as a reference, this study categorizes the research samples into two distinct groups: transportation hub cities and non-transportation hub cities, and carries out benchmark regression analyses to deeply examine the varying impacts of AI on urban green energy efficiency across different transportation settings. The econometrics regression outcomes are displayed in columns (5) and (6) of Table 12.
As per the regression findings in columns (5) and (6) of Table 12, artificial intelligence notably enhances the green energy efficiency in both non-transportation hub cities and transportation hub cities at the 1% significance level. A comparison of the Robot coefficients reveals that the influence is more pronounced in transportation hub cities than in their non-transportation counterparts. Nonetheless, because of the restricted sample size of transportation hub cities, the statistical significance does not reach a certain threshold.
Figure 10 illustrates the regression findings from the neural network model concerning this heterogeneity, with non-transportation hub cities depicted on the left and transportation hub cities on the right. In line with the regression outcomes in Figure 10, artificial intelligence bolsters urban green energy efficiency, exerting a more significant effect in transportation hub cities, which corresponds well with the insights gleaned from the econometrics regression results.
(6)
Individual City Analysis
Based on the aforementioned heterogeneity analysis, this paper further conducts a city-by-city neural network regression analysis of the relationship between AI and urban green energy efficiency. The regression slopes are extracted as indicators of impact magnitude, and a bar chart ranking the cities is generated to clearly display the differences in effects among cities and quantify the geographical heterogeneity. The results are shown in Figure 11.
According to the results presented in Figure 11, there is a significant variation in the impact of AI on urban green energy efficiency across different cities. Based on this, the paper identifies the five cities with the highest and the five with the lowest slopes, as shown in Table 13.
Analysis of the Ten Extreme Values Based on Table 13:
① Dongguan: Dongguan has a slope value of −0.1653, showing the most significant negative impact of AI on green energy efficiency among all cities. This may be due to its focus on labor-intensive manufacturing, which has a weaker capacity to absorb and apply AI technologies. As a result, the introduction of these technologies has not effectively translated into improved green energy efficiency and may have even led to resource wastage in the short term due to technological adjustment issues.
② Longnan: Longnan has a slope value of −0.1389, and its industrial structure is relatively simple, mainly relying on traditional agriculture and a small amount of mineral resources. The demand for AI in such industries is low and the application is difficult, making it difficult for AI to play a role in the field of green energy, and may even interfere with the original production process and affect energy efficiency due to inapplicable technical means.
③ Sanya: Sanya has a slope value of −0.1299. As a tourist city, its energy consumption is primarily concentrated in the tourism service industry, which is highly affected by tourist flow and seasonal changes. The application of AI in the tourism sector may not have sufficiently extended to energy management, resulting in no visible improvement in green energy efficiency and even potential negative effects due to excessive energy consumption during peak tourist seasons.
④ Ziyang: Ziyang has a slope value of −0.1214. Located inland with a relatively lower level of economic development, it lacks adequate technological infrastructure and talent reserves. This limits the promotion and application of AI technologies in the energy field, leading to negative impacts on green energy efficiency even if relevant technologies are introduced.
⑤ Maoming: With a slope value of −0.1037, Maoming faces many challenges in its industrial transformation as a resource-based city. The application of artificial intelligence technology in traditional industries may be restricted by the existing production model and interest pattern, and it is difficult to fully play its role, and may even reduce energy efficiency due to inadaptability in the process of industrial adjustment.
⑥ Zhengzhou: Zhengzhou has a slope value of 0.1637. As a transportation hub and the core city of the Central Plains Economic Zone, it has dense logistics and information flows, facilitating the rapid dissemination and application of AI technologies. Government policy support for the integration of AI and green energy, combined with a diversified industrial base, enables AI to effectively optimize energy allocation and enhance green energy efficiency.
⑦ Langfang: With a slope value of 0.1649, Langfang is close to Beijing and actively undertakes industrial transfer and technology spillover from the Beijing-Tianjin region. Under the Beijing-Tianjin-Hebei coordinated development strategy, Langfang focuses on the development of high-tech industries, and artificial intelligence is widely used in energy management, which can effectively improve the efficiency of green energy utilization.
⑧ Heze: Heze has a slope value of 0.1850. In recent years, Heze has increased its technological investment and actively introduced emerging technologies such as AI to create new energy industry clusters. A series of preferential policies by the government have attracted several new energy companies, promoting the development and utilization of green energy and enabling AI to play a significant role in enhancing green energy efficiency.
⑨ Shenzhen: Shenzhen has a slope value of 0.1939. As a center for technological innovation, Shenzhen boasts strong research capabilities and a complete innovation ecosystem. The widespread application of AI technologies in Shenzhen’s energy sector, through innovations such as smart grids and energy management systems, has effectively improved energy utilization efficiency. Additionally, Shenzhen’s industrial diversification and market vitality provide ample space for the integrated development of AI and green energy.
⑩ Jinan: Jinan has a slope value as high as 0.1951. As the capital of Shandong Province, Jinan has significant advantages in the integrated development of AI and green energy. As a provincial capital, it enjoys policy support as well as funding and talent resources. Local companies actively explore innovative applications of AI in the green energy field, including intelligent energy monitoring and green transportation. These factors collectively drive the improvement of Jinan’s green energy efficiency, making it stand out in this area.
Overall, the impact of AI on urban green energy efficiency exhibits significant heterogeneity. Due to differences in industrial structure, technological level, policy support, and other factors, the effects of AI in promoting green energy efficiency vary across cities.

6.2. Mechanism Tests

To verify Hypothesis 2 and explore the indirect impact of artificial intelligence (AI) on urban green energy efficiency, this paper constructs a mediating econometric model to investigate the mechanisms through which AI affects urban green energy efficiency. Furthermore, a chain mediation model is employed to further explore the interactions between these mechanisms.
(1)
Industrial Chain Resilience
Industrial chain resilience indicates how well the industrial chain adapts, recovers, and innovates in response to external disruptions. The theoretical discussions [78] highlight that this resilience significantly mediates AI’s effect on urban green energy efficiency. Following the assessment of industrial chain resilience, this study uses stepwise regression for empirical analysis, with findings presented in columns (1) and (2) of Table 14. The mechanism test results in columns (1) and (2) of Table 14 show that AI significantly enhances urban industrial chain resilience at the 1% significance level, demonstrating AI’s capacity to fortify the urban industrial chain’s resilience. In column (2), AI also notably promotes urban green energy efficiency at the 1% significance level, with the industrial chain resilience variable (ICR) coefficient at 0.527, significantly positive at the 1% significance level. Overall, it is evident that AI boosts urban green energy efficiency by improving urban industrial chain resilience, marking it as a pivotal and essential mechanism.
(2)
Green Finance
Green finance is instrumental in the role AI plays concerning urban green energy efficiency. Significant financial investments are necessary for AI development and application in the green energy sector, involving R&D, equipment procurement, and talent development. Green finance offers funding avenues for green energy projects, lowers financing costs, and advances AI implementation in green energy [79]. This study empirically examines green finance’s mediating role, based on theoretical analysis, with results in columns (3) and (4) of Table 14.
As per the test results in columns (3) and (4) of Table 14, AI significantly advances urban green finance at the 1% significance level, indicating AI’s role in enhancing city green finance indices. In column (4), AI significantly promotes urban green energy efficiency at the 1% significance level, with green finance’s coefficient being significantly positive at the 1% significance level. It is concluded that AI elevates urban green energy efficiency by improving city green finance indices.
(3)
Environmental Regulation Intensity
Environmental regulation intensity acts as a guiding and regulatory force in AI’s impact on urban green energy efficiency. Stringent and rational environmental regulations encourage businesses to adopt AI technologies to enhance green energy efficiency, meeting environmental standards. As environmental regulation intensity increases, businesses face heightened pressure to curb energy use and emissions. AI technologies assist businesses in refining energy management and boosting energy efficiency, thus achieving energy conservation and emission reduction goals [80]. Based on theoretical analysis, this study empirically tests environmental regulation intensity’s mediating role, with results shown in columns (5) and (6) of Table 14.
The test results in columns (5) and (6) of Table 14 indicate that AI significantly promotes city environmental regulation intensity indices at the 1% significance level, showing AI’s ability to strengthen city environmental regulation intensity, validating theoretical insights. In column (6), AI significantly fosters urban green energy efficiency improvement at the 1% significance level, with the Regulation coefficient at 0.013, significantly positive at the 1% significance level. In summary, it is concluded that AI enhances urban green energy efficiency by increasing city environmental regulation intensity.
Far from operating in silos, the three mechanisms weave a synergistic tapestry across the entire energy continuum—generation, distribution, and consumption. At the production node, bolstered supply chain resilience stabilizes the flow of critical components for wind turbines and PV arrays while accelerating technological refresh cycles, ensuring that green-generation hardware is both continuously available and progressively upgraded. This same resilience, by synchronizing upstream and downstream actors, trims logistical frictions and curbs energy losses during transport and dispatch. In parallel, green finance channels targeted capital: it bankrolls the construction of renewable hubs and smart-grid retrofits on the supply side, and simultaneously extends credit lines that allow firms to swap out legacy equipment and households to adopt intelligent, energy-sipping appliances on the demand side. Meanwhile, heightened environmental regulation acts as a double-edged catalyst. It coerces high-energy industries to embed AI-monitored, low-carbon processes at the point of production, while steering consumers—via labeling standards and subsidy schemes—toward high-efficiency household devices at the point of use. Together, these interlocking levers generate a compounding, chain-wide uplift in green-energy performance.
(4)
Chain Mediation Effects
The previous discussion on mechanisms focused on the individual effects of each mechanism. However, considering the potential interactions between these mechanisms, this paper employs a chain mediation model to further examine these interactions. The model is constructed as follows:
Mediator2kt = ų0 + ų1 Robotkt + ų2 Controlskt + ζk + δt + µkt
Mediator1kt = ƫ0 + ƫ1 Mediator2kt + ƫ2 Robotkt + ƫ3 Controlskt + ζk + δt + εkt
Effkt = ȡ0 + ȡ1 Robotkt + ȡ2 Mediator1kt + ȡ3 Controlskt + ζk + δt + εkt
In Models (5), (6), and (7), if ų1, ƫ1, ȡ1 and ȡ2 are all significant, then the second mechanism variable interacts with the first mechanism variable. The explanatory variable first affects Mechanism Variable 1, which then affects Mechanism Variable 2. Through the combined influence of Mechanism 1 and Mechanism 2, the dependent variable is ultimately affected.
Finally, through a series of empirical tests, this paper finds significant chain mediation and interaction effects between environmental regulation intensity and industrial chain resilience, as well as between green finance and industrial chain resilience. Although there is no direct chain mediation effect between environmental regulation intensity and green finance, the inclusion of their interaction term in the benchmark regression model reveals a strong interaction effect between the two.
The relationship between environmental control intensity and industrial chain robustness is detailed in Table 15, columns (1) to (3). Based on the findings from Table 15, it is evident that the utilization of artificial intelligence substantially enhances the level of urban environmental governance at a 1% significance threshold. In column (2), it is shown that environmental governance intensity notably advances the intermediary factor of industrial chain robustness at the 1% significance level. Column (3) reveals that this intermediary factor of industrial chain robustness significantly boosts urban sustainable energy efficiency also at the 1% significance level. This suggests that the combination of environmental governance intensity and industrial chain robustness drives urban economic growth. Furthermore, the confidence intervals estimated via Bootstrap do not encompass 0, confirming the reliability of the chain mediation effect.
The interplay between green finance and industrial chain robustness is depicted in Table 15, columns (4) to (6). As per the data in Table 15, column (4) indicates that the incorporation of artificial intelligence markedly advances green finance at a 1% significance level. Column (5) highlights that environmental governance intensity substantially enhances the green finance variable at the 1% significance level. Column (6) establishes that the intermediary factor of industrial chain robustness significantly elevates urban green energy efficiency at the 1% significance level. This points to the idea that environmental governance intensity and industrial chain robustness together stimulate urban green energy efficiency. Moreover, the Bootstrap-derived confidence intervals exclude 0, affirming the solidity of the chain mediation.
(5)
Mechanism Synergies
Beyond examining how artificial intelligence enhances urban sustainable energy efficacy and the interrelations among these mechanisms, this study also accounts for the interactions among these factors within the standard regression framework <1>. This approach evaluates how the combined effect of two mechanism variables influences the outcome variable. The findings from the analysis of these combined effects are detailed in Table 16.
In Table 16, Interaction1, Interaction2, Interaction3, and Interaction4 correspond to the combined effects of environmental governance intensity with green finance, environmental governance intensity with industrial chain robustness, green finance with industrial chain robustness, and the combination of all three, respectively. The results from columns (1) to (4) of Table 16 indicate that the coefficients for Interaction1 through Interaction4 are significantly positive across various significance thresholds. This demonstrates that the collaborative impact of these mechanisms further advances urban sustainable energy efficiency.
In summary, based on the empirical regression results of the mechanism tests, Hypothesis 2 proposed in this paper is validated.

