Deep Learning-Based Study of Carbon Emissions Peak Pathways in Chinese Building Sector: Incorporating Legal and Policy Text Quantification

Abstract

The decarbonization process of the carbon emissions in the Chinese building sector exerts a profound impact on the achievement of the national goals of carbon peak and carbon neutrality. Currently, there is limited literature quantifying the impact of laws and policies on the achievement of carbon peak in the Chinese building sector and further utilizing deep learning technology to characterize the carbon emissions peak path under uncertainty in the Chinese building sector. To address this issue, a quantitative framework of legal and policy incentive intensity is constructed to capture both the immediate effects and the long-term evolution of laws and policies, and the index of legal and policy incentive intensity for carbon emissions in the building sector in China from 2010 to 2022 is calculated. Based on this, a dynamic scenario forecasting model for carbon emissions in the Chinese building sector is developed by integrating a CNN-BiLSTM-AM model with the Monte Carlo simulation algorithm, embedded within the scenario analysis method. The model projects the dynamic trajectories of carbon emissions in the Chinese building sector under different scenarios from 2023 to 2050 and identifies effective schemes for controlling carbon emissions in the Chinese building sector. Results indicate that the growth in legal and policy incentive intensity was most significant during the 12th Five-Year Plan period in China. During the 13th Five-Year Plan in China, the legal and policy system became increasingly mature, leading to a diminishing marginal effect of newly issued policies. A negative growth in legal and policy incentive intensity was observed in 2020 due to the impact of the COVID-19 pandemic. From 2021 to 2022, the annual growth rate of policy intensity began to rebound. Under the current scenario, carbon emissions in the Chinese building sector are projected to reach its carbon peak in 2036 (±1), with a peak level of 28.617 (±1.047) × 10⁸ t CO₂. Energy consumption per unit floor space, population size, legal and policy incentive intensity, integrated carbon emission factor, and floor space per capita are identified as the most critical factors influencing the timing and value of carbon peaking. The research methodology employed in this study not only provides scientific insights for the emission reduction efforts in the building sector but is also applicable to related studies in other industries’ energy conservation and emission reduction. It holds universal value for environmental policymakers and strategic planners.

1. Introduction

Achieving carbon peak (CP) and carbon neutrality (CN) represents China’s solemn commitment to the international community and serves as an intrinsic requirement for implementing the new development philosophy, fostering a new development paradigm, and advancing high-quality development [1,2]. The realization of the “CP and CN” targets necessitates concerted efforts across all sectors, particularly within the building sector, which constitutes one of the major global sources of greenhouse gas emissions [3,4]. Existing studies reveal that over the past two decades of the 21st century, carbon emissions in the Chinese building sector (CECBS) have contributed approximately 1.49 billion tons of CO₂. By 2020, the share of CECBS in China’s total carbon emissions had exceeded 20% [5,6]. This indicates that the trajectory of carbon peaking within the building sector will play a pivotal role in achieving China’s national climate target of reaching CP by 2030. Managing CECBS while sustaining economic growth has become one of the critical challenges the Chinese government must address to ensure the timely realization of the 2030 CP goal [7,8]. Therefore, this study aims to comprehensively quantify the impact of laws and policies on achieving CP in CECBS and further predict the path of CP and its uncertainty under different scenarios. The goal is to provide scientific evidence for the government’s efforts in carbon emission reduction in the building sector. To achieve this objective, the study raises the following critical questions:

• How can the impact of laws and policies on achieving CP in CECBS be quantified?
• How can the path of CP in CECBS be characterized under the influence of uncertainty?
• How can effective measures to accelerate the CP in CECBS be identified?

A substantial body of research has been conducted on the decarbonization transition of CECBS [9,10,11]. With respect to influencing factors, existing studies explore the key determinants of CECBS from multiple perspectives, including PZ and structure [12,13], urban–rural spatial distribution [14,15], age composition of residents [16], household composition [17], transformation of the energy structure [18,19], electrification level [20,21], changes in income [22,23], FSC [24,25], heating system types [26,27], population mobility and interprovincial migration [28], climate change [29], economic growth [30,31], trends in energy–carbon emission decoupling [32,33], and energy consumption behavior patterns [34,35]. It is noteworthy that legal documents and policies play a critical role in supporting and guiding energy conservation and emission reduction efforts in the building sector, thereby contributing significantly to the realization of CP and CN targets in CECBS [36]. However, current research in this area primarily relies on indirect indicators—such as government financial expenditures—to assess the incentive intensity of legal documents and policies on energy conservation and emission reduction in the building sector [37], or applies difference-in-difference methods to examine the impact of individual laws or policies after their implementation. Comprehensive quantitative evaluations of the impact of legal documents and policies on achieving the CP in CECBS remain absent.

Meanwhile, in the study of the evolutionary trends of CECBS, the selection of forecasting tools directly determines the accuracy and credibility of the results. Due to differences in assumption logic and computational mechanisms, various modeling approaches often yield significantly divergent projections for the future trajectory of CECBS. Commonly adopted forecasting methods include energy planning models (e.g., LEAP) [38,39], the IPAT framework [40,41], the Stochastic Impacts by Regression on Population, Affluence and Technology (STIRPAT) model [42,43], the Environmental Kuznets Curve (EKC) theory [44,45], system dynamics models [46,47], and integrated assessment models [48]. Although these methods offer substantial value in policy scenario formulation and pathway simulation, they are generally characterized by structural rigidity and a high dependence on theoretical assumptions. Consequently, they often fail to adequately capture the interactions among variables and the nonlinear feedback mechanisms inherent in complex systems.

To overcome the limitations of traditional models in terms of adaptability and complexity handling, recent years have witnessed a growing interest among scholars in incorporating artificial intelligence into carbon emission modeling. For instance, certain studies have employed machine learning algorithms to fit and forecast the carbon emissions of buildings in selected Chinese cities [49,50]. These approaches can automatically identify latent patterns within data during model training, thereby reducing subjective bias from manually defined equations. However, most machine learning models still rely on shallow network structures and exhibit limited capacity to capture high-order nonlinear relationships among variables, often failing to accurately characterize the dynamic evolution of carbon emission systems [51]. In contrast, deep learning methods, which have rapidly advanced in recent years, demonstrate significant potential in forecasting carbon emissions in the building sector. Leveraging multi-layered network architectures, these techniques autonomously extract and abstract information from raw data, enabling the recognition of long-term trends and complex dependencies that are typically imperceptible to conventional models. Compared to traditional approaches, deep learning methods are more data-driven, exhibit strong generalization capabilities, and incorporate dynamic response mechanisms, making them particularly well-suited for processing high-dimensional, time-variant, and highly nonlinear data. Existing studies have indicated that models based on deep neural networks can achieve high levels of accuracy in predicting CECBS [37]. Nevertheless, current research predominantly addresses the future trajectory of CECBS from the perspective of static parameter changes, without incorporating the influence of uncertainties on its dynamic evolution. To date, studies employing deep learning methods for dynamic forecasting of CECBS remain scarce.

A comprehensive review reveals several existing gaps in current research on CECBS. On one hand, at present, no comprehensive quantitative study has been conducted to assess the impact of legal documents and government policies on achieving CP in CECBS. On the other hand, research employing deep learning methods for the dynamic prediction of carbon emissions in CECBS remains notably scarce.

To address these issues and bridge the existing research gaps, this study constructs an integrated quantitative framework for legal and policy incentive intensity, which captures both the immediate effects and long-term evolution of legal and policy measures. Utilizing this framework, the study quantifies the legal and policy incentive intensity index related to CECBS from 2010 to 2022. Building upon this, a dynamic scenario-based prediction model for CECBS is developed by combining a CNN-BiLSTM-AM model with Monte Carlo simulation and embedding scenario analysis. Using this model, this study forecasts the dynamic trajectory of CECBS under various scenarios from 2023 to 2050 and identifies effective measures to control CECBS.

The main contributions of this study are as follows:

• A comprehensive quantitative framework for legal and policy incentive intensity is developed, capable of capturing both the immediate effects and long-term evolution of legal and policy measures. By quantifying the policy implications and incorporating an influence attenuation factor, this framework characterizes the temporal accumulation and decay of policy effects, thereby enabling a scientific quantification of the intensity of a large number of legal and policy incentives over an extended period.
• A dynamic scenario-based prediction model for CECBS is constructed by integrating the CNN-BiLSTM-AM model with the Monte Carlo simulation algorithm within a scenario analysis approach. This method addresses the existing gap in the field, where deep learning techniques have rarely been employed for dynamic scenario forecasting of CECBS.

2. Data and Methodology

2.1. Data Sources

All energy consumption data used in this study are derived from the energy balance sheets in the China Energy Statistical Yearbook [52]. The carbon emission factors for primary energy sources required to calculate CECBS are taken from the Standard for Building Carbon Emission Calculation (GB/T 51366-2019) [53]. The methods for calculating the carbon emission factors of secondary energy sources, such as electricity and heat, follow the 2021 Annual Report on China’s Building Energy Consumption and Carbon Emissions [54]. Population and economic data are obtained from the China Statistical Yearbook [55]. The number of green invention patents granted is sourced from the National Intellectual Property Administration (https://www.cnipa.gov.cn/, accessed on 1 July 2025).

2.2. Quantification Method for Legal and Policy Incentive Intensity Considering Cumulative Effects

Legal documents and policies support and guide energy-saving and emission-reduction efforts in the building sector, playing an indispensable role in advancing CECBS toward “CP and CN” targets. However, existing research in this field largely focuses on using government fiscal inputs and other quantifiable indicators to indirectly measure legal and policy incentive intensity for CECBS or employs difference-in-difference models to explore the impact of individual laws or policies. To date, no comprehensive study has quantified the impact of legal documents and policies on achieving CP in CECBS. To address this gap, a comprehensive quantification framework for legal and policy incentive intensity is developed to reflect both immediate effects and long-term evolution. The framework quantifies policy implications and incorporates an influence-decay factor to characterize the temporal accumulation and attenuation of policy effects, thereby achieving a scientific measurement of incentive intensity. This approach not only quantifies the instantaneous impact of individual laws and policies but also reveals the additive effects of multiple laws and policies over time, providing more precise tools for predicting CECBS and evaluating policy effectiveness.

Currently, mainstream browser automation testing frameworks include Selenium, Playwright, and Puppeteer, all of which can achieve automatic web content crawling. Among these, Selenium stands out as a mature, well-documented open-source solution. It offers broad support across languages and browsers, a highly active developer community, and a rich ecosystem of plugins, making it capable of meeting automation needs in complex scenarios. Selenium provides high stability and ease of use, accurately simulating real user behavior to ensure the capture of complete, rendered textual data. Therefore, this study first selects the Selenium framework to automatically crawl policy documents related to carbon reduction on the Peking University Law Database (https://pkulaw.com/, accessed on 1 July 2025). The retrieval process uses keywords such as “carbon emissions,” “carbon reduction,” “carbon tax,” “carbon sink,” “energy conservation,” “energy consumption,” and “greenhouse gases.” Subsequently, this study conducts a rigorous screening of the collected documents, eliminating those that do not meet the predetermined criteria. The documents are then manually scored based on their legal and policy authority levels as well as the scope of their policy targets. This study assigns distinct ranges of policy quantification scores to different categories of documents according to their hierarchical level of authority, from highest to lowest: laws, administrative regulations, intra-party regulations and systems, departmental rules, group provisions, and industry regulations. Furthermore, for documents of the same hierarchical level, this study assigns different scores based on the extent to which the document focuses on CECBS, as well as the document’s objectives and feasibility. The specific scoring criteria and detailed descriptions of each level are shown in

Table 1. Evaluation criteria for legal and policy texts.

Score	Evaluation Criteria
1	(a) Documents with the hierarchical level of authority of laws that relate to emission reduction. (b) Documents with the hierarchical level of authority of administrative regulations and intra-party regulations that fully focus on CECBS.
0.75	(a) Emission-reduction documents with the hierarchical level of authority of administrative regulations and intra-party regulations that do not fully focus on CECBS. (b) Documents with the hierarchical level of authority of departmental rules that specify emission-reduction targets for CECBS and set concrete, executable policy measures.
0.5	(a) Documents with the hierarchical level of authority of departmental rules that establish overall emission-reduction directions and provide some guidance on decarbonization but insufficiently cover key areas of CECBS and offer limited support. (b) Documents with the hierarchical level of authority of group standards and industry regulations that fully focus on CECBS.
0.25	(a) Emission-reduction documents with the hierarchical level of authority of group standards and industry regulations that do not fully focus on CECBS. (b) Documents with the hierarchical level of authority of departmental rules that merely relay information and omit any concrete targets or policy measures for curbing CECBS or carbon emissions in related field.