6.3. Spatial Effect Test

(1)
Research Design
When examining the impact of artificial intelligence on urban green energy efficiency, it is crucial to investigate spatial spillover effects. Cities are interconnected and influence each other in terms of economy, technology, and talent. The application of AI and the improvement of green energy efficiency can generate spillover effects within a region [81]. When a city achieves significant success in optimizing energy management through AI, it attracts neighboring cities to learn and emulate its practices. Additionally, technology diffusion and talent mobility can drive overall efficiency improvements in the region [82]. Therefore, studying spatial spillover effects can provide a more comprehensive understanding of the comprehensive impact of AI on regional green energy development, offering a scientific basis for policymaking and promoting coordinated regional green development.
Based on this, this paper conducts a spatial econometric empirical test of the spatial spillover effects of AI applications on green energy efficiency. Since the spatial economic-geographic nested matrix, which combines economic differences and geographical distances, can reflect the spatial dependence of economic interactions between regions, this paper employs this matrix for spatial econometric analysis. The method for constructing the matrix is detailed in Appendix D.
(2)
Global Spatial Autocorrelation Test
Initially, the study utilizes the Global Moran’s I index to analyze the spatial relationship between the deployment of artificial intelligence and urban sustainable energy productivity. Table 17 demonstrates that the Moran’s I values for AI deployment and sustainable energy efficiency are not zero, highlighting a notable spatial interconnection. This implies that AI has a dual influence: it directly enhances local sustainable energy productivity and potentially impacts the productivity in nearby cities via spatial diffusion effects, thus influencing the region’s overall sustainability. Consequently, to precisely gauge AI’s role in urban sustainable energy productivity, this paper builds a spatial econometric model to delve into its propagation dynamics across geographic spaces.
(3)
Spatial Econometric Model Test and Result Analysis
Table 18 presents the findings from the LM and Robust LM tests, which confirm the statistical significance of both spatial error and spatial lag components. To manage the model’s intricacies, the Spatial Durbin Model (SDM) is employed to explore the effects of spatial diffusion, with detailed results provided in Table 19.
Drawing on the regression outcomes from the spatial econometric model shown in Table 19, column (1) reveals that the coefficient associated with the Robot variable stands at 0.023, significant at the 1% level. This signifies a strong positive direct impact of AI on enhancing local urban green energy efficiency, underlining AI’s pivotal role in improving a city’s green energy performance. Regarding the indirect impact, column (2) indicates that the Robot variable’s coefficient is 0.008, also significant at the 1% level, suggesting a substantial positive indirect effect of AI on the green energy efficiency of nearby cities. This underscores AI’s dual contribution to both local enhancements and the amplification of green energy efficiency in neighboring areas through spatial diffusion processes. Column (3) displays the total effect, with the Robot variable’s coefficient at 0.031, significant at the 1% level, signifying a considerable overall positive influence of AI on green energy efficiency. This combined impact, incorporating both direct and indirect effects, confirms the presence of spillover effects and the comprehensive positive impact of AI on urban green energy efficiency, thereby validating the Hypothesis 3 presented in this study.
(4)
Risk Scanning and Deliberation
While the headline finding—that AI propels urban green-energy efficiency upward—is robust, any policy prescription would be half-baked if it ignored the lurking “green-inequality” hazards. Heterogeneity diagnostics across 271 cities surface three distinct risk archetypes:
  • Algorithmic bias that magnifies regional cleavages.
Our mechanism tests show AI leverages green finance to lift efficiency, yet the very algorithms that enable this boon may embed discriminatory credit rationing. Regressions in Table 10 and Table 11 reveal that coastal and mega-cities enjoy dense data infrastructures and a richer inventory of high-quality green projects. AI credit-scoring models trained on these lopsided samples naturally tilt toward well-rated incumbents. By contrast, smaller or northeastern cities—characterized by sparse data and modest project scales—are algorithmically tagged as “high-risk”, pushing their green-finance costs in the wrong direction and throttling efficiency gains where they are needed most.
2.
Technological thresholds that calcify the “digital chasm”.
Again, drawing on Table 10 and Table 11, transport-hub and core cities register steeper marginal effects. Armed with abundant compute power and talent pipelines, these hubs can continuously refine AI-driven energy-management systems, driving marginal abatement costs ever lower. Non-hub cities, bereft of specialized O&M teams, may deploy the identical algorithm yet suffer performance decay because of sluggish parameter tuning. Figure 9 corroborates this asymmetry: although AI coefficients for transport hubs are the highest, the underlying sample counts only 233 city-years, signaling that the technological dividend is pooling in a handful of nodal cities and potentially sharpening spatial polarization in green-energy efficiency.
3.
Policy arbitrage and “green crowding-out”
Mechanism tests in Table 14, columns (5) and (6), confirm that AI intensifies environmental regulation, which in turn elevates efficiency. Yet this virtuous loop is vulnerable to gaming. Firms in tightly regulated jurisdictions can weaponise AI to hover precisely beneath compliance thresholds, relocating energy-intensive processes to neighboring cities where oversight is lax and data patchy. The spatial Durbin model in Table 19 shows that indirect spillovers are net positive, but the city-level heterogeneity diagnostics in Table 13 sound a cautionary note: Dongguan and Longnan display negative slopes, hinting at possible displacement effects—local abatement achieved at the cost of cross-border emissions.
By flagging these risks, the paper completes its diagnostic arc. Corresponding safeguards and countermeasures are laid out in the concluding section to ensure that AI’s green dividend is shared rather than skewed.

7. Discussion

Regarding impact, scholars have extensively studied AI’s beneficial role in enhancing urban green energy efficiency. Hao et al., examining production efficiency, discovered that AI’s application substantially boosts energy extraction and usage efficiency, cutting urban energy consumption and thus promoting urban green energy efficiency [83]. Yang et al. noted that AI’s use in smart grid development and energy management optimization has spurred the intelligence upgrade of urban energy systems, enhancing the precision of energy allocation and utilization, and positively affecting urban green energy efficiency [84]. Wu et al. highlighted AI’s capacity to optimize energy supply and demand matching through big data analytics and intelligent forecasting, reducing waste and improving urban green energy efficiency [85]. These investigations underscore AI’s substantial potential in boosting urban energy utilization efficiency and energy structure optimization, offering strong technical backing for urban green energy progression.
Concerning the mechanisms of impact, existing research has thoroughly explored how AI bolsters urban green energy efficiency. Liu et al. posited that AI, by streamlining energy production processes and enabling intelligent operation and maintenance of energy equipment, increases energy output rates, thus fostering urban green energy efficiency improvement [86]. Li et al. suggested that AI applications offer customized energy management solutions for consumers, guiding them to adjust energy use patterns, reducing consumption, and enhancing urban green energy utilization efficiency [87]. Wang et al., through empirical research, found that innovative AI applications in energy domains, such as smart microgrids and virtual power plants, integrate distributed energy resources effectively, heighten urban energy systems’ flexibility and adaptability, and, thus, enhance green energy efficiency [88]. Ji et al. also probed AI’s role in fostering energy technology innovation and industrial upgrading, indirectly spurring urban green energy efficiency advancement [89].
This study further broadens and deepens the comprehension of AI’s relationship with urban green energy efficiency. Although prior studies predominantly focus on singular impact mechanisms, this study systematically delineates AI’s influence across three dimensions: green finance, industrial chain resilience, and environmental regulation intensity. While prior literature has explored AI’s macro mechanisms, this study delves into specific pathways through which AI enhances urban green energy efficiency by bolstering green finance, strengthening industrial resilience, and intensifying environmental regulation. Empirical findings confirm that AI significantly improves urban green energy efficiency, a conclusion robust to multiple robustness checks. Additionally, AI directly affects local green energy efficiency and exerts spatial spillover effects, underscoring its substantial value in regional green energy development.
Moreover, this study examines AI impact heterogeneity on urban green energy efficiency. Results indicate variable effect magnitudes across cities with differing regional locations, sizes, and transportation conditions, yet AI consistently demonstrates a positive influence. In eastern regions, larger, and central cities, AI’s enhancing effect on green energy efficiency is more pronounced, providing crucial evidence for localized AI strategy formulation.
In conclusion, this study validates AI’s positive effect on urban green energy efficiency and achieves significant research perspective and content breakthroughs. However, it has limitations. First, variable selection may have omitted important factors like policy and socio-cultural aspects, potentially influencing the AI-urban green energy efficiency relationship. Second, indicator measurement may be limited; this study mainly examines AI’s impact through green finance, industrial chain resilience, and environmental regulation intensity, possibly not capturing AI’s full potential in other areas.
Based on the above analysis, future research could improve by: First, expanding the data scope to include more cities globally and extending the research period to reflect AI’s long-term impact and dynamic changes on urban green energy efficiency comprehensively. Second, enriching variable and indicator selections by incorporating additional factors affecting urban green energy efficiency, such as policy and socio-cultural elements. Third, adopting a multidimensional measurement approach to construct a more comprehensive integrated indicator system for accurately assessing AI’s impact on urban green energy efficiency.

8. Conclusions and Policy Recommendations

Amidst the swift advancement of artificial intelligence (AI) technologies, it is crucial to delve into the opportunities and challenges AI presents for improving urban green energy efficiency, aiding countries in aligning their energy strategies with green growth initiatives. Utilizing panel data from 271 cities across China spanning 2010 to 2022, this study integrates neural network models with econometric techniques to thoroughly examine AI’s impact on urban green energy efficiency. The findings indicate that: ① AI substantially bolsters urban green energy efficiency, a finding that persists following an array of robustness and endogeneity assessments, including outlier management, alternative estimation methods, and instrumental variable approaches. ② Analysis of heterogeneity demonstrates varying impacts of AI across different geographic areas, city sizes, and city types such as central, coastal, or transport hubs, yet the general trend remains positive. The influence of AI on green energy efficiency is particularly evident in the eastern regions, larger cities, and central cities. ③ Mechanistic evaluations uncover that AI fosters urban green energy efficiency via several avenues, namely by elevating green finance levels, fortifying industrial chain resilience, and intensifying environmental regulations. Additionally, an analysis of synergistic mechanisms pinpoints considerable interactions between green finance, industrial chain resilience, and environmental regulation intensity, all of which collaboratively contribute to the enhancement of urban green energy efficiency. ④ Spatial spillover evaluations reveal that AI’s influence on urban green energy efficiency extends positively to adjacent cities.
Based on the above research conclusions, the following policy recommendations are proposed:
① Encourage AI Utilization in Green Energy: Governments need to implement specific policies that incentivize the integration of AI technologies by businesses to enhance green energy production, distribution, and consumption. This could involve offering financial subsidies for the adoption of AI in green energy initiatives like smart grids and distributed energy management systems. Additionally, tax incentives could be provided for projects that utilize AI to boost renewable energy efficiency, thereby cutting operational costs, fueling innovation, and fostering the advancement of the green energy sector, as well as improving overall energy efficiency.
② Boost Urban Green Energy Infrastructure: Tailor infrastructure development strategies to suit the unique attributes of different urban environments. In larger cities, prioritize the establishment of AI data centers and intelligent energy monitoring systems to elevate the sophistication of urban energy management. For medium, small cities, and transportation hubs, focus on enhancing the green energy access and distribution network, employing AI to streamline energy allocation and minimize transmission and distribution losses. In coastal regions, exploit geographic advantages to develop offshore wind and marine energy facilities, using AI to enhance the efficiency and reliability of energy generation, thus optimizing the urban energy mix and increasing the share of green energy in the total energy consumption.
③ Elevate Urban Green Finance Development: Stimulate financial institutions to create a broader array of green financial products, including specialized loans and green bonds for AI-assisted green projects. Governments could set up risk compensation funds to mitigate financial institutions’ concerns regarding green energy project risks, thereby channeling more investment into green sectors. Additionally, enhance financial regulatory innovation by employing AI to improve transparency and efficiency within the green finance market, curb financial risks, and ensure the stability of green finance, providing robust financial backing for enhancing urban green energy efficiency.
④ Foster Industrial Chain Synergy: Governments should establish industry-academia-research collaboration platforms to stimulate partnerships between AI and green energy companies, hastening the commercialization of technological breakthroughs. Encourage leading enterprises to guide the industrial chain, advocating for collective AI adoption among upstream and downstream businesses to establish green energy industry clusters. Through policy guidance, facilitate the connection and integration of industrial chains across different cities, refine regional industrial structures, and improve the resilience of these chains. This approach will enable resource sharing, create synergies, and enhance regional green energy efficiency and economic outcomes.
⑤ Enhance Environmental Regulations and Policy Synergy: Strengthen environmental regulations and increase scrutiny over high-energy-consuming and high-polluting enterprises. Implement AI to enhance environmental monitoring and enforcement, ensuring effective implementation of energy-saving and emission-reduction measures by businesses. Sync environmental regulations with green development goals, using tax adjustments and financial incentives to promote the adoption of green energy technologies among businesses, steering industrial structures towards green and low-carbon development. It is also crucial to align environmental regulations across regions to create a unified front against environmental challenges, thus fostering regional green energy efficiency and sustainable development.
⑥ Improve the Legal and Standards System for AI: Accelerate the formulation of special regulations for the application of AI in the green energy field, clarify the responsibilities and obligations of all parties, and ensure the healthy development of AI technologies; establish unified green energy data standards and interface specifications to promote data sharing and collaboration between different systems, improving the efficiency of AI applications; strengthen the regulation of AI algorithms to ensure their fairness, transparency, and interpretability, preventing energy allocation injustices caused by algorithmic biases, and ensuring the fairness and efficient operation of the urban green energy system.
⑦ Crafting a multi-layer policy architecture that interlocks algorithmic governance, technology inclusivity, and regulatory synergy.
Algorithmic governance: roll out a nationwide, green-finance algorithm-audit protocol. Mandate that AI credit-scoring engines ingest balanced training sets containing historical data from rust-belt Northeast and small-to-mid-size cities in central–western China; embed a “regional fairness weight” to correct the built-in tilt toward high-grade projects. Require every green-project evaluation model to publish a fairness dashboard—if resource-based or less-developed regions host micro-scale renewable schemes, the algorithm must assign them non-trivial weights, shaving their financing threshold.
On the technology-inclusion front, we propose anchoring regional AI-energy service hubs to the national supercomputing spine. These hubs would deliver off-the-shelf algorithmic plug-ins and round-the-clock O&M assistance to non-hub cities, dissolving the “digital moat” forged by skill shortages. Tailored AI-plus-retrofit boot camps would also be rolled out for the Northeast and other heavy-industry heartlands, equipping plant managers with the hands-on know-how to deploy and fine-tune intelligent monitoring systems.
Regulatory synergy: weave a cross-jurisdictional energy-flow ledger on blockchain, creating an immutable trail that flags any migration of carbon-intensive processes. Complement this with a regional environmental-governance compact: when algorithmic bias misallocates green resources, trigger a compensatory transfer-payment scheme under central fiscal tools, narrowing inter-city green-energy investment gaps and safeguarding the universal and sustainable dividends of AI.

Author Contributions

Conceptualization, Y.D.; Software, T.L.; Validation, Y.D. and W.S.; Resources, T.L.; Data curation, Y.D. and T.L.; Writing—original draft, Y.D. and T.L.; Writing—review & editing, J.L.; Visualization, W.S.; Funding acquisition, J.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the China College Students’ Innovation and Entrepreneurship Training Program (research on “Turning “Carbon” into “Gold”: A Study on the Pathways for Corporate Transformation Based on the Decision-Making Mechanism of Carbon Trading Pricing—Taking Typical Energy Enterprises along the Yellow River as Examples”), with grant number 202410445021. Additionally, this research is also funded by the Undergraduate Research Fund of Shandong Normal University (the research topic is “Research on the Effect of Digital Inclusive Finance on Poverty Reduction and Income Increase in Rural Areas from the Perspective of Common Prosperity”), with the grant number BKJJ2025012.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

This paper selects the panel data at the urban level in China and the panel data at the urban level from 2011–2022. The data mainly comes from the China Urban Statistical Yearbook, China Energy Yearbook, China Environmental Yearbook, China Financial Yearbook and the government work reports of cities.