. In addition, to ensure the consistency of the scoring for legal and policy documents, this study invites three experts to independently score the policy texts according to the scoring rules. A Kappa consistency test is then performed to assess the inter-rater consistency between the experts. The results show that the average Kappa value between pairs of experts is 0.91, which passes the Kappa consistency test, ensuring consistency in document scoring. Through the aforementioned measures, this study obtains a quantified score representing the legal and policy incentive intensity of individual policies.

However, it is noteworthy that policy effects do not vanish in the short term; rather, they exhibit temporal persistence and attenuation, indicating the presence of cumulative policy effects. The cumulative effect of policies refers to the gradual accumulation of their influence over time following the successive introduction of policy measures within a specific domain, instead of relying on the isolated impact of a single policy. The formation of this effect is primarily influenced by policy continuity, sectoral adaptability, and the overlay effect. First, policies typically do not yield significant effects immediately after their implementation. Instead, as stakeholders within the industry progressively adapt and adjust, the impacts of these policies gradually emerge and usually do not dissipate completely in the short term. Second, in the process of promoting the CP in CECBS, the government frequently adjusts its policies in response to the developmental needs of CECBS and technological advancements. Newly introduced policies often partially inherit or reinforce the objectives of existing ones, thereby generating overlay effects. Lastly, initial policy stimuli may exert a strong guiding influence on CECBS; however, as the industry gradually adapts, the marginal effects of these policies tend to diminish. Therefore, when quantifying legal and policy incentive intensity, it is essential to establish a reasonable time decay mechanism to accurately reflect the long-term impact of policies. Neglecting the temporal cumulative effect of policies may lead to biased assessment results. Thus, the time dimension must be fully considered when quantifying legal and policy incentive intensity to ensure an accurate representation of policy implications.

It is important to note that legal and regulatory instruments possess unique binding power and stability during their implementation process. Unlike other policy measures, their effectiveness does not gradually diminish over time. Specifically, laws and regulations typically take the form of mandatory provisions and long-term legal obligations, characterized by strong persistence and irreversibility. Therefore, this study assumes that the influence of laws and regulations on CECBS remains stable over time and does not exhibit diminishing marginal effects as time progresses.

Building on the work of Zhang et al. [56] and Yu et al. [57], this study introduces a decay factor $\lambda$ to assess the temporal cumulative effect of legal and policy incentive intensity. By aggregating the annual policy intensity and its corresponding effects, an index of legal and policy incentive intensity accounting for the time effect is ultimately derived. The specific formulation is presented in Equation (1). In this study, the decay factor $\lambda$ is set to 0.95.

$$LP{I}_t=\lambda ·LP{I}_{t-1}+{\displaystyle \sum _{i=1}^{n}lp{i}_i}$$

(1)

where $LP{I}_t$ denotes legal and policy incentive intensity for the year $t$; $LP{I}_{t-1}$ denotes legal and policy incentive intensity for the year $t-1$; $lp{i}_i$ denotes legal and policy incentive intensity of the $i$ text in year $t$; and $n$ denotes total number of legals and policies issued in year $t$.

2.3. Dynamic Scenario Forecasting Model for CECBS Based on Deep Learning

2.3.1. Model Framework

The dynamic scenario forecasting model for CECBS based on deep learning integrates the CNN-BiLSTM-AM model—a type of deep learning model—with the Monte Carlo simulation algorithm, embedding it within a scenario analysis framework. Figure 1 presents the model framework of the aforementioned model. Firstly, this study incorporates legal and policy incentive intensity into the influencing factor analysis framework and constructs a model feature system for forecasting CECBS from five dimensions. Secondly, the CNN-BiLSTM-AM algorithm is employed to capture the complex nonlinear relationships among the historical patterns of CECBS and the various influencing factors within the feature system. Subsequently, Monte Carlo simulation and scenario analysis are introduced to establish a dynamic scenario-based forecasting model for CECBS driven by deep learning. Finally, this model is utilized to predict the dynamic trajectory and peak status of CECBS from 2023 to 2050 under different scenarios. Furthermore, FAST is used to evaluate the influence of different factors on the time of CP and the value of CP in CECBS.

2.3.2. Construction of the Model Feature System

To identify the influencing factors of CECBS and thereby enhance the accuracy of future CECBS projections, this study designs a model feature system specifically for CECBS forecasting. This system encompasses five dimensions. It includes ten factors influencing CECBS: PZ, UR, DIC, FSC, ECFS, ICEF, STEC, NGPGC, ECEPEC, and legal and policy incentive intensity.

Specifically, the population structure dimension comprises two variables: PZ and UR. PZ indicates the total number of residents in a region and directly impacts the demand for buildings and the corresponding level of CECBS. As PZ increases, demand for residential buildings, public facilities, and infrastructure expands, leading to higher energy consumption and increased CECBS. The UR reflects the proportion of rural-to-urban migration in the process of urban development. A higher UR implies a concentration of construction activities and energy use in urban areas. Urbanization intensifies building density and promotes the prevalence of high-rise buildings, thereby increasing the carbon emission intensity in the building sector. Consequently, these two variables influence carbon emissions not only in terms of volume but also by shaping the spatial distribution of energy use.

The economic development dimension includes two variables: DIC and FSC. As a critical indicator of residents’ economic status, DIC directly reflects their capacity to purchase and utilize building products. An increase in income levels generally leads to a preference for higher-quality and larger residential and commercial spaces, thereby contributing to greater building energy consumption and CECBS. FSC, another key indicator of living standards and urban expansion, increases with improvements in residents’ quality of life. While such increases often drive demand for enhanced building energy efficiency technologies, they also typically lead to higher ECFS and elevated carbon emissions.

The energy structure dimension consists of two variables: ECFS and the ICEF. ECFS measures the energy use intensity within the building sector, reflecting the efficiency of energy utilization throughout the construction and operation phases. A reduction in this indicator generally signifies improved energy efficiency and decreased carbon emissions. The ICEF quantifies the impact of the energy types used in the building sector on CECBS, thus indicating the implications of energy structure changes for CECBS. Optimizing the energy structure—such as progressively replacing traditional fossil fuels with clean energy—represents a fundamental pathway toward low-carbon development in CECBS.