Conflicts of Interest

The authors declare no conflict of interest.

Appendix A

For the accounting of the explanatory variable “artificial intelligence” in this paper, with reference to existing literature, the specific accounting methods are as Equation (A1):
R o b o t i t = r o b o t s t o c k _ c i t y i . t / e m p i , t = 2004 = [ θ i . k . t = 2004 × j = 1 J ( E k , j , t = 2004 × R o b o t s j , t ) ] / e m p i , t = 2004
Here, R o b o t i t represents the penetration rate of industrial robots in city i in year t; r o b o t s t o c k _ c i t y i . t is the stock of robots in city i in year t; e m p i , t = 2004 denotes the number of employees in the manufacturing sector of city i in 2004 (unit: thousand people); θ i . k . t = 2004 indicates the proportion of manufacturing employees in city i in 2004 relative to the total number of manufacturing employees in its province; E k , j , t = 2004 represents the proportion of employees in industry j of city k in 2004 relative to the total number of employees in industry j nationwide; R o b o t s j , t is the stock of robots in industry j in year t.

Appendix B

For the accounting of the explained variable “green energy efficiency” in this paper, the non-radial SBM-ML method is adopted, and the specific accounting methods are as follows:
First, a single-period SBM-DDF underlying efficiency model is constructed to calculate the static efficiency of decision-making units at time t relative to the frontier, with the calculation steps as follows. Firstly, construct the decision-making units (DMU). Each municipal administrative region is set as a DMU. Let X R m × n be the input matrix, Y g R S 1 × n be the desirable output matrix, and Y b R S 2 × n be the undesirable output matrix. The corresponding slack variables are denoted by, S R m × n , S g R S 1 × n and S b R S 2 × n , respectively. There are n DMU in total, with m inputs, S1 desirable outputs, and S2 undesirable outputs.
Second, define the constraints and construct the objective function. In the matrix, each row represents a variable, and each column represents a DMU. It can be expressed as the k-th input variable for the i-th DMU. Each DMU consists of m inputs X = x 1 , x 2 , , x m , S outputs, including S1 types of desirable outputs Y g = y 1 g , y 2 g , , y n g and S2 types of undesirable outputs b = y 1 b , y 2 b , , y n b .The target parameter λ i i = 1, 2 , , n , representing the weight vector of cross-sectional observations, is set to solve for each specific DMU. The SBM model is constructed as shown in Equation (A2). Here, ρ is the efficiency score, and the current DMU is considered efficient if and only if ρ = 1. Finally, the ultimate green energy efficiency is derived based on Equation (A3).
m i n ρ = 1 + 1 S 1 + S 2 i = 1 S 1 S i g y i k g + i = 1 S 2 S i b y i k b 1 1 m i = 1 m S i x i k s . t . X λ + S = x k Y g λ S g = y k g Y b λ S b = y k b λ 0 , S 0 , S g 0 , S b 0
M L t t + 1 = 1 + D 0 t x t , y t , b t ; y t , b t 1 + D 0 t x t + 1 , y t + 1 , b t + 1 ; y t + 1 , b t + 1 × 1 + D 0 t + 1 x t , y t , b t ; y t , b t 1 + D 0 t + 1 x t + 1 , y t + 1 , b t + 1 ; y t + 1 , b t + 1
In Equation (A3), t and t + 1 represent periods; denotes the Malmquist–Luenberger index from period t to period t + 1; D 0 t and D 0 t + 1 represent the distance functions under the technological frontiers in periods t and t + 1, respectively. The subscript 0 denotes the proportional parameter of input contraction and output expansion in the directional distance function. x t , y t , b t correspond to the input vector, desirable output variable, and undesirable output variable in period t, respectively; ( y t , b t ) represents the direction vector, indicating the direction of “expansion of desirable outputs and contraction of undesirable outputs”. The same logic applies to period t + 1.

Appendix C

This paper adopts a neural network model to fit the relationship between artificial intelligence and urban green energy efficiency. The specific construction process and adjustment process of the neural network model are as follows:
(1). Data loading and preprocessing The dataset is loaded from Excel, where “eff” is defined as the target variable, “Robot” as the independent variable (the independent variable in the test set is also “Robot”), “Fin”, “Finance”, “Science”, “structure”, and “Passenger” as control variables, and “Location conditions”, “City centre”, “Coastal and inland division”, “Transportation facilities”, and “City sizes” as geographical features (dummy variables). The str.strip and dropna functions from the third-party library pandas are used to clean the column names of the data table and remove missing values. One-hot encoding is applied to the categorical variables in the data table, and then the column names of the encoded geographical features are obtained to form input features together with the independent variables and control variables. Z-score standardization is used for continuous variables (independent variables and control variables).
(2). Construction and training of the main model Considering the time series characteristics of the panel data used in the study, a time-stratified splitting method is adopted, with data from 2010 to 2019 as the training set and data from 2020 to 2022 as the test set. Secondly, a fully connected neural network is constructed. The first layer is a dense layer, which takes the input feature list, has 128 output channels, uses the ReLu activation function, and is responsible for feature extraction and nonlinear transformation. The second layer is a Dropout layer, which randomly drops 30% of neurons (during training) to prevent overfitting. The third layer is a dense layer with 64 output channels and the ReLu activation function, used for deep feature extraction. The fourth layer is a dense layer with 32 output channels and the ReLu activation function, which further compresses and extracts deep features. The fifth layer is a dense layer with 1 output channel, outputting continuous values (for regression). The entire training process consists of 200 epochs with a training batch size of 32, using the Adam adaptive learning rate optimization, with the mean squared error (MSE) as the loss function and the mean absolute error (MAE) as the monitoring indicator. The EarlyStopping mechanism is employed to terminate the training early when the validation loss does not decrease for 10 epochs. All values are converted to the float32 numeric type using pandas.to_numeric() and DataFrame.apply(). If a value cannot be converted to a number, it is set to NaN. Then numpy.isnan() is used to detect missing values, and rows containing NaN are deleted. Figure A1 shows the training loss curve, where “Training Loss” represents the training loss and “Validation Loss” represents the validation set loss (an important indicator for evaluating the model’s performance on data not involved in training). Both losses use the same mean squared error (MSE), and the monitoring indicator is the mean absolute error (MAE). It can be seen from the figure that both the Train Loss and Val Loss decrease, indicating that the model learns well.
Figure A1. Training Loss Curve.
Figure A1. Training Loss Curve.
Sustainability 17 07205 g0a1
(3) Robustness Test Load the test dataset (independent variables, control variables, and dummy variables (encoded geographical features) of the test set), and convert all values to the float32 numeric type using pandas.to_numeric() and DataFrame.apply(). Input the test dataset into the main model, calculate and output the MSE of the robustness test, so as to evaluate the performance of the main model on the test set.
(4) Geographical Heterogeneity Analysis Iterate through each geographical feature, obtain all one-hot encoded columns of the current geographical feature, select one geographical feature to establish a subgroup, screen the data of the current subgroup, check whether the length of the sample size is less than 30, divide the training/test dataset, clean NaN values, conduct a secondary check on the sample size, construct a subgroup model with the same structure as the main model, train and evaluate the subgroup model, visualize the results (display scatter plots of actual values and predicted values, and relationship diagrams between Robot and eff within the subgroup), and save the results and model parameters at the same time.
(5) Visualization Screen out the data from 2010 to 2022, draw a dynamic relationship diagram, calculate the linear regression parameters of Robot and eff for each year, and create the animated image robot_eff_2010_2022.gif. Then, compare the impacts among cities, calculate the slope of each city, and create a bar chart for city comparison. Finally, conduct a nonlinear relationship analysis (LOWESS), and draw scatter plots and LOWESS fitting curves.

Appendix D

To verify the spatial spillover effect of the application of artificial intelligence technology, this paper constructs a spatial economic and geographical nested matrix, and the specific construction method is as follows:
W i j = f ( G D P i , G D P j , d i j )
In Equation (A4), the element in the i-th row and j-th column represents the interaction intensity between regions i and j. G D P i and G D P j denote the economic scales of regions i and j, respectively, while d i j represents the geographical distance between regions i and j. This study uses the centroid distance. Additionally, the diagonal elements of the d i j matrix are set to zero to reflect zero distance, and an exponential decay transformation is applied to the distance matrix to reflect the distance decay effect.
Next, the economic difference matrix is constructed as follows:
E i j 0, 1 , and a larger value indicates a more significant economic difference.
Then, the nested matrix is constructed as Equation (A5):
W i j = E i j × e d i j
Finally, to ensure that the sum of weights in each row is 1 and to avoid numerical overflow, the nested matrix is standardized as Equation (A6):
W i j * = W i j j = 1 n W i j
After normalization, the diagonal elements are set to W i i * = 1 , ensuring that the matrix is a diagonal matrix, meaning the weight of a region to itself is 0.

References

  1. Li, J.; Zheng, W.; Guo, X. Detecting multi-modal GAI-manipulated tourism review. Tour. Manag. 2025, 111, 105220. [Google Scholar] [CrossRef]
  2. Saura, J.R.; Buzinskiene, R. Behavioral economics, artificial intelligence and entrepreneurship: An updated framework for management. Int. Entrep. Manag. J. 2025, 21, 67. [Google Scholar] [CrossRef]
  3. Dwivedi, Y.K. Generative Artificial Intelligence (GenAI) in entrepreneurial education and practice: Emerging insights, the GAIN Framework, and research agenda. Int. Entrep. Manag. J. 2025, 21, 82. [Google Scholar] [CrossRef]
  4. Haque, A.K.M.B.; Islam, N.; Mikalef, P. To Explain or Not To Explain: An Empirical Investigation of AI-based Recommendations on Social Media Platforms. Electron. Mark. 2025, 35, 2. [Google Scholar] [CrossRef]
  5. Helmer, J.; Hawa, J.; Plewa, C. Digital technology as market shaper: A typology of digital technology roles for shaping markets. Electron. Mark. 2025, 35, 6. [Google Scholar] [CrossRef]
  6. Jung, Y.; Lee, S.-H.; Shin, J.K. Automation & the Future of Work: When Artificial Intelligence Meets Schumpeterian Innovators. Korean Econ. Rev. 2025, 41, 5–41. [Google Scholar] [CrossRef]
  7. Lin, M.; Lin, T.; Jones, L.; Liu, X.; Xing, L.; Sui, J.; Zhang, J.; Ye, H.; Liu, Y.; Zhang, G.; et al. Quantitatively Assessing Ecological Stress of Urbanization on Natural Ecosystems by Using a Landscape-Adjacency Index. Remote Sens. 2021, 13, 1352. [Google Scholar] [CrossRef]
  8. Wang, L.; Chen, Y. Artificial intelligence and corporate investment efficiency: Evidence from China. Electrom. Mark. 2025, 68, 101314. [Google Scholar] [CrossRef]
  9. Luo, J.; Cao, Q.; Zhang, S.; Gu, D. Generative AI usage among investor types: The role of personality and perceptions. Financ. Res. Lett. 2025, 82, 107604. [Google Scholar] [CrossRef]
  10. Qi, Y.; Su, H. Can artificial intelligence mitigate corporate fraud? Exploring the influence of institutional cross-holdings and financial misallocation. Pac.-Basin Financ. J. 2025, 92, 102822. [Google Scholar] [CrossRef]
  11. Jarotschkin, V.; Soykoth, M.W.; Chaker, N.N. Artificial intelligence in sales research: Identifying emergent themes and looking forward. J. Bus. Res. 2025, 198, 115383. [Google Scholar] [CrossRef]
  12. Ren, B.; Qiu, Z.; Liu, B. Supply chain decarbonisation effects of artificial Intelligence: Evidence from China. Int. Rev. Econ. Financ. 2025, 101, 104198. [Google Scholar] [CrossRef]
  13. Tian, H.; Zhao, L.; Li, Y.; Wang, W. Can enterprise green technology innovation performance achieve “corner overtaking” by using artificial intelligence?-Evidence from Chinese manufacturing enterprises. Technol. Forecast. Soc. Change 2023, 194, 122732. [Google Scholar] [CrossRef]
  14. Xu, Y.; Dai, H.; Yan, W. Identity Disclosure and Anthropomorphism in Voice Chatbot Design: A Field Experiment. Manag. Sci. 2024. [Google Scholar] [CrossRef]
  15. Bai, S.; He, H.; Han, C.; Yang, M.; Shang, W.-L.; Fan, W. What is the focus of energy supply chain relationship management during geopolitical risks? Evidence from the stock market based on transaction cost economics. Energy Econ. 2025, 148, 108629. [Google Scholar] [CrossRef]
  16. Majrashi, K. Employees’ perceptions of the fairness of AI-based performance prediction features. Cogent Bus. Manag. 2025, 12, 2456111. [Google Scholar] [CrossRef]
  17. Hui, X.; Reshef, O.; Zhou, L. The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market. Organ. Sci. 2024, 35, 1977–1989. [Google Scholar] [CrossRef]
  18. Mariani, M.M.; Machado, I.; Magrelli, V.; Dwivedi, Y.K. Artificial intelligence in innovation research: A systematic review, conceptual framework, and future research directions. Technovation 2023, 122, 102623. [Google Scholar] [CrossRef]
  19. Glikson, E.; Woolley, A.W. Human Trust in Artificial Intelligence: Review of Empirical Research. Acad. Manag. Ann. 2020, 14, 627–660. [Google Scholar] [CrossRef]
  20. Dayioglu, M.A.; Türker, U. Digital Transformation for Sustainable Future—Agriculture 4.0: A Review. J. Agric. Sci. 2021, 27, 373–399. [Google Scholar] [CrossRef]
  21. Lin, J.; Zeng, Y.; Wu, S.; Luo, X. How does artificial intelligence affect the environmental performance of organizations? The role of green innovation and green culture. Inf. Manag. 2024, 61, 103924. [Google Scholar] [CrossRef]
  22. Long, Y.; Liu, L.; Yang, B. The effects of enterprise digital transformation on low-carbon urban development: Empirical evidence from China. Technol. Forecast. Soc. Change 2024, 201, 123259. [Google Scholar] [CrossRef]
  23. Wang, Z.; Zhang, T.; Ren, X.; Shi, Y. AI adoption rate and corporate green innovation efficiency: Evidence from Chinese energy companies. Energy Econ. 2024, 132, 107499. [Google Scholar] [CrossRef]
  24. Liu, C.; Tang, C.; Liu, Y. Does the transformation of energy structure promote green technological innovation? A quasi-natural experiment based on new energy demonstration city construction. Geosci. Front. 2024, 15, 101615. [Google Scholar] [CrossRef]
  25. Tan, L.; Yang, Z.; Irfan, M.; Ding, C.J.; Hu, M.; Hu, J. Toward low-carbon sustainable development: Exploring the impact of digital economy development and industrial restructuring. Bus. Strategy Environ. 2024, 33, 2159–2172. [Google Scholar] [CrossRef]
  26. Du, K.; Cheng, Y.; Yao, X. Environmental regulation, green technology innovation, and industrial structure upgrading: The road to the green transformation of Chinese cities. Energy Econ. 2021, 98, 105247. [Google Scholar] [CrossRef]
  27. Zhang, J.S.; Jiang, Y.C.; Li, X.; Huo, M.Y.; Luo, H.; Yin, S. An adaptive remaining useful life prediction approach for single battery with unlabeled small sample data and parameter uncertainty. Reliab. Eng. Syst. Saf. 2022, 222, 108357. [Google Scholar] [CrossRef]
  28. Wang, B.; Wang, J.; Dong, K.; Nepal, R. How does artificial intelligence affect high-quality energy development? Achieving a clean energy transition society. Energy Policy 2024, 186, 114010. [Google Scholar] [CrossRef]
  29. Zhou, W.; Zhang, Y.; Li, X. Artificial intelligence, green technological progress, energy conservation, and carbon emission reduction in China: An examination based on dynamic spatial Durbin modeling. J. Clean. Prod. 2024, 446, 141142. [Google Scholar] [CrossRef]
  30. Dong, X.; Zhou, N.; Zhao, X.; Yang, S. The impact of artificial intelligence on corporate green innovation: Can “increasing quantity” and “improving quality” go hand in hand? J. Environ. Manag. 2025, 376, 124439. [Google Scholar] [CrossRef]
  31. Zhao, S.; Zhang, L.; Peng, L.; Zhou, H.; Hu, F. Enterprise pollution reduction through digital transformation? Evidence from Chinese manufacturing enterprises. Technol. Soc. 2024, 77, 102520. [Google Scholar] [CrossRef]
  32. Kotecha, N.; Chanona, A.d.R. Leveraging graph neural networks and multi-agent reinforcement learning for inventory control in supply chains. Comput. Chem. Eng. 2025, 199, 109111. [Google Scholar] [CrossRef]
  33. Johri, A.; Singh, R.K.; Kushwaha, B.P.; Alhumoudi, H.; Alakkas, A.; Khoja, M. Leveraging open innovation for e-commerce success: The contingent role of accounting information systems and artificial intelligence. J. Innov. Knowl. 2025, 10, 100737. [Google Scholar] [CrossRef]
  34. Grewal, D.; Okazaki, S.; Guha, A.; Liu-Thompkins, Y. Generative AI: Delivering on the Promises and Understanding the Perils. J. Public Policy Mark. 2025, 44, 299–308. [Google Scholar] [CrossRef]
  35. Khatib, A.M.G.A. Beyond linearity: A critical review of the finance-growth nexus. Cogent Econ. Financ. 2025, 13, 2514690. [Google Scholar] [CrossRef]
  36. Ospina-Holguin, J.H.; Padilla-Ospina, A.M. Reinforcement learning meets technical analysis: Combining moving average rules for optimal alpha. Cogent Econ. Financ. 2025, 13, 2490818. [Google Scholar] [CrossRef]
  37. Chauhan, J.K.; Ahmed, T.; Sinha, A. A novel deep learning model for stock market prediction using a sentiment analysis system from authoritative financial website’s data. Connect. Sci. 2025, 37, 2455070. [Google Scholar] [CrossRef]
  38. Chenavaz, R.Y.; Dimitrov, S. Artificial intelligence and dynamic pricing: A systematic literature review. J. Appl. Econ. 2025, 28, 2466140. [Google Scholar] [CrossRef]