The technological capability dimension comprises two variables: STEC and the NGPGC. STEC reflects the level of investment by governments or society in technological innovation and research and development, particularly in support of energy conservation and emission reduction technologies. Increasing investment in science and technology research facilitates breakthroughs and the widespread application of new low-carbon technologies, thereby advancing the green transformation of the building sector. The NGPGC serves as a key indicator of technological innovation outcomes. Growth in the number of patents indicates an accelerated pace of technological advancement, especially in innovations related to building energy efficiency and green construction, which are crucial for the low-carbon development of CECBS.

The policy support dimension includes two variables: ECEPEC and legal and policy incentive intensity. ECEPEC represents the fiscal support provided by the government for energy-saving and emission-reducing initiatives, directly influencing the research and development of low-carbon technologies, the promotion of green buildings, and the effectiveness of policy implementation. Higher fiscal investment generally signals more effective support for carbon reduction measures in the building industry. Legal and policy incentive intensity reflects the extent to which various laws and regulations issued by the government are enforced, as well as the specific requirements directed at the building sector. This metric captures both the restrictive and guiding functions of policies. Enhancing the intensity of legal and policy incentives can encourage carbon reduction efforts among enterprises and social entities, thereby advancing the gradual realization of green development pathways in CECBS.

2.3.3. CNN-BiLSTM-AM Model

Convolutional Neural Network (CNN)

CNN demonstrates significant advantages in processing time series data due to its ability to automatically extract the most critical local features from input data. Specifically, CNN performs convolution operations on the input using a local receptive field mechanism, enabling the capture of short-term variations and latent patterns within the data—an aspect that is particularly crucial for feature extraction in time series analysis. Unlike traditional neural networks, CNN employs a weight-sharing strategy, whereby each convolutional kernel learns only a limited number of parameters. This approach greatly reduces computational complexity and mitigates the risk of overfitting. Moreover, CNN models local spatial correlations in the data through its convolutional layers during the feature extraction process. This capability allows the model to effectively focus on localized dependencies within the input while filtering out irrelevant noise or redundant information. In the preprocessing of time series data, convolution operations not only eliminate excessive noise but also enhance the recognition of meaningful patterns in the signal, thereby improving the generalization ability and predictive performance of subsequent models. The pooling layer further compresses the dimensions of the feature maps, reducing feature dimensionality and enhancing model stability. This ensures a balance between computational efficiency and information retention. Given these advantages, CNN is employed in this study as a feature extraction module, enabling the model to uncover latent structures in the data and provide clear, efficient feature representations for downstream deep learning components. The computational process is shown in Equation (S1) in the Supplementary Materials.

Bidirectional Long Short-Term Memory Network (BiLSTM)

LSTM demonstrates strong performance in processing time series data due to its ability to effectively capture long-term dependencies through structures such as forget gates, input gates, and output gates. Although LSTM possesses unique advantages in maintaining long-term memory, it remains constrained by its unidirectional modeling structure. Specifically, the output of LSTM relies solely on historical information, lacking the capacity to incorporate future context. This limitation means that predictions can only be made based on past time series data, thereby reducing the model’s expressive power and forecasting accuracy in complex temporal tasks. In scenarios requiring a comprehensive understanding of both past and future contexts in time series data, this unidirectional flow often leads to diminished predictive performance. In response to these limitations, BiLSTM has been developed.

BiLSTM addresses the unidirectional constraint of traditional LSTM models by integrating both forward and backward LSTM layers. The primary advantage of the BiLSTM model lies in its ability to learn temporal dependencies in two directions: from past to present (forward information) and from future to present (backward information). Specifically, BiLSTM generates two hidden states at each time step: one from the forward LSTM layer and another from the backward LSTM layer. This bidirectional architecture enables the model to incorporate contextual information from both temporal directions at every point in time, resulting in a more comprehensive understanding of the sequence data. Overall, compared to LSTM, the BiLSTM architecture significantly enhances the model’s ability to learn complex patterns in time series by capturing bidirectional dependencies, thereby improving generalization performance and predictive accuracy in high-dimensional time series forecasting tasks. The computation of its parameters is shown in Equations (S2)–(S8) in the Supplementary Materials.

Attention Mechanism (AM)

AM represents a significant advancement in the field of deep learning, particularly in applications such as time series forecasting and natural language processing, where it has achieved notable success. The core concept of AM involves computing attention weights for the input data to automatically focus on the most critical features, thereby enabling the model to process information with greater efficiency and precision. Specifically, the mechanism learns the interrelations among different parts of the input data and dynamically adjusts the model’s focus accordingly. This approach allows for more flexible handling of complex tasks, especially when dealing with long sequences or high-dimensional data, significantly enhancing the model’s representational capacity and processing performance. In the AM model, the process begins with calculating the similarity between output vectors and input vectors, followed by scaling and normalization. The resulting weights are then used to perform a weighted summation with the value vectors. The detailed computation process is shown in Equation (S9) in the Supplementary Materials.

CNN-BiLSTM-AM Model

In conclusion, the CNN model exhibits notable advantages in feature extraction, with its capability to learn local features enabling efficient identification of essential information during data processing. The BiLSTM model enhances the learning of complex temporal patterns by incorporating both forward and backward information from the input data, thereby demonstrating superior generalization and predictive accuracy in high-dimensional time series forecasting tasks. However, despite its strength in capturing temporal dependencies, BiLSTM exhibits relatively limited capacity in local feature extraction. The integration of CNN and BiLSTM leverages the local feature extraction strengths of CNN and the global temporal modeling capacity of BiLSTM, enabling more comprehensive and effective data processing. This combination enhances the model’s ability to accurately capture both short-term and long-term dependencies within time series data. Simultaneously, the AM model, by computing attention probability distributions, emphasizes critical input features, further strengthening the model’s ability to extract temporal characteristics from the data. Based on the above analysis, these three methods demonstrate complementary advantages. Therefore, the present study embeds the CNN model into the BiLSTM structure and introduces the AM model to construct the CNN-BiLSTM-AM model. A schematic of the model architecture is illustrated in Figure 2. In this framework, the CNN model is employed to extract features from the input data; the BiLSTM model captures and learns long-term dependencies within the time series and outputs future predictions; and the AM model enhances the model’s ability to capture temporal relationships, thereby further improving predictive accuracy.

2.3.4. Model Performance Evaluation

To evaluate the performance of the CNN-BiLSTM-AM model, this study employs several metrics, including MAE, MSE, MAPE, and R², as evaluation indicators. Simultaneously, this study calculates the relative error ($E_i{|}_t$) between the model’s predicted CECBS for the actual CECBS for the corresponding historical periods. In addition, to confirm the superior performance of the proposed CNN-BiLSTM-AM model, the same set of tests is applied to Support Vector Regression (SVR), LSTM, BiLSTM, and Gated Recurrent Unit (GRU) networks for comparison. The ratio of the training set to the test set is 8:2. The formulas for these evaluation indicators are shown in Equations (S10)–(S14) in the Supplementary Materials.

2.3.5. Scenario Design

After effectively capturing the relationships between CECBS and various influencing factors, this study constructs three distinct scenarios to further examine the development of CECBS under different conditions of legal and policy frameworks, technological advancement, and socio-economic progress. These scenarios include a baseline scenario, a high carbon scenario, and a low carbon scenario. In the baseline scenario, all parameters follow their current developmental trajectories. The high-carbon scenario assumes that government efforts prioritize socio-economic growth while placing comparatively less emphasis on energy conservation and emission reduction in the building sector. In this context, the implementation intensity of energy-saving policies declines relative to the baseline, and the advancement of energy-saving technologies lags behind. Key factors such as FSC, ECFS, and ICEF increase compared to the corresponding values in the baseline scenario. In contrast, the low-carbon scenario reflects a governmental focus on promoting energy conservation and emission reduction in the building sector alongside economic development. The strength of policy implementation in this scenario significantly surpasses that in the baseline, accompanied by a higher level of technological advancement. Consequently, core indicators such as FSC, ECFS, and ICEF exhibit substantial reductions relative to the baseline scenario during the same period.

2.3.6. Monte Carlo Simulation

Although static scenario analysis facilitates the exploration of variations in key variables under differing conditions of legal and policy incentive intensity, technological advancement, and socio-economic development, it often fails to adequately capture the stochastic fluctuations induced by future changes in the policy, economy, and technology. Monte Carlo simulation, renowned for its random sampling and statistical analysis capabilities, is employed here to capture the full range of uncertainties in CECBS projections by producing a multitude of stochastic scenarios. This methodology effectively delineates the potential trajectories of CECBS under various future conditions. In this study, Monte Carlo sampling is integrated within the CNN-BiLSTM-AM predictive framework to develop a deep learning-based dynamic scenario model for CECBS. Through the application of repeated random sampling and probabilistic assessments across multiple scenarios, this approach capitalizes on the pattern recognition capabilities of neural networks while enhancing the model’s ability to assess and quantify risk uncertainty. Based on the studies by Zhang et al. [47], this study sets the following uncertainty parameter (σ) rules under a 95% confidence level. The maximum and minimum values of each input variable in both the high-carbon and low-carbon scenarios should fall within the range covered by ±2σ for the respective period. That is, after applying a ±2σ change to each input variable, the maximum value within the range must be greater than the maximum value for the current period in the static scenario, and the minimum value within the range must be lower than the minimum value for the current period. The calculation process is shown in Equation (S15) in the Supplementary Materials.

2.3.7. Sensitivity Analysis of the Model

This study utilizes FAST to assess the influence of various variables on the uncertainty of the output in the dynamic scenario prediction model for CECBS. Through Fourier transformation, FAST breaks down the model output into sinusoidal components at different frequencies. By examining the amplitude of these frequency components, the method quantitatively determines the contribution of each variable to the output variance, thus providing an evaluation of its impact on future CECBS projections. The principal computational procedures are shown in Equations (S16)–(S19) in the Supplementary Materials.

2.3.8. Akima Interpolation

Owing to the limited historical data available for CECBS, the existing dataset is inadequate for training the model. In line with the approach proposed by Zhang et al. [37,56], this study enhances the training dataset by employing the Akima interpolation technique on the original historical data, significantly boosting the model’s performance. The detailed formula for the Akima interpolation method can be found in [58].

Furthermore, to better demonstrate the rationality of using Akima interpolation, we designed a small experiment. Specifically, this study trained the CNN-BiLSTM-AM model using only the newly interpolated data as the training and validation sets. The original true data does not participate in this process. Subsequently, we input the original data into the trained model to compare the error between the model’s predicted data and the true data. The experimental results are shown in Figure 3. As observed, the maximum error of the model does not exceed 2%. This indicates that even when trained exclusively with interpolated data, the model can effectively learn the complex relationships within the original true data.

Meanwhile, to further demonstrate that the model retains strong extrapolation ability after learning from interpolation data, this study designs an additional experiment. In this experiment, we train the model using only the 2010–2020 data (including interpolation data), while the actual data for 2021 and 2022 and the interpolation data for 2020–2022 do not participate in model training. After training, this study inputs the true values of the input variables for 2021 and 2022 into the trained CNN-BiLSTM-AM model to evaluate the relative error between predicted and actual values of CECBS. The results show that the relative error between the model’s predicted values and the actual data for 2021 is 0.35%, and for 2022 is 0.94%. Both relative errors are below 1%. This indicates that by learning from interpolation data, the model still possesses strong extrapolation ability and can accurately capture the historical patterns of CECBS. In summary, this study’s use of Akima interpolation for data expansion is reasonable.

3. Results and Discussion

3.1. Historical Evolution of Legal and Policy Incentive Intensity

The types and quantities of legal and policy documents collected in this study for the years 2010–2022 are shown in

Table 2. Types and quantities of legal and policy documents.