  39. Gumbo, L.; Chude-Okonkwo, U.A.K. Regulatory sandbox as a frontier for innovation and sustainability: A systematic review. Cogent Bus. Manag. 2025, 12, 2510555. [Google Scholar] [CrossRef]
  40. Dong, Y. Implementation of artificial intelligence for brand equity. Cogent Bus. Manag. 2025, 12, 2471538. [Google Scholar] [CrossRef]
  41. Kusmaryanto, S.; Santoso, C.B. A scoping review of middle managers in the digital transformation era in public sector organizations: Are they still needed? Cogent Bus. Manag. 2025, 12, 2461734. [Google Scholar] [CrossRef]
  42. Rammer, C.; Fernandez, G.P.; Czarnitzki, D. Artificial intelligence and industrial innovation: Evidence from German firm-level data. Res. Policy 2022, 51, 104555. [Google Scholar] [CrossRef]
  43. Ooi, K.-B.; Tan, G.W.-H.; Al-Emran, M.; Al-Sharafi, M.A.; Capatina, A.; Chakraborty, A.; Dwivedi, Y.K.; Huang, T.-L.; Kar, A.K.; Lee, V.-H.; et al. The Potential of Generative Artificial Intelligence Across Disciplines: Perspectives and Future Directions. J. Comput. Inf. Syst. 2025, 65, 76–107. [Google Scholar] [CrossRef]
  44. Mikalef, P.; Gupta, M. Artificial intelligence capability: Conceptualization, measurement calibration, and empirical study on its impact on organizational creativity and firm performance. Inf. Manag. 2021, 58, 103434. [Google Scholar] [CrossRef]
  45. Yang, C.-H. How Artificial Intelligence Technology Affects Productivity and Employment: Firm-level Evidence from Taiwan. Res. Policy 2022, 51, 104536. [Google Scholar] [CrossRef]
  46. Belhadi, A.; Mani, V.; Kamble, S.S.; Khan, S.A.R.; Verma, S. Artificial intelligence-driven innovation for enhancing supply chain resilience and performance under the effect of supply chain dynamism: An empirical investigation. Ann. Oper. Res. 2024, 333, 627–652. [Google Scholar] [CrossRef]
  47. Janiesch, C.; Zschech, P.; Heinrich, K. Machine learning and deep learning. Electron. Mark. 2021, 31, 685–695. [Google Scholar] [CrossRef]
  48. Kong, H.; Yin, Z.; Baruch, Y.; Yuan, Y. The impact of trust in AI on career sustainability: The role of employee-AI collaboration and protean career orientation. J. Vocat. Behav. 2023, 146, 103928. [Google Scholar] [CrossRef]
  49. Shao, Y.; Huang, C.; Song, Y.; Wang, M.; Song, Y.H.; Shao, R. Using Augmentation-Based AI Tool at Work: A Daily Investigation of Learning-Based Benefit and Challenge. J. Manag. 2024. [Google Scholar] [CrossRef]
  50. Noy, S.; Zhang, W. Experimental evidence on the productivity effects of generative artificial intelligence. Science 2023, 381, 187–192. [Google Scholar] [CrossRef] [PubMed]
  51. Lee, C.-C.; Lee, C.-C. How does green finance affect green total factor productivity? Evidence from China. Energy Econ. 2022, 107, 105863. [Google Scholar] [CrossRef]
  52. Liu, Z.; Deng, Z.; He, G.; Wang, H.; Zhang, X.; Lin, J.; Qi, Y.; Liang, X. Challenges and opportunities for carbon neutrality in China. Nat. Rev. Earth Environ. 2022, 3, 141–155. [Google Scholar] [CrossRef]
  53. Zhao, X.; Benkraiem, R.; Abedin, M.Z.; Zhou, S. The charm of green finance: Can green finance reduce corporate carbon emissions? Energy Econ. 2024, 134, 107574. [Google Scholar] [CrossRef]
  54. Li, Z.; Wang, J. The Dynamic Impact of Digital Economy on Carbon Emission Reduction: Evidence City-level Empirical Data in China. J. Clean. Prod. 2022, 351, 131570. [Google Scholar] [CrossRef]
  55. Lyu, Y.; Wang, W.; Wu, Y.; Zhang, J. How does digital economy affect green total factor productivity? Evidence from China. Sci. Total Environ. 2023, 857, 159428. [Google Scholar] [CrossRef]
  56. Zhang, D. Green financial system regulation shock and greenwashing behaviors: Evidence from Chinese firms. Energy Econ. 2022, 111, 106064. [Google Scholar] [CrossRef]
  57. Wang, J.; Liu, Y.; Wang, W.; Wu, H. How does digital transformation drive green total factor productivity? Evidence from Chinese listed enterprises. J. Clean. Prod. 2023, 406, 136954. [Google Scholar] [CrossRef]
  58. Liu, Q.; Chen, Z.; Xiao, B. The common prosperity effect of industrial robot application: From the perspective of the urban-rural income gap. Econ. Theory Econ. Manag. 2024, 44, 90–106. [Google Scholar]
  59. Wang, Y.; Dong, W. The evolution of China’s labor market structure: From the perspective of task-biased technological progress. Soc. Sci. China 2023, 11, 45–64+205. [Google Scholar]
  60. Shi, D. Regional differences and energy-saving potential of China’s energy efficiency. China Ind. Econ. 2006, 10, 49–58. [Google Scholar] [CrossRef]
  61. Du, Y.H.; Chen, W.L.; Dai, X.J.; Li, J. Research on the Impact of Carbon Emission Trading Policies on Urban Green Economic Efficiency-Based on Dual Macro and Micro Perspectives. Sustainability 2025, 17, 2670. [Google Scholar] [CrossRef]
  62. Shao, S.; Meiting, F.; Lili, Y. Economic Structure Adjustment, Green Technological Progress and China’s Low-Carbon Transition Development——An Empirical Study from the Perspective of Overall Technological Frontier and Spatial Spillover Effect. Manag. World 2022, 38, 46–69+4–10. [Google Scholar] [CrossRef]
  63. Li, J.; Min, Y.; Wang, X. Can the China-Europe Railway Network Promote Industrial Structure Upgrading?—Evidence from a Quasi-Natural Experiment in 285 Chinese Cities. J. Ind. Econ. Res. 2021, 3, 69–83. [Google Scholar] [CrossRef]
  64. Sun, H.; Zhu, J. Can the Establishment of Free Trade Zones Enhance the Resilience of China’s Industrial Chain?—Testing the Mediating Mechanism of Multiple Innovation Element Agglomeration. Mod. Econ. Res. 2023, 11, 72–84. [Google Scholar] [CrossRef]
  65. Yan, T.; Ma, L. The Carbon Emission Reduction Effect of Green Finance—An Empirical Analysis Based on Provincial Panel Data in China. Financ. Theory Teach. 2024, 42, 19–25. [Google Scholar] [CrossRef]
  66. Shao, S.; Ge, L.; Zhu, J. How to Achieve Harmonious Coexistence between Humanity and Nature: Environmental Regulation and Environmental Welfare Performance from a Geographic Perspective. Manag. World 2024, 40, 119–146. [Google Scholar]
  67. Chen, N. The Impact of Value-Added Tax Reform on Local Government Public Finance Revenue and Countermeasures. Financ. Econ. 2022, 20, 140–142. [Google Scholar] [CrossRef]
  68. Wen, Z.L.; Ye, B.J. Mediation Effect Analysis: Methodology and Model Development. Adv. Psychol. Sci. 2014, 22, 731–745. [Google Scholar] [CrossRef]
  69. Kavousighahfarokhi, A.; Hannan, M.A.; Abu, S.M.; Ker, P.J.; Wong, R.T.K.; Jang, G. Grid-integrated electric vehicle charging station technologies and data dissemination through internet of energy use. Renew. Sustain. Energy Rev. 2025, 220, 115904. [Google Scholar] [CrossRef]
  70. Zhang, Q.; Li, Y.; Zhang, Q.; Ma, F.; Lü, X. Application of deep dehumidification technology in low-humidity industry: A review. Renew. Sustain. Energy Rev. 2024, 193, 114278. [Google Scholar] [CrossRef]
  71. Berkessa, Z.A.; Läärä, E.; Waldmann, P. A Review of Causal Methods for High-Dimensional Data. IEEE Access 2025, 13, 11892–11917. [Google Scholar] [CrossRef]
  72. Wang, L.; Shao, J. The energy saving effects of digital infrastructure construction: Empirical evidence from Chinese industry. Energy 2024, 294, 130778. [Google Scholar] [CrossRef]
  73. Ai, H.; Tan, X.; Mangla, S.K.; Emrouznejad, A.; Liu, F.; Song, M. Renewable energy transition and sustainable development: Evidence from China. Energy Econ. 2025, 143, 108232. [Google Scholar] [CrossRef]
  74. Jiang, R.A.; Yang, S.H.; Wen, J. High-Speed Rail Opening and High-Quality Economic Development—Mechanism and Effect. South China J. Econ. Res. 2023, 7, 70–89. [Google Scholar] [CrossRef]
  75. Zhao, T.; Zhang, Z.; Liang, S.K. The Digital Economy, Entrepreneurial Activity, and High-Quality Development—Empirical Evidence from Chinese Cities. Manag. World 2020, 36, 65–76. [Google Scholar] [CrossRef]
  76. Luo, F.Y.; Yang, S.Q.; Li, S.Q. The Impact of the Digital Economy on the Integration of the Factor Market. Ind. Econ. Rev. 2023, 3, 31–53. [Google Scholar] [CrossRef]
  77. Medium and Long-Term Railway Network Plan. 2016. Available online: https://www.gov.cn/xinwen/2016-07/20/content_5093165.htm (accessed on 7 July 2025).
  78. Zhao, Z.; Zhao, Y.; Shi, X.; Zheng, L.; Fan, S.; Zuo, S. Green innovation and carbon emission performance: The role of digital economy. Energy Policy 2024, 195, 114344. [Google Scholar] [CrossRef]
  79. Jiakui, C.; Abbas, J.; Najam, H.; Liu, J.; Abbas, J. Green technological innovation, green finance, and financial development and their role in green total factor productivity: Empirical insights from China. J. Clean. Prod. 2023, 382, 135131. [Google Scholar] [CrossRef]
  80. Irfan, M.; Razzaq, A.; Sharif, A.; Yang, X. Influence mechanism between green finance and green innovation: Exploring regional policy intervention effects in China. Technol. Forecast. Soc. Change 2022, 182, 121882. [Google Scholar] [CrossRef]
  81. Nepal, R.; Zhao, X.; Dong, K.; Wang, J.; Sharif, A. Can artificial intelligence technology innovation boost energy resilience? The role of green finance. Energy Econ. 2025, 142, 108159. [Google Scholar] [CrossRef]
  82. Shang, Y.; Raza, S.A.; Huo, Z.; Shahzad, U.; Zhao, X. Does enterprise digital transformation contribute to the carbon emission reduction? Micro-level evidence from China. Int. Rev. Econ. Financ. 2023, 86, 1–13. [Google Scholar] [CrossRef]
  83. Hao, X.; Li, Y.; Ren, S.; Wu, H.; Hao, Y. The role of digitalization on green economic growth: Does industrial structure optimization and green innovation matter? J. Environ. Manag. 2023, 325, 116504. [Google Scholar] [CrossRef] [PubMed]
  84. Yang, X.; Wang, H.; Yan, T.; Cao, M.; Han, Y.; Pan, Y.; Feng, Y. The road to inclusive green growth in China: Exploring the impact of digital-real economy integration on carbon emission efficiency. J. Environ. Manag. 2024, 370, 122989. [Google Scholar] [CrossRef] [PubMed]
  85. Wu, L.; Zhu, C.; Wang, G. The impact of green innovation resilience on energy efficiency: A perspective based on the development of the digital economy. J. Environ. Manag. 2024, 355, 120424. [Google Scholar] [CrossRef] [PubMed]
  86. Liu, Y.; Song, P. Digital Transformation and Green Innovation of Energy Enterprises. Sustainability 2023, 15, 7703. [Google Scholar] [CrossRef]
  87. Li, H.; Su, Y.; Ding, C.J.; Tian, G.G.; Wu, Z. Unveiling the green innovation paradox: Exploring the impact of carbon emission reduction on corporate green technology innovation. Technol. Forecast. Soc. Change 2024, 207, 123562. [Google Scholar] [CrossRef]
  88. Wang, W.; Wang, J.; Wu, H. The impact of energy-consuming rights trading on green total factor productivity in the context of digital economy: Evidence from listed firms in China. Energy Econ. 2024, 131, 107342. [Google Scholar] [CrossRef]
  89. Ji, C.-Y.; Tan, Z.-K.; Chen, B.-J.; Zhou, D.-C.; Qian, W.-Y. The impact of environmental policies on renewable energy investment decisions in the power supply chain. Energy Policy 2024, 186, 113987. [Google Scholar] [CrossRef]
Figure 1. Research framework diagram.
Figure 1. Research framework diagram.
Sustainability 17 07205 g001
Figure 2. Simulation Results of the Neural Network Model (2010–2015).
Figure 2. Simulation Results of the Neural Network Model (2010–2015).
Sustainability 17 07205 g002
Figure 3. Simulation Results of the Neural Network Model (2016–2021).
Figure 3. Simulation Results of the Neural Network Model (2016–2021).
Sustainability 17 07205 g003
Figure 4. Simulation Results of the Neural Network Model (2022).
Figure 4. Simulation Results of the Neural Network Model (2022).
Sustainability 17 07205 g004
Figure 5. Dimensionality Reduction of Model Variables.
Figure 5. Dimensionality Reduction of Model Variables.
Sustainability 17 07205 g005
Figure 6. Regional Heterogeneity.
Figure 6. Regional Heterogeneity.
Sustainability 17 07205 g006
Figure 7. Heterogeneity by City Size.
Figure 7. Heterogeneity by City Size.
Sustainability 17 07205 g007
Figure 8. Heterogeneity Analysis of Whether a City is a Central City.
Figure 8. Heterogeneity Analysis of Whether a City is a Central City.
Sustainability 17 07205 g008
Figure 9. Heterogeneity Analysis of Whether a City is a Coastal City.
Figure 9. Heterogeneity Analysis of Whether a City is a Coastal City.
Sustainability 17 07205 g009
Figure 10. Heterogeneity Analysis of Whether a City is a Transportation Hub.
Figure 10. Heterogeneity Analysis of Whether a City is a Transportation Hub.
Sustainability 17 07205 g010
Figure 11. Slope Values for Each City.
Figure 11. Slope Values for Each City.
Sustainability 17 07205 g011
Table 1. Indicators for the Calculation of Green Energy Efficiency.
Table 1. Indicators for the Calculation of Green Energy Efficiency.
Indicators for the Calculation of Green Energy Efficiency
Primary IndicatorSecondary IndicatorThird IndicatorUnit
Input IndicatorsLabor InputYear-end Employment in Municipal DistrictsTen Thousand People
Capital InputCapital Stock Calculated by Perpetual Inventory MethodTen Thousand Yuan
Energy InputNighttime Light Data Simulated Measurement, then Decomposing Provincial Energy Consumption Data by Light Data Value to Prefecture-level Cities, and Dividing by Regional Gross ProductTen Thousand Tons of Coal Equivalent
Desired Output IndicatorRegional Gross ProductActual Regional Gross ProductTen Thousand Yuan
Undesired Output IndicatorUrban WasteIndustrial SO2 EmissionsTons
Industrial Wastewater DischargeTen Thousand Tons
Industrial Smoke and Dust EmissionsTons
Table 2. Measurement of Control Variables.
Table 2. Measurement of Control Variables.
VariableSymbolDefinition
Fiscal Investment IntensityFinThe share of urban capital expenditure within the overall expenditure of the public sector
Financial Development LevelFinanceThe ratio of the balance of deposits and loans held by financial institutions at year-end to the regional gross output
Urban Industrial StructureStructureThe share of the service sector’s production value in the total production value of the city’s industries
Urban Science and Technology ExpenditureScienceThe ratio of urban R&D expenditure to the overall expenditures of the public sector
Population Mobility StatusPassengerThe logarithm of the total number of passengers (in units of ten million)
Table 3. Baseline regression results.
Table 3. Baseline regression results.
VariablesEff
(1)(2)(3)(4)(5)(6)
Robot0.021 ***0.020 ***0.020 ***0.019 ***0.026 ***0.026 ***
(0.001)(0.002)(0.002)(0.002)(0.003)(0.003)
Fin 0.0050.0040.005 *0.007 **0.007 **
(0.003)(0.003)(0.003)(0.003)(0.003)
Finance −0.001−0.001−0.001−0.001
(0.001)(0.001)(0.001)(0.001)
Science 0.475 ***0.475 ***0.499 ***
(0.177)(0.177)(0.177)
structure −0.152 ***−0.152 ***
(0.046)(0.046)
Passenger −0.009 ***
(0.003)
CityYesYesYesYesYesYes
YearYesYesYesYesYesYes
Constant0.229 ***0.226 ***0.227 ***0.223 ***0.245 ***0.258 ***
(0.008)(0.008)(0.008)(0.008)(0.011)(0.012)
N352335233523352335233523
R2-adj0.6920.7000.7000.7370.7690.790
Standard errors in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.1. The asterisks in the following tables have the same meaning.
Table 4. Robustness Tests (1).
Table 4. Robustness Tests (1).
VariablesOutlier Treatment
Eff
Truncated RegressionCensored RegressionExclude the Year 2020
(1)(2)(3)
Robot0.025 ***0.025 ***0.025 ***
(0.003)(0.003)(0.003)
Control variableYesYesYes
CityYesYesYes
YearYesYesYes
Constant0.255 ***0.255 ***0.258 ***
(0.011)(0.012)(0.012)
N352331703242
R2-adj0.8590.8340.881
Note: Standard errors in parentheses, *** p < 0.01.
Table 5. Robustness Tests (2).
Table 5. Robustness Tests (2).
VariablesReplace the Dependent VariableReplace the Independent VariableBoth Are Replaced
Eff-CCREffEff-CCR
(1)(2)(3)
Robot0.026 ***
(0.004)
AI 0.146 ***0.133 ***
(0.017)(0.021)
Control variableYesYesYes
CityYesYesYes
YearYesYesYes
Constant0.533 ***0.120 ***0.406 ***
(0.015)(0.018)(0.023)
N352335233523
R2-adj0.6170.7950.671
Note: Standard errors in parentheses, *** p < 0.01.
Table 6. Robustness Tests (3).
Table 6. Robustness Tests (3).
VariablesSmart City PilotKey Air Control Zone PilotMeanwhile Exclude
EffEffEff
(1)(2)(3)
Robot0.029 ***0.027 ***0.029 ***
(0.003)(0.003)(0.003)
Smart_pilot−0.022 *** −0.022 ***
(0.007) (0.007)
Key_pilot −0.002−0.001
(0.006)(0.006)
Control variableYesYesYes
CityYesYesYes
YearYesYesYes
Constant0.255 ***0.258 ***0.255 ***
(0.012)(0.012)(0.012)
N352335233523
R2-adj0.8220.7950.821
Note: Standard errors in parentheses, *** p < 0.01.