Type	Laws	Administrative Regulations	Intra-Party Regulations and Systems
Number	5	17	5
Type	Departmental rules	Group provisions	Industry regulations
Number	1122	6	151

, and Figure 4 illustrates the changes in legal and policy incentive intensity from 2010 to 2022. Overall, legal and policy incentive intensity shows a rapid growth trend throughout the study period. From 2010 to 2022, legal and policy incentive intensity increased from 67 to 435.73, an increase of approximately 550%, with an average annual growth rate of 17.5%. This significant upward trend reflects the continuous strengthening of the Chinese government’s policy support for CECBS. Specifically, from 2010 to 2015, legal and policy incentive intensity surged from 67 to 337.46, with an impressive average annual growth rate of 25.6%. This demonstrates the government’s strong emphasis on energy conservation and emission reduction during the “Twelfth Five-Year Plan” period. The rapid policy growth during this phase likely correlates with the escalating global climate change agenda and Chinese commitment to international emission reduction agreements, such as the post-Kyoto Protocol framework. Concurrently, at the national level, the “Twelfth Five-Year Plan” explicitly sets energy-saving targets for the building sector, leading to the promulgation of a series of policy documents that support green building and low-carbon development, which in turn significantly boosts legal and policy incentive intensity. During the “Thirteenth Five-Year Plan” period (2016–2020), legal and policy incentive intensity continues to grow, though at a slower pace, rising from 363.59 to 399.89, with the average annual growth rate decreasing to approximately 3.9%. In this stage, the policy system gradually improves, and the marginal effect of new policies weakens. However, the intensity still maintains positive growth, indicating that the government continuously optimizes its support within the existing policy framework. From 2021 to 2022, legal and policy incentive intensity further increases to 435.73, and the average annual growth rate rebounds to 5.5%. This resurgence may directly relate to the “CP and CN” targets proposed in the “Fourteenth Five-Year Plan,” as the government intensifies policy investment in low-carbon development during the post-pandemic era.

From the perspective of the growth rate of legal and policy incentive intensity, the growth rate of legal and policy incentive intensity peaked in 2011 at 98.73% and subsequently decreased year by year, reaching 1.13% in 2019. This trend indicates that the growth of legal and policy incentive intensity underwent a process from rapid expansion to gradual stabilization. The high growth rate in 2011 likely stems from the concentrated and forceful release of initial policies, such as the issuance of documents like the “Notice of the Department of Building Energy Efficiency and Science and Technology of the Ministry of Housing and Urban-Rural Development on Printing and Distributing the Key Work of the Department of Building Energy Efficiency and Science and Technology of the Ministry of Housing and Urban-Rural Development in 2011,” which quickly established the policy foundation for carbon emission reduction.

It is worth noting that, in 2020, the legal and policy incentive intensity decreased from 399.89 in 2019 to 391.9, with a growth rate of −2.00%. This marks the only year with negative growth during the study period. This decline is highly correlated with the global outbreak of the COVID-19 pandemic. The pandemic caused a significant slowdown in economic activity, and the building sector experienced a substantial impact due to construction stagnation and supply chain disruptions, which likely limited the effectiveness of related policies. Concurrently, in 2020, the government shifted its policy focus towards economic recovery and social stability. For instance, it stimulated the economy through large-scale infrastructure projects rather than prioritizing carbon emission reduction targets, which led to a temporary reduction in investment in energy conservation and emission reduction-related policies. Furthermore, the decrease in energy consumption in the building sector during the pandemic may have further reduced the apparent demand for policy incentives. However, in 2021 and 2022, the legal and policy incentive intensity rapidly rebounded, increasing to 406.55 and 435.73, with growth rates recovering to 3.74% and 7.18%, respectively. This rebound indicates that the government readjusted its policy priorities post-pandemic, increasing support for carbon emission reduction. In particular, the “Fourteenth Five-Year Plan” explicitly proposed the goal of achieving CP before 2030, which drove the intensive introduction of policies in areas such as green buildings and low-carbon technologies. For example, new energy-saving subsidies or mandatory emission reduction standards may have been introduced in 2021, stimulating the rapid recovery of legal and policy incentive intensity.

3.2. Effectiveness Test of the CNN-BiLSTM-AM Model

To assess the effectiveness of the CNN-BiLSTM-AM model developed in this research, five different models—SVR, LSTM, BiLSTM, GRU, and CNN-BiLSTM-AM—were evaluated. Their performance was tested on the test set using metrics such as MAE, MSE, MAPE, R², and the relative error ($E_i{|}_t$) between predicted and actual values for 2021 and 2022. The results, presented in

Table 3. Model performance test results.

Model	MAE	MSE	MAPE(%)	R2	${\left.\mathit{E}\right|}_{2021}(\mathit{\%})$	${\left.\mathit{E}\right|}_{2022}(\mathit{\%})$
SVR	0.492	1.488	3.209	0.933	−3.95	−3.28
LSTM	0.352	0.581	1.955	0.961	−1.44	−1.41
BiLSTM	0.236	0.299	1.394	0.971	−0.73	−1.03
GRU	0.224	0.241	1.232	0.987	−0.82	−0.97
CNN-BiLSTM-AM	0.134	0.029	0.673	0.994	−0.69	−0.60

, demonstrate that the CNN-BiLSTM-AM model consistently outperforms the other models across all evaluation metrics, highlighting its superior predictive capability for CECBS. Additionally, the CNN-BiLSTM-AM model shows better performance than the BiLSTM model without the combined modules, confirming the rationale and necessity of incorporating these modules in this study. Overall, the CNN-BiLSTM-AM model shows strong predictive performance for CECBS and is capable of effectively simulating its development trajectory through parameter adjustments in simulation experiments.

3.3. Dynamic Evolution Paths of CECBS from 2023 to 2050 Under Different Scenarios

Deep learning-based dynamic scenario prediction model for CECBS fully considers the nonlinear evolution characteristics of future CECBS under the influence of multiple factors across different scenarios, including legal policies, technological advancements, and socioeconomic development. It also effectively captures the fluctuation range and probability distribution of the prediction interval by simulating perturbations to key parameters using the Monte Carlo method. The model sets the prediction period from 2023 to 2050, systematically depicting three potential dynamic development paths for CECBS—baseline, high carbon, and low carbon—under various socioeconomic environmental impacts. Figure 5 illustrates the specific results. Overall, CECBS exhibits an “inverted U-shaped” change across all three scenarios between 2023 and 2050. However, the development trends, value of CPs, and times of CP for CECBS show significant differences.

Specifically, in the baseline scenario, CECBS continues to grow rapidly from 2023 to 2030, with an average annual growth rate of 1.76%. From 2030 to 2035, the growth rate of CECBS begins to decline, decreasing to an average annual rate of 1.12%. In this scenario, CECBS achieves its CP in 2036 (±1), with a peak value of 28.617 (±1.047) × 10⁸ t CO₂. This result indicates that, without strong policy intervention and technological advancements, CECBS will continue to increase before 2030. Furthermore, because current policies do not effectively curb the growth rate, the time of CP for the building sector significantly exceeds the 2030 target. After peaking, CECBS begin to decrease year by year, with the rate of decrease increasing annually. By 2050, CECBS will decrease to 23.525 (±1.811) × 10⁸ t CO₂, returning to the emission level of 2022. This study finds that, under current conditions, the base level of CECBS remains substantial by 2050, indicating a considerable distance from achieving the CN target. Therefore, in the baseline scenario, both the time of CP and the carbon neutrality time for CECBS differ significantly from the targets. The government must implement more robust policy measures to close this gap.

In the high-carbon scenario, CECBS are not effectively controlled, and their growth rate from 2023 to 2030 further climbs to an average of 2.21% annually. The growth rate remains above 1% per year until 2036. CECBS then achieves its CP in 2038 (±1), with a peak value of 30.336 (±1.422) × 10⁸ t CO₂. In this scenario, the time of CP is delayed by two years compared to the baseline scenario and by eight years compared to the target of CP. The peak value increases by 1.719 × 10⁸ t CO₂ relative to the baseline scenario. After reaching its peak, the rate of decarbonization for CECBS remains slow. By 2043, the carbon emission value will still remain above 29 × 10⁸ t CO₂. If the period from 2033 is considered, the plateau phase for carbon emission reduction extends for a full 11 years, significantly longer than both the baseline and low-carbon scenarios. This result indicates that loosening government control over CECBS could result in severe outcomes, undermining the effectiveness of prior carbon emission reduction measures.

In the low-carbon scenario, the growth rate of CECBS will be further and effectively controlled, with the carbon emission growth rate falling below 1% after 2030. Under this scenario, CECBS will reach its CP around 2034 (±1), with a peak value of 26.749 (±0.733) × 10⁸ t CO₂. Compared to the baseline scenario, its time of CP is two years earlier, and the peak value is reduced by 1.868 × 10⁸ t CO₂. By 2050, the carbon emission value for CECBS will decrease to 20.566 × 10⁸ t CO₂, falling back to the emission level of 2017. This study observes that even in the low-carbon scenario, it remains challenging for CECBS to achieve the climate target of peaking by 2030. This result demonstrates the difficulty of achieving the target of CP for CECBS and the urgency of implementing strong energy-saving and decarbonization efforts in the building sector. Governments at all levels should further enhance the enforcement of policies and optimize relevant energy policies based on their own actual conditions, encourage the development and application of new energy technologies in the building sector, and promote the early realization of the CP for CECBS.

3.4. Identifying Effective Measures for Controlling CECBS and Policy Implications

To quantify the influence of various factors on future values of CP and times of CP for CECBS, and thereby identify effective control measures, this study employs FAST on the dynamic carbon emission prediction model. Figure 6 illustrates the results. Overall, the impact levels of various influencing factors on both the time of CP and the value of CP differ to some extent. However, the most critical factors influencing both the time of CP and value of CP are ECFS, PZ, legal and policy incentive intensity, ICEF, and FSC.

Specifically, ECFS impacts the value of CP by 38.66% and the time of CP by 25.90%. It stands as the most significant factor influencing the value of CP and the second most significant factor affecting the time of CP. This result reveals the structural weight of building energy efficiency in the CP process for CECBS. Currently, Chinese building energy consumption accounts for a relatively high proportion of its total end-use energy consumption. Changes in ECFS directly determine the carbon emission intensity corresponding to unit building activity. If China can continuously promote energy-efficient building design, optimize the energy efficiency of systems such as air conditioning and lighting, and enhance overall energy efficiency through energy-saving renovations during the building operation phase, it can achieve a leap in energy use efficiency without compromising residential and office comfort. Particularly during the “Fourteenth Five-Year Plan” period and beyond, the green building development path urgently needs to shift towards “energy efficiency improvement” as its core objective, promoting the transformation of high-energy-consuming buildings to low-carbon models. It is crucial to note that improving energy efficiency during the building operation phase is more long-term and systemic than in the earlier construction phase. This means that optimizing ECFS is not only a crucial lever for short-term emission control but also an indispensable core driver for achieving the long-term CP strategic goal.

PZ influences the value of CP by 20.90% and the time of CP by 37.57%. This makes it the second most significant factor affecting the value of CP and the most significant factor impacting the time of CP. This result indicates that changes in total population not only relate to shifts in overall building demand but also directly affect the evolution of total energy consumption and CECBS pathways. While the Chinese urbanization process will gradually slow in the future, regional population migration and the development of metropolitan areas will remain significant trends. This, coupled with the transformation of building usage demands due to an aging population, profoundly alters the CECBS. Therefore, formulating forward-looking building plans that account for population growth is crucial. This involves improving population prediction models and guiding public building and residential development to reasonably match population changes, thereby controlling the total CECBS per capita. Additionally, changes in population structure should prompt a transformation in building functions and operation and maintenance methods; for example, architectural designs for the elderly may focus more on indoor environmental regulation, indirectly influencing energy consumption intensity.

Legal and policy incentive intensity influences the value of CP by 11.84% and the time of CP by 10.37%. This makes it the fourth most significant factor affecting the value of CP and the third most significant factor impacting the time of CP. This result demonstrates the irreplaceable role of laws and policies as external guiding forces in establishing a low-carbon pathway for the building sector. Effective legal and policy incentives not only reduce the cost of low-carbon technologies but also provide necessary financial support for building energy conservation and carbon emission reduction projects through fiscal subsidies, tax reductions, and green credit. To expand policy benefits, China should construct a differentiated and quantifiable incentive indicator system. It is important to note that policy incentives include not only traditional economic measures like subsidies and tax exemptions but also multi-level intervention methods such as regulations, standards, administrative orders, and informational guidance. While building energy conservation policies have established a preliminary framework over the past few decades, significant gaps still exist in institutional coordination, implementation rigidity, and regional adaptability. In the future, China urgently needs to build a nationally unified yet flexible policy system to guide local governments in strengthening green building development tailored to local conditions. Particular attention should be paid to the “time effect” and “behavioral response mechanism” of policies, avoiding situations where short-term stimuli lead market entities to lose enthusiasm for long-term energy-saving investments. Furthermore, China needs to strengthen the dynamic adjustment capability of policies to respond to external environmental changes caused by energy price fluctuations, technological updates, and urban development strategy adjustments. The effectiveness of policy incentives ultimately depends on their ability to precisely stimulate green transformation behaviors in building enterprises and users.

The ICEF influences the value of CP by 12.63% and the time of CP by 10.29%. This places it as the third most significant factor affecting the value of CP and the fourth most significant factor impacting the time of CP. This result reflects the crucial position of energy structure in CECBS. Currently, Chinese building energy consumption primarily relies on coal and industrial waste heat, necessitating a significant increase in clean energy penetration. This study recommends that, at the national level, China continue to accelerate the proportion of non-fossil energy in electricity and heat supply. This involves establishing a “green electricity certificate” system and setting carbon emission intensity limits for heat sources, thereby encouraging regional energy centers to integrate more wind, solar, hydropower, and geothermal energy into centralized cooling/heating systems. Simultaneously, China should promote “combined heat and power + energy storage” projects to enhance energy conversion efficiency, reduce carbon emission intensity at the building energy consumption end, and provide solid energy assurance for national CP targets.

Concurrently, FSC influences the value of CP by 7.60% and the time of CP by 7.14%. Its impact level is only surpassed by legal and policy incentive intensity and the ICEF. This result demonstrates that the scale of residential and public building spaces has a moderate pulling effect on total carbon emissions and their timing. Amidst slowing urbanization and upgrading residential consumption structures, China should curb the blind expansion of FSC by promoting urban renewal and increasing the density of existing communities. In new urban developments, China should advocate for mixed-use land and vertical space development to enhance intensive land use. Furthermore, China needs to promote “multi-plan integration” in urban design, emphasizing “low-carbon community” and “shared space” concepts. This guides residents to reduce demand for non-essential floor space under complete public service facilities, thereby curbing the growth of CECBS at its source.

Additionally, the technology level dimension, composed of STEC and NGPGC, collectively impacts the value of CP by 6.22% and the time of CP by 6.75%. These are also significant factors influencing the future value of CP and the time of CP. This indicates that technological innovation provides a continuous and potential driving effect in controlling CECBS. To achieve this, China should increase research and development investment by both the state and enterprises in areas such as green building materials, structural optimization, and intelligent operation and maintenance systems, while also improving the collaborative mechanism among industry, academia, research, and application. China should support the establishment of regional technology centers and the formulation of industry standards, promoting the pilot application of “digital twin buildings” and “blockchain + carbon management” technologies in select cities. This enhances the accuracy of carbon emission monitoring and accounting throughout the entire life cycle of buildings. By fostering an innovation-friendly institutional environment, China can continuously generate low-carbon technological achievements with independent intellectual property rights, providing endogenous impetus for CP in CECBS.

4. Conclusions

This study constructs a comprehensive quantitative framework for legal and policy incentive intensity, which reflects both the immediate effects and long-term evolutionary processes of laws and policies. Using this framework, this study quantifies the legal and policy incentive intensity index for CECBS from 2010 to 2022. Building upon this, this study integrates the CNN-BiLSTM-AM model with the Monte Carlo simulation algorithm and embeds scenario analysis to develop a deep learning-based dynamic scenario prediction model for CECBS. Utilizing this model, this study predicts the dynamic development trajectories of CECBS under different scenarios from 2023 to 2050 and identifies effective measures for controlling CECBS. The main conclusions of this research are as follows:

(1)

From 2010 to 2022, the legal and policy incentive intensity generally exhibited a rapid growth trend. It reflects the continuous strengthening of the Chinese government’s policy support for decarbonization in CECBS. The growth in legal and policy incentive intensity was most pronounced during the “Twelfth Five-Year Plan” period. During the “Thirteenth Five-Year Plan” period, legal and policy incentive intensity continued to grow, but at a slower pace, as the relevant legal and policy systems matured and the marginal effects of new policies diminished. Affected by the COVID-19 pandemic, legal and policy incentive intensity experienced negative growth in 2020. From 2021 to 2022, the average annual growth rate of legal and policy incentive intensity began to rebound.

(2)

Under current conditions, CECBS will achieve CP around 2036 (±1), with a peak value of 28.617 (±1.047) × 10⁸ t CO₂. Even in the low-carbon scenario, CECBS still finds it challenging to achieve the climate target of peaking on time by 2030. This result highlights the difficulty of achieving the CP target for CECBS and the urgency of implementing strong energy-saving and decarbonization efforts in the building sector. Governments at all levels should further enhance policy enforcement and optimize relevant energy policies based on their specific conditions, encourage the development and application of new energy technologies in the building sector, and promote the early realization of CP for CECBS.

(3)

The impact levels of various influencing factors on the time of CP and the value of CP differ to some extent. However, the most critical factors influencing both the time of CP and value of CP are ECFS, PZ, legal and policy incentive intensity, ICEF, and FSC. Additionally, the technology level dimension, consisting of STEC and NGPGC, also represents a significant factor affecting the future value of CP and the time of CP. The government should focus on controlling ECFS and the ICEF. It should guide residents to reduce demand for non-essential floor space under complete public service facilities. Furthermore, while building energy conservation policies has established a preliminary framework, significant gaps still exist in institutional coordination, implementation rigidity, and regional adaptability. In the future, China urgently needs to build a nationally unified yet flexible policy system to guide local governments in strengthening green building development tailored to local conditions.

5. Existing Limitations and Future Research

The methods used in this study present several limitations that future work can address. First, the legal and policy text data come primarily from the PKULaw website, which—despite collecting the vast majority of energy conservation and carbon reduction-related legal and policy texts—still suffers from incomplete coverage or delayed updates, potentially affecting the accuracy of the legal and policy incentive intensity index. At the same time, we search using only the keywords most closely related to carbon emissions, which may still omit a small number of documents. Future research can incorporate additional databases and expand the keyword list to further reduce omissions. Second, the scoring of legal and policy texts in this study inevitably involves some subjectivity. Future work can develop semi-automated scoring approaches to further minimize subjective bias. Third, we select the time-decay factor based largely on existing studies and expert judgment. Future research can devise more rigorous, data-driven methods for determining this parameter. Fourth, the parameters and scenario designs in our dynamic scenario prediction model for CECBS can be further expanded. Doing so will enable more precise forecasting of CECBS.