Table 7. Endogeneity Analysis (1).
Table 7. Endogeneity Analysis (1).
VariablesIV1IV2
(1)(2)(3)(4)
RobotEffRobotEff
AIPC0.221 ***
(0.017)
AIDE 0.095 ***
(0.020)
Robot 0.255 *** 0.657 ***
(0.024) (0.142)
Control variableYesYesYesYes
CityYesYesYesYes
YearYesYesYesYes
Constant3.488 ***0.990 ***2.119 ***2.355 ***
(0.015)(0.093)(0.170)(0.489)
N3523352335233523
R2-adj0.7160.7790.7020.763
Chi-squared(1)17.98 37.85
(0.133)(0.238)
Kleibergen-Paap rk
Wald F Value
22.15 50.46
(16.38) (16.38)
Note: The Cragg–Donald statistic is presented with the Stock-Yogo critical values in parentheses, which are used to determine whether the instrumental variables are weak. If the Cragg–Donald statistic is greater than the corresponding Stock-Yogo critical value, the instrumental variables can be considered not weak. The Robust score chi2 statistic is shown with the p-value in parentheses, which is used to test the exogeneity of the variables. If the p-value is less than the significance level, the null hypothesis is rejected, indicating that the original explanatory variable is endogenous. Additionally, the values in parentheses represent robust standard errors, and the stars indicate the level of statistical significance. Standard errors in parentheses, *** p < 0.01
Table 8. Endogeneity Analysis (2).
Table 8. Endogeneity Analysis (2).
VariablesDifference GMMSystem GMMLag by One PeriodLag by Two Periods
EffEffEffEff
(1)(2)(3)(4)
L.posttreat 0.018 ***
(0.003)
L2.posttreat 0.018 ***
(0.004)
L.eff0.262 ***0.790 ***
(0.002)(0.115)
Robot0.008 ***0.005 ***
(0.000)(0.000)
Control variableYesYesYesYes
CityYesYesYesYes
YearYesYesYesYes
Constant0.042 ***0.249 ***0.056 ***0.065 ***
(0.012)(0.017)(0.015)(0.017)
N2981325232522981
AR(1)0.0000.000--
AR(2)0.1330.410--
Hansen0.3150.124--
R2-adj--0.7410.763
Note: Standard errors in parentheses, *** p < 0.01.
Table 9. Endogeneity Analysis (3).
Table 9. Endogeneity Analysis (3).
VariablesPlease Translate the Following Text into English
Eff
(1)
Pc_Robot0.027 ***
(0.002)
Pc_Control variableYes
CityYes
YearYes
Constant0.172 ***
(0.008)
N3523
R2-adj0.803
Note: Standard errors in parentheses, *** p < 0.01.
Table 10. Heterogeneity Analysis (1).
Table 10. Heterogeneity Analysis (1).
VariablesRegional Location
Eff
EastCentralWestNortheast
(1)(2)(3)(4)
Robot0.025 ***0.017 ***0.014 ***−0.012 ***
(0.004)(0.006)(0.003)(0.002)
Control variableYesYesYesYes
CityYesYesYesYes
YearYesYesYesYes
Constant0.060−0.568 ***0.203 *0.465 ***
(0.162)(0.163)(0.105)(0.066)
N11059751040403
R2-adj0.6110.6230.6360.638
Note: Standard errors in parentheses, *** p < 0.01, * p < 0.1.
Table 11. Heterogeneity Analysis (2).
Table 11. Heterogeneity Analysis (2).
VariablesCity Size
Eff
MegacitiesLarge CitiesMedium-Small CitiesSmall Cities
(1)(2)(3)(4)
Robot0.0150.054 ***0.015 ***0.007 ***
(0.026)(0.013)(0.005)(0.003)
Control variableYesYesYesYes
CityYesYesYesYes
YearYesYesYesYes
Constant−2.464 **−0.400−0.670 ***−0.145 **
(1.170)(0.617)(0.160)(0.073)
N9118210012249
R2-adj0.5870.3260.4720.570
Note: Standard errors in parentheses, *** p < 0.01, ** p < 0.05.
Table 12. Heterogeneity Analysis (3).
Table 12. Heterogeneity Analysis (3).
VariablesWhether It Is a Central CityWhether It Is a Coastal CityWhether It Is a Transportation Hub
EffEffEff
NoYesNoYesNoYes
(1)(2)(3)(4)(5)(6)
Robot0.011 ***0.017 *0.009 ***0.021 ***0.011 ***0.018
(0.002)(0.010)(0.003)(0.004)(0.002)(0.015)
Control variableYesYesYesYesYesYes
CityYesYesYesYesYesYes
YearYesYesYesYesYesYes
Constant−0.158 **−1.753 ***−0.398 ***0.192−0.142 **−2.888 ***
(0.066)(0.408)(0.083)(0.131)(0.065)(0.625)
N3081442206714563290233
R2-adj0.6510.6190.5640.5470.6800.461
Note: Standard errors in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.1.
Table 13. Heterogeneity Analysis (4).
Table 13. Heterogeneity Analysis (4).
FeatureCityProvinceSlope
Largest slopeJinanShandong Province0.195
ShenzhenShandong Province0.193
HezeShandong Province0.184
LangfangHebei Province0.164
ZhengzhouHenan Province0.163
Smallest slopeMaomingGuangdong−0.103
ZiyangSichuan Province−0.121
SanyaHainan Province−0.129
LongnanGansu Province−0.138
DongguanGuangdong Province−0.165
Table 14. Mechanism Tests.
Table 14. Mechanism Tests.
VariablesIndustrial Chain ResilienceGreen FinanceEnvironmental Regulation Intensity
ICREffGreenFinanceEffRegulationEff
(1)(2)(3)(4)(5)(6)
Robot0.016 ***0.018 ***0.031 ***0.022 ***0.029 **0.029 ***
(0.001)(0.003)(0.001)(0.004)(0.011)(0.003)
ICR 0.527 ***
(0.046)
GreenFinance 0.137 **
(0.067)
Regulation 0.013 ***
(0.001)
Control variableYesYesYesYesYesYes
CityYesYesYesYesYesYes
YearYesYesYesYesYesYes
Constant−0.018 ***0.267 ***0.184 ***0.233 ***0.881 ***0.255 ***
(0.005)(0.011)(0.003)(0.017)(0.044)(0.013)
N352335233523352335233523
R2-adj0.6630.7220.6380.7810.7950.786
Note: Standard errors in parentheses, *** p < 0.01,** p < 0.05.
Table 15. Chain Mediation Effects.
Table 15. Chain Mediation Effects.
VariablesRegulation Intensity→Industrial Chain ResilienceGreen Finance→Industrial Chain Resilience
RegulationICREffGreen FinanceICREff
(1)(2)(3)(4)(5)(6)
Robot0.029 **0.016 ***0.018 ***0.031 ***0.011 ***0.018 ***
(0.011)(0.001)(0.003)(0.001)(0.001)(0.003)
ICR 0.527 *** 0.527 ***
(0.046) (0.046)
GreenFinance 0.168 ***
(0.027)
Control variableYesYesYesYesYesYes
CityYesYesYesYesYesYes
YearYesYesYesYesYesYes
Constant0.881 ***−0.016 ***0.267 ***0.184 ***−0.048 ***0.267 ***
(0.044)(0.005)(0.011)(0.003)(0.007)(0.011)
N352335233523352335233523
R2-adj0.7950.6640.7220.6380.7750.722
Bootstrap-Z2.672.85
(0.0000175~0.0002164)(0.0000209~0.0002389)
Note: Standard errors in parentheses, *** p < 0.01,** p < 0.05.
Table 16. Mechanism Synergies Effects.
Table 16. Mechanism Synergies Effects.
VariablesMechanism Synergy
Eff
(1)(2)(3)(4)
Robot0.028 ***0.021 ***0.018 ***0.021 ***
(0.003)(0.003)(0.003)(0.003)
Interaction10.022 **
(0.009)
Interaction2 0.506 ***
(0.049)
Interaction3 0.962 ***
(0.082)
Interaction5 1.019 ***
(0.093)
Control variableYesYesYesYes
CityYesYesYesYes
YearYesYesYesYes
Constant0.253 ***0.266 ***0.268 ***0.268 ***
(0.012)(0.012)(0.011)(0.012)
N3523352335233523
R2-adj0.6680.6310.6350.636
Note: Standard errors in parentheses, *** p < 0.01, ** p < 0.05.
Table 17. Test of Spatial Correlation.
Table 17. Test of Spatial Correlation.
YearRobotEff
Moran’IMoran’I
20100.10500.1250
20110.12800.1520
20120.15600.1860
20130.18200.2130
20140.20500.2350
20150.19800.2560
20160.22300.2280
20170.24500.2790
20180.26200.2980
20190.27800.3120
20200.28500.3250
20210.29200.3380
20220.30500.3520
Table 18. LM Test.
Table 18. LM Test.
Test IndicatorTest MethodStatistic Valuep-Value
LM_error testLM_error25.630.000
LM_error_robust12.790.000
LM_lag testLM_lag32.470.000
LM_lag_robust8.630.003
Table 19. Regression Results of the Spatial Durbin Model.
Table 19. Regression Results of the Spatial Durbin Model.
VariablesEff
DirectIndirectTotal
(1)(2)(3)
Robot0.023 ***0.008 ***0.031 ***
(0.001)(0.003)(0.004)
Spatial Autoregressive Coefficient0.374 ***--
(0.036)--
sigma2_e0.007 ***--
(0.000)--
Controlsyes--
cityyesyesyes
yearyesyesyes
N352335233523
R20.2110.2110.211
Note: Standard errors in parentheses, *** p < 0.01.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Du, Y.; Liu, T.; Shang, W.; Li, J. Research on the Impact of Artificial Intelligence on Urban Green Energy Efficiency: An Empirical Test Based on Neural Network Models. Sustainability 2025, 17, 7205. https://doi.org/10.3390/su17167205

AMA Style

Du Y, Liu T, Shang W, Li J. Research on the Impact of Artificial Intelligence on Urban Green Energy Efficiency: An Empirical Test Based on Neural Network Models. Sustainability. 2025; 17(16):7205. https://doi.org/10.3390/su17167205

Chicago/Turabian Style

Du, Yuanhe, Tianhang Liu, Wei Shang, and Jia Li. 2025. "Research on the Impact of Artificial Intelligence on Urban Green Energy Efficiency: An Empirical Test Based on Neural Network Models" Sustainability 17, no. 16: 7205. https://doi.org/10.3390/su17167205

APA Style

Du, Y., Liu, T., Shang, W., & Li, J. (2025). Research on the Impact of Artificial Intelligence on Urban Green Energy Efficiency: An Empirical Test Based on Neural Network Models. Sustainability, 17(16), 7205. https://doi.org/10.3390/su17167205

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop