Driving Green Technology Innovation via National Innovative City Policy—Evidence from a Combined DID, LSTM, and GRU Counterfactual Framework

Abstract

In the context of global climate governance, green technology innovation is essential for urban sustainable development. To address the limitations of traditional linear econometric models, this study investigates the impact of the National Innovative City Pilot Policy on green innovation using a novel framework combining a Multi-period Difference-in-Differences model and Deep Learning Counterfactual Prediction. Analyzing panel data from 100 eastern Chinese cities between 2004 and 2023, the research reveals that the policy significantly and robustly enhances innovation levels in pilot cities. Furthermore, the policy operates through a dual-track synergistic governance mechanism, successfully combining government scientific and technological support with environmental regulation. Additionally, heterogeneity analysis reveals that the policy exerts a more pronounced driving effect on green innovation in small-to-medium-sized cities and regions with lower industrial upgrading levels. Finally, deep learning counterfactual trajectories demonstrate that the policy dividend exhibits a non-linear, long-term cumulative effect that expands over time—a dynamic that traditional linear models often underestimate. Ultimately, this study provides solid empirical evidence that a combined governance system of incentives and constraints effectively promotes innovation-driven, sustainable urban transitions.

1. Introduction

Under the background of global climate governance, green technology innovation is widely regarded as a key pathway to achieving sustainable urban transitions and deep emission reductions [1]. Moving toward net-zero targets requires continuous low-carbon technological substitution, which necessitates policy systems that cover the entire “R&D–diffusion–application” chain [2]. Existing studies highlight that green innovation exhibits significant externalities and uncertainties, making it highly sensitive to policy interventions. Different types of policy instruments can significantly alter innovation incentives, presenting complex interactive relationships and transmission pathways [3,4].

China’s National Innovative City Pilot Policy (NICP) provides an important quasi-natural experiment to study this dynamic. By systematically embedding innovation capacity building into local governance, recent empirical evidence shows that NICP can promote green technological progress [5] and improve the innovation environment at the enterprise level [6]. Furthermore, research on policy synergy suggests that when NICP overlaps with other low-carbon policies, it generates a stronger promotion effect, indicating the effectiveness of combining “incentives and constraints” [7,8,9].

However, there are still three deficiencies in the existing literature. First, mechanism identification remains relatively “single-channel,” lacking a systematic examination of the “dual-track governance” logic—specifically, the parallel operation of government sci-tech support and environmental regulation. Second, the green innovation response may possess significant non-linear and long-term cumulative characteristics, meaning traditional linear models might underestimate long-term dividends.

Third, under the setting of staggered policy implementation, causal identification requires rigorous handling. Extensive econometric research has pointed out that traditional Two-Way Fixed Effects (TWFE) models may produce severe biases due to heterogeneous treatment effects and “forbidden comparisons” [10,11,12]. In response, the recent literature has developed various robust linear estimators to correct these dynamic biases [13,14,15], alongside more credible methods for parallel trend inference applied in recent urban pilot evaluations [16]. While these robust linear frameworks are highly valuable for causal identification, they fundamentally rely on strict linear parameter settings, making it difficult to fully capture the complex, non-linear dynamic evolutionary laws in macro-economic systems.

To overcome these limitations, this paper adopts a novel cross-methodology approach based on panel data from 100 eastern Chinese cities between 2004 and 2023. We utilize a multi-period Difference-in-Differences (DID) framework as the classical baseline for causal identification. Innovatively, we introduce a deep learning counterfactual prediction framework (LSTM/GRU) to address the potential dynamic biases of TWFE and accurately capture the non-linear treatment trajectories. The use of deep learning for counterfactual estimation is supported by recent advancements in longitudinal causal inference. For instance, Counterfactual Recurrent Networks (CRN) can estimate time-varying treatment effects through balanced representations [17], while Time Series Deconfounders utilize RNN architectures to mitigate hidden confounding [18]. Recent systematic reviews further validate this integration of causal inference and deep learning [19]. To address the credibility and “underspecification” challenges of modern machine learning [20], we combine these predictive models with strict structural constraints and cross-validate them against our robust DID estimates.

Through this framework, the marginal contributions of this paper are mainly reflected in four aspects: (1) Providing direct net policy effect evidence for the NICP on green technology innovation, aligning with sustainable transition goals; (2) Systematically constructing and testing the “dual-track governance mechanism” (government sci-tech support and environmental regulation); (3) Innovatively introducing deep learning counterfactuals to break through traditional linear constraints, thereby fully identifying the long-term, non-linear cumulative effects of the policy; (4) Conducting a multidimensional heterogeneity analysis based on city size and industrial structure to provide targeted empirical evidence for differentiated policy implementation.

2. Literature Review and Hypothesis Development

2.1. Literature Review

Green Technology Innovation (GTI) is widely regarded as the core engine to break the “zero-sum game” between economic growth and environmental protection, and to achieve dual-carbon goals and urban sustainable development [21]. However, GTI possesses ‘dual externalities’—the positive externalities of both technology spillovers and environmental benefits. Coupled with long R&D cycles and high investment risks, relying solely on market mechanisms often leads to severe resource misallocation and a lack of innovation motivation [22]. Existing literature exploring the driving factors of GTI mainly follows two distinct threads: The first is the environmental regulation perspective based on the “Porter Hypothesis.” Classical theories and a large number of recent cross-national empirical studies show that rationally designed environmental regulations (such as environmental tax laws, pollution assessments, etc.) can effectively stimulate the “innovation compensation” effect of enterprises, prompting resources to tilt toward eco-friendly technologies, thereby enhancing enterprises’ market competitiveness and total factor productivity in the long run [23,24,25]. The second is the financial and fiscal support perspective based on alleviating resource constraints. Studies have found that direct government technological subsidies, green credit, and the broad coverage of digital finance can significantly alleviate the internal financing constraints faced by enterprises, providing necessary liquidity support for long-cycle, high-risk green R&D [26,27,28].

Simultaneously, as a comprehensive spatial carrier for implementing the national innovation-driven development strategy, the macro-economic and environmental synergistic effects of the National Innovative City Pilot Policy (NICP) have attracted widespread academic attention in recent years. Existing studies have confirmed the positive role of this comprehensive pilot policy in enhancing urban total factor productivity (TFP) [27,29,30], promoting the advanced and rationalized industrial structure [31], improving urban land green utilization efficiency, and increasing green total factor energy efficiency [32,33].

With deepening research, some scholars have begun to focus on the direct causal relationship between the NICP and green innovation. The latest empirical evidence indicates that this policy can significantly increase the output scale and patent quality of green technology innovation by improving the overall urban innovation environment and promoting the spatial agglomeration of innovation factors (such as high-skilled talents and R&D funds) [34,35]. However, the existing literature still presents two distinct limitations when exploring policy dividends: First, regarding transmission mechanisms, previous studies mostly focused on single-dimensional spillover paths (such as merely increasing R&D investment or foreign direct investment drives), rarely placing them under a systematic deconstruction within the dual-track parallel collaborative governance framework of “government sci-tech support” and “government environmental regulation” [36,37]. Second, concerning the evaluation methods of policy effects, the vast majority of macro-policy evaluations rely heavily on the traditional multi-period Difference-in-Differences (DID) model. Although DID has a solid econometric foundation for causal identification, its strict linear parameter settings and parallel trend assumptions make it difficult to fully capture the complex, non-linear dynamic evolutionary laws in macro-economic systems [38,39]. In recent years, with the deep integration of Causal Inference and Machine Learning, using Deep Neural Networks (DNN) or other high-dimensional machine learning models to reconstruct counterfactual trajectories has become a frontier method for longitudinal causal evaluation and long-cycle complex policy effect prediction [40,41,42]. Therefore, this paper attempts to make up for the shortcomings in the above research dimensions by introducing the dual-track mechanism and the deep learning counterfactual framework.

2.2. Hypothesis Development

The NICP provides a macro-institutional guarantee for the research and development of green technologies by coordinating scientific and technological resources and optimizing the innovation environment. Based on this, the paper proposes the baseline hypothesis:

H1 (Baseline Effect):

The innovative city pilot policy significantly enhances the level of urban green technology innovation.

Green technology innovation is highly prone to market failure, thus requiring financial intervention from the government. After the implementation of this policy, local governments usually increase financial appropriations for scientific and technological innovation, forming a resource compensation effect. Multiple micro and macro empirical studies have pointed out that government-led fiscal sci-tech expenditures and green subsidies can effectively hedge against enterprise R&D risks and cross the “valley of death” of technological innovation [37]. Therefore, this paper proposes:

H2 (Mechanism 1: Government Sci-Tech Support):

The policy effectively alleviates the financial constraints of enterprise green R&D by increasing local government expenditures on science and technology, providing direct financial support for green innovation.

Besides positive resource incentives, pilot cities are often accompanied by stricter ecological assessments and a “green” orientation. According to the “Porter Hypothesis,” reasonably strong environmental regulations can stimulate the “innovation compensation” effect of enterprises. Innovative cities typically embed energy consumption and emission constraints into their assessment systems. This stringent institutional arrangement increases compliance costs for highly polluting enterprises, thereby forcing market entities to actively redirect capital toward green process improvements and clean technology R&D [21,23,24]. Therefore, this paper proposes:

H3 (Mechanism 2: Government Environmental Regulation):

The policy increases the compliance costs of highly polluting enterprises by strengthening the level of local environmental regulation (corresponding to the mechanism variable ln_er), thereby forcing enterprises to undergo green process improvements and clean technology transformations.

The formation of innovation networks and the accumulation of technological stock is a non-linear and long-term process. The innovative city pilot is not merely a short-term injection of funds, but a deep reshaping of the logic of urban governance, industrial structure, and talent structure. With the deepening of digital governance and the diffusion of knowledge spillover effects, policy dividends will break free from traditional linear decay laws, presenting significant long-memory characteristics and positive feedback loops [41,42]. Therefore, this paper proposes:

H4 (Long-Cycle Cumulative Effect Hypothesis):

The promotional effect of the national innovative city pilot policy on green technology innovation is not a short-term “pulse” stimulus but exhibits a non-linear, long-cycle cumulative effect that gradually strengthens over time.

3. Materials and Methods

3.1. Data Sources and Sample Selection

This paper selects 102 prefecture-level and above cities in eastern China from 2004 to 2023 as the initial research sample. Given that Sansha City and Danzhou City were established relatively late and have severe data deficiencies, they are excluded from this study, resulting in a final panel dataset of 100 cities. The data in this study are multi-sourced and detailed: the list of national innovative city pilots and their approval times are derived from official announcements by the National Development and Reform Commission (NDRC) and the Ministry of Science and Technology (MOST) of China; the core patent data measuring green technology innovation (including applications and grants) are obtained from the Chinese Research Data Services (CNRDS) Platform. To construct the mediating variable of government environmental regulation intensity, we utilized a reliable measurement method drawing on localized environmental regulation studies [43] and based on environmental protection word frequencies [44]. Specifically, using Python 3.13.9 web scraping technology to extract the annual government work reports of various municipalities, this paper draws on the text mining approach of relevant scholars to accurately measure local environmental regulation intensity. Specifically, Python is used to segment municipal government work reports and count the total frequencies of 15 core environmental regulation keywords, including ‘environmental protection’, ‘eco-friendly’, and ‘pollution’ (The complete list of all 15 core keywords and the exact calculation formula are detailed in Appendix A

Table A2. Keyword List and Measurement Formula for Environmental Regulation Intensity.

Category	Description/Details
	Data Source
Text Segmentation Tool	Python-based text segmentation libraries.
15 Core Keywords	(Environmental protection), (Environmental protection—abbreviated), (Pollution), (Energy consumption), (Emission reduction), (Pollution discharge), (Ecology), (Green), (Low carbon), (Air), (Chemical oxygen demand), (Sulfur dioxide), (Carbon dioxide), PM1, PM2.
Calculation Formula (Step 1)	ER = (Total word count of the reportTotal frequency of 15 keywords) × 100
Logarithmic Transformation (Step 2)	ln_er = ln(ER + 1) (Subjected to 1% bidirectional Winsorization) + 2

) [44]. The remaining macroeconomic control variables are extracted from the China City Statistical Yearbook and the CSMAR database.

In the data preprocessing stage, this paper draws on common practices for handling green innovation panel data in relevant literature, using logarithmic interpolation or linear interpolation to supplement a small number of missing variables, drawing on data handling practices in recent green innovation literature [25]. To eliminate the interference of extreme outliers on empirical estimation, a 1% upper and lower tail-trimming (Winsorize) is applied to all continuous variables. Since the core dependent variable, total green invention patent applications (ln_green_invention), and the difference-in-differences term (DID) have both been converted to logarithmic forms, the estimated elasticity coefficients inherently possess intuitive economic significance; therefore, rigorous winsorization based on logarithmic transformation further ensures the authenticity and robustness of the counterfactual analysis and empirical regression results [45]. The complete list of the sample cities, the list of national pilot cities, detailed missing rate statistics, and the specific calculation methods for environmental regulation are provided in the Supplementary Materials.

3.2. Variables Measurement

Dependent Variable: As shown in the Appendix A

Table A1. Variable Definitions and Data Processing Methods.

Category	Variable Name	Variable	Definition and Data Processing Method
Dependent Variable	Total green invention patent applications	ln_green_invention	[Core Dependent Variable] Original applications + 1, natural logarithm.
	Total green patent applications	ln_green_patents	[Robustness Check] Original applications + 1, natural logarithm.
	Total green patent grants	ln_green_grant	[Robustness Check] Original grants + 1, natural logarithm.
	Green invention patent grants	ln_green_inv_grant	[Robustness Check] Original grants + 1, natural logarithm.
	Raw green invention patents	green_invention_total	[Robustness Check] Retain original value, no logarithm applied.
Independent Variable	Difference-in-differences term	did	The net effect of the NICP, calculated by Treat × Post.
Control Variables	Economic development level	ln_pgdp	Per capita GDP, natural logarithm after linear interpolation for missing values.
	Financial deepening degree	fin_deep	Ratio of loan balance to GDP, winsorized, no logarithm.
	Human capital	ln_students	Number of enrolled students in higher education institutions, natural logarithm after interpolation and winsorized.
	Foreign capital dependence	open_fdi	Ratio of FDI to GDP, unified units, winsorized, no logarithm.
	City size	ln_pop	Total population at year-end, natural logarithm after interpolation and winsorized.
	Informatization level	ln_internet	Number of broadband internet users, + 1, natural logarithm after interpolation and winsorized.
	Industrial structure upgrading	ind_upg	Ratio of tertiary to secondary industry output value, winsorized, no logarithm.
	Environmental pollution	ln_so2	Industrial sulfur dioxide emissions, natural logarithm after linear interpolation.
Mediating Variables	Government sci-tech expenditure	ln_sci_exp	[Mechanism 1] Natural logarithm after interpolation and winsorized.
	Local environmental regulation intensity	ln_er	[Mechanism 2] Based on environmental word frequency proportion in government reports, + 1, natural logarithm and winsorized.

, the core dependent variable in this paper is the total number of green invention patent applications (ln_green_invention). Compared to overall patents with a broader scope, focusing on invention patents can more accurately filter out the noise of “strategic innovation,” thereby authentically and objectively reflecting the substantive level of a city’s high-quality green technology innovation [46].

The identification and classification of these green patents within the CNRDS database strictly adhere to the ‘IPC Green Inventory’ guidelines established by the World Intellectual Property Organization (WIPO). To ensure a rigorous and standardized filtration process that distinguishes authentic green technological innovations from non-green patents, the specific technological categories applied in this study are detailed in Appendix A

Table A3. WIPO IPC Green Inventory Classification Standards for Green Patents.

Main Category	Included Technological Sub-Fields	Example IPC Classes
1. Alternative Energy Production	Biofuels, Integrated gasification combined cycle (IGCC), Fuel cells, Wind energy, Solar energy, Geothermal energy, Hydro energy.	F03D, F24J, H01M, C10L
2. Transportation	Hybrid vehicles, Electric vehicles, Energy-efficient vehicle technologies, Non-fossil fuel propulsion systems.	B60K, B60L, B60W, Y02T
3. Energy Conservation	Energy storage systems, Efficient lighting (e.g., LEDs), Thermal insulation for buildings, Power supply circuitry optimization.	H01M, F21V, E04B, H02J
4. Waste Management	Air pollution control, Water and wastewater treatment, Solid waste disposal and recycling, Hazardous waste management.	B09B, C02F, F23G, B01D
5. Agriculture/Forestry	Alternative pesticides, Soil improvement, Energy-efficient agricultural machinery, Sustainable forestry management.	A01N, A01B, A01G
6. Administrative, Regulatory or Design Aspects	Environmental monitoring systems, Carbon emissions trading frameworks, Eco-design, Smart grid management.	G06Q, G01N, H02J
7. Nuclear Power Generation	Nuclear fusion reactors, Nuclear engineering safety systems, Reactor design and optimization.	G21B, G21C, G21D

.

The processing method is adding 1 to the original application volume and taking the natural logarithm. Additionally, this paper introduces total green patent applications (ln_green_patents), total green patent grants (ln_green_grant), green invention patent grants (ln_green_inv_grant), and the raw number of green invention patent applications (green_invention_total) as robustness check indicators.

Independent Variable: The difference-in-differences term (did) is used to measure the net effect of the innovative city pilot policy. This variable is calculated by multiplying the treatment group dummy variable (treat_group) and the policy time dummy variable (post). Because both the core dependent variable (ln_green_invention) and the DID term are in logarithmic forms, the regression coefficient of this core explanatory variable inherently possesses highly intuitive economic significance, clearly characterizing the semi-elasticity effect of the policy shock on green innovation [8].

Control Variables: To control for omitted variable bias as much as possible, this paper introduces a series of control variables, including economic development level (ln_pgdp), degree of financial deepening (fin_deep, measured by the ratio of loan balance to GDP and winsorized), human capital (ln_students), foreign capital dependence (open_fdi), city size (ln_pop), informatization level (ln_internet), industrial structure upgrading (ind_upg, calculated by the ratio of tertiary to secondary industry output value), and environmental pollution (ln_so2). Among them, financial deepening and foreign capital dependence are the core drivers determining regional capital flows, while human capital and informatization levels form the foundational base for technological innovation and absorption. When evaluating the net effect of urban macroeconomic policies on green innovation, strictly controlling for the above heterogeneous characteristics covering economic, demographic, and openness dimensions can effectively strip away the interference of external economic environments; this aligns highly with the control variable screening paradigm in mainstream international environmental economics empirical research [47]. Furthermore, using the ratio of tertiary to secondary industry output values to quantify industrial structure upgrading (ind_upg) can acutely capture the dynamic trend of urban industries transitioning towards service-orientation and cleaner production, which is a classic measurement for quantifying macroeconomic industrial structure optimization [48].

Mechanism Variables: These include government scientific and technological expenditure (ln_sci_exp) and local environmental regulation intensity (ln_er). On one hand, government sci-tech expenditure is used to measure the government’s technological support and financial guidance mechanism. Authoritative literature shows that local governments’ fiscal sci-tech expenditures can not only directly alleviate the financing constraints faced by innovation subjects but also exert a leverage effect to guide social capital into the R&D field, making it one of the core paths driving regional technological innovation [49]. On the other hand, to avoid the endogenous defects of traditional measurement indicators and accurately quantify local governments’ true environmental willingness, this paper measures environmental regulation intensity by calculating the proportion of environmental-related word frequencies in government work reports. This quantitative indicator based on text mining has been widely validated in academia as a reliable proxy variable for measuring government environmental attention and regulatory intensity [44].

3.3. DID Empirical Model

Considering that the NICP was implemented in batches, this paper constructs a Multi-period Difference-in-Differences (Multi-period DID) model to evaluate the Average Treatment Effect (ATE) of the policy on green technology innovation in eastern cities. The baseline regression model, which utilizes the traditional Two-Way Fixed Effects (TWFE) estimator, is specified as follows:

$$\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\mathrm{I}\mathrm{n}\mathrm{n}\mathrm{o}{\mathrm{v}}_{\mathrm{it}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1\mathrm{D}\mathrm{I}{\mathrm{D}}_{\mathrm{i}\mathrm{t}}+\sum {\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(1)

where GreenInnov_it represents the green technology innovation level of city i in year t (using ln_green_invention in the baseline regression); the core explanatory variable DID_it is a policy dummy variable, indicating whether city i has been established as an innovative pilot city in year t; Control_k,it is a series of control variables; μ_i represents city individual fixed effects, used to absorb inherent city characteristics that do not change over time; γ_t represents year fixed effects, used to control for macroeconomic cycle shocks; and ε_it is the random error term. The coefficient α₁ is the core focus of this paper; if it is significantly positive, it indicates that the policy effectively promoted green technology innovation.

To ensure the reliability of the empirical estimates, we conducted essential panel data diagnostics prior to regression. The Variance Inflation Factor (VIF) test yields a mean VIF of 2.03, with a maximum value of 4.09, which is well below the conventional threshold of 10, indicating no severe multicollinearity among the variables (detailed results are provided in Appendix A

Table A4. Variance Inflation Factor (VIF) Test Results.

Variable	VIF	1/VIF (Tolerance)
ln_internet	6.67	0.149931
ln_pgdp	5.06	0.197442
ln_pop	3.57	0.279945
ln_students	3.43	0.291795
fin_deep	3.14	0.318017
ind_upg	2.35	0.425381
ln_so2	2.08	0.48035
did	1.9	0.525053
open_fdi	1.49	0.670641

). Furthermore, given that our dataset is a macro-panel of cities spanning multiple years, the random error terms within the same city are likely to be serially correlated over time. To rigorously account for this intra-group correlation and potential heteroskedasticity, and to prevent the over-estimation of statistical significance, all baseline and robustness regressions in this study explicitly employ robust standard errors clustered at the city level.

3.4. Deep Learning Counterfactual Prediction Model

While traditional causal inference methods, such as the Synthetic Control Method (SCM) or Matrix Completion, are widely used in policy evaluation, they often rely on strict linear assumptions and struggle to capture long-term sequential dependencies in complex macro-panel data. Therefore, to overcome these limitations and break through the strict linear parameter settings of the traditional multi-period DID model, this paper innovatively introduces Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to construct a non-linear counterfactual framework based on PyTorch 2.12.0, as shown in Figure 1. To break through the strict linear parameter settings of the traditional multi-period DID model and to more accurately capture the complex non-linear dynamics and long-memory characteristics of time series among macroeconomic variables, this paper innovatively introduces Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to construct a non-linear counterfactual framework based on PyTorch, as shown in Figure 1. It is crucial to clarify the methodological boundaries of this study: the multi-period DID model serves as the core framework for rigorous causal identification. The LSTM and GRU deep learning models are not utilized as independent causal estimators; rather, they are employed to provide supportive descriptive evidence by forecasting a robust, non-linear baseline trajectory. This allows for a dynamic trajectory comparison to cross-validate the DID estimates and illustrate long-term trends. To ensure the reproducibility of this study, the complete Python source code for the LSTM and GRU networks, alongside the Stata code used for the DID empirical regressions, are available in the Supplementary Materials.

3.4.1. Data Windowing and Feature Construction

To enable the deep learning network to learn the time dependency of urban green technology innovation, this paper adopts the Sliding Window method to serialize the panel data. Setting the time window length to $\mathrm{w}=4$, for the prediction of city $\mathrm{i}$ at time $\mathrm{t}$, its input feature sequence ${\mathrm{X}}_{\mathrm{i}\mathrm{t}}$ contains the set of all control variables and mechanism variables over the past $\mathrm{w}$ periods:

$${\mathrm{X}}_{\mathrm{i}\mathrm{t}}=[{\mathrm{x}}_{\mathrm{i},\mathrm{t}-\mathrm{w}+1},{\mathrm{x}}_{\mathrm{i},\mathrm{t}-\mathrm{w}+2},\dots ,{\mathrm{x}}_{\mathrm{i},\mathrm{t}}]$$

(2)

where the feature vector ${\mathrm{x}}_{\mathrm{i},\mathrm{t}}$ contains standardized dimension indicators such as economic development level, financial deepening degree, and environmental regulation. The model is trained using a pure dataset consisting of the control group and pre-intervention samples to avoid Data Leakage.

3.4.2. Deep Learning Network Architecture and Hyperparameters Setting

Compared to traditional feedforward neural networks, LSTM and GRU can effectively avoid the vanishing or exploding gradient problems in long sequence training. As shown in

Table 1. Network Structure and Hyperparameter Settings for Deep Learning Models (LSTM/GRU).

Category	Parameter	Value	Purpose and Core Function
Network Structure Simplification	Time Window (seq_length)	4	Adapts to the short time-series characteristics of macro-panel data.
	Number of Layers (num_layers)	1	Discards deep structures to fundamentally reduce the model’s capacity for overfitting.
	Hidden Dimension (hidden_dim)	16	Significantly reduces dimensionality to limit the over-memorization of noise, enhancing training stability on short panels.
Regularization Constraints	Dropout Rate (dropout)	0.3	Introduces a 30% probability of neuron deactivation during the hidden state output phase to prevent overfitting.
	L2 Regularization Coefficient (weight_decay)	1 × 10−4	Strongly suppresses weight parameter inflation, serving as the core mechanism to prevent gradient explosion.
Training Strategy Optimization	Batch Size (batch_size)	16	Accommodates small sample characteristics, balancing computational efficiency with the accuracy of gradient updates.
	Initial Learning Rate (learning_rate)	0.0005	Adopts an extremely low learning rate to ensure more stable optimization steps during gradient descent.
	Maximum Epochs (num_epochs)	150	Sets an upper limit for training to provide sufficient space for loss function convergence.
	Early Stopping Patience (patience)	50	Monitors validation set performance to terminate training early and lock in the optimal generalized model weights.
Model Output	Counterfactual Prediction	ln_green_invention_it	Smoothly mapped through a fully connected layer to output the logarithmic value of green innovation for city i in period t.

, hyperparameters and network structures are set through the PyTorch framework.

As shown in the LSTM architecture in Figure 2, by introducing a Cell State and forget, input, and output gates, the model can independently determine the retention and discarding of historical economic feature information, thereby accurately capturing the long-term dynamic evolution laws of innovation [50].

Given the input sequence ${\mathrm{x}}_{\mathrm{t}}$ and the previous hidden state ${\mathrm{h}}_{\mathrm{t}-1}$, the internal information forward propagation mechanism of a single-layer LSTM with a hidden dimension set to 16 is as follows:

Forget Gate determines the information discarded from the cell state:

$${\mathrm{f}}_{\mathrm{t}}=\mathsf{\sigma }({\mathrm{W}}_{\mathrm{f}}\cdot [{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+{\mathrm{b}}_{\mathrm{f}})$$

(3)

Input Gate and candidate cell state jointly determine the writing of new information:

$${\mathrm{i}}_{\mathrm{t}}=\mathsf{\sigma }({\mathrm{W}}_{\mathrm{i}}\cdot [{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+{\mathrm{b}}_{\mathrm{i}})$$

(4)

$${\tilde{\mathrm{C}}}_{\mathrm{t}}=\tanh ({\mathrm{W}}_{\mathrm{C}}\cdot [{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+{\mathrm{b}}_{\mathrm{C}})$$

(5)

Cell State Update equation:

$${\mathrm{C}}_{\mathrm{t}}={\mathrm{f}}_{\mathrm{t}}\odot {\mathrm{C}}_{\mathrm{t}-1}+{\mathrm{i}}_{\mathrm{t}}\odot {\tilde{\mathrm{C}}}_{\mathrm{t}}$$

(6)

Output Gate and final hidden state generation:

$${\mathrm{o}}_{\mathrm{t}}=\mathsf{\sigma }({\mathrm{W}}_{\mathrm{o}}\cdot [{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+{\mathrm{b}}_{\mathrm{o}})$$

(7)

$${\mathrm{h}}_{\mathrm{t}}={\mathrm{o}}_{\mathrm{t}}\odot \tanh ({\mathrm{C}}_{\mathrm{t}})$$

(8)

where $\mathsf{\sigma }$ represents the Sigmoid activation function, $\odot$ is the Hadamard Product, and $\mathrm{W}$ and $\mathrm{b}$ are the weight matrices and bias terms the network needs to learn, respectively.

As shown in the GRU architecture in Figure 2, acting as an efficient variant of LSTM, GRU merges the cell state and hidden state and controls information flow through an Update Gate and a Reset Gate [51]. This simplified gating mechanism reduces the number of parameters, further enhancing the model’s generalization ability on short macro-panel data.

Its core state updates are controlled by the Update Gate and Reset Gate:

$${\mathrm{z}}_{\mathrm{t}}=\mathsf{\sigma }({\mathrm{W}}_{\mathrm{z}}\cdot [{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+{\mathrm{b}}_{\mathrm{z}})$$

(9)

$${\mathrm{r}}_{\mathrm{t}}=\mathsf{\sigma }({\mathrm{W}}_{\mathrm{r}}\cdot [{\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+{\mathrm{b}}_{\mathrm{r}})$$

(10)

The candidate hidden state and final hidden state update formulas are:

$${\tilde{\mathrm{h}}}_{\mathrm{t}}=\tanh (\mathrm{W}\cdot [{\mathrm{r}}_{\mathrm{t}}\odot {\mathrm{h}}_{\mathrm{t}-1},{\mathrm{x}}_{\mathrm{t}}]+\mathrm{b})$$

(11)

$${\mathrm{h}}_{\mathrm{t}}=(1-{\mathrm{z}}_{\mathrm{t}})\odot {\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{z}}_{\mathrm{t}}\odot {\tilde{\mathrm{h}}}_{\mathrm{t}}$$

(12)

To robustly suppress model overfitting and ensure high generalization ability, we implemented a strict combination of structural constraints and training strategies, as detailed in the hyperparameter settings in Table 1. Specifically, in the final stage of feature extraction, we introduce a Dropout layer with a 30% dropout rate after the network extracts the hidden state at the last time step. Furthermore, we apply an L2 Regularization coefficient (weight decay) of 1 × 10⁻⁴ to prevent gradient explosion, and utilize an Early Stopping mechanism with a patience of 50 epochs to continuously monitor the validation set and lock in the optimal generalized model weights [52]. Subsequently, through a Fully Connected Layer, high-dimensional hidden features are mapped back to a one-dimensional space, outputting the target variable—the counterfactual logarithmic predicted value of green technology innovation level (${\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_{\mathrm{i}\mathrm{t}}$):$\underset{\mathrm{i}\mathrm{t}}{\widehat{\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}}={\mathrm{W}}_{\mathrm{f}\mathrm{c}}\cdot \mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}({\mathrm{h}}_{\mathrm{T}})+{\mathrm{b}}_{\mathrm{f}\mathrm{c}}$.

3.4.3. Estimation of Counterfactual ATT

After the model has been fully optimized via the training and validation sets, this paper inputs the real macro characteristics of the experimental group cities post-policy implementation into the optimal model to obtain the counterfactual innovation level ${\widehat{\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}}_{\mathrm{i}\mathrm{t}}^{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}}$, assuming the “innovative city pilot policy was not implemented.”

Ultimately, the individual Average Treatment Effect on the Treated (ATT) for city $\mathrm{i}$ in period $\mathrm{t}$ is defined as the difference between the actual observed logarithmic value of green invention patent applications ($\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}^{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l}}$) and the counterfactual predicted value:

$$\mathrm{A}\mathrm{T}{\mathrm{T}}_{\mathrm{i}\mathrm{t}}=\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}^{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l}}-{\widehat{\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}}_{\mathrm{i}\mathrm{t}}^{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}}$$

(13)

Building upon this, by averaging the treatment effects across periods for all experimental group cities after policy implementation (i.e., relative years $\mathrm{t}\in \left[\mathrm{0,5}\right]$), the overall average treatment effect from the deep learning perspective is derived:

$$\mathrm{A}\mathrm{T}{\mathrm{T}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}}={\displaystyle \frac{1}{{\mathrm{N}}_{\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}\times {\mathrm{T}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}}}}\sum _{\mathrm{i}\in \mathrm{t}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}\sum _{\mathrm{t}=0}^5(\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}^{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l}}-{\widehat{\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}}_{\mathrm{i}\mathrm{t}}^{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}})$$

(14)

3.4.4. Model Training and Convergence Validation

As shown in Figure 3, both the Train Loss and Validation Loss of the LSTM model show a rapid downward trend during the initial iteration phase, indicating that the network is effectively capturing the non-linear mapping relationship between macroeconomic features and green innovation. As the number of training epochs increases, the two curves quickly flatten out and stably converge to an extremely low level of around 0.1. It is worth noting that the validation loss curve always closely follows the training loss curve, and no rebound phenomenon occurs throughout the entire training cycle. The GRU model achieved the same effect after testing.

4. Main Empirical Results

4.1. Baseline Regression Analysis

Column 1 of

Table 2. Baseline Regression Analysis.

Variable	ln_green_invention	ln_green_invention	Year-by-Year PSM-DID	Random Placebo Check
did	0.3630 ***	0.3895 ***	0.3697 ***	−0.0772
	−0.0744	−0.0702	−0.0666	−0.0804
ln_pgdp		0.7056 ***	0.7292 ***	0.7506 ***
		−0.1372	−0.1364	−0.1455
fin_deep		0.1862 ***	0.1906 ***	0.2139 ***
		−0.0571	−0.0574	−0.0605
ln_students		0.1622	0.1664	0.1395
		−0.1003	−0.1097	−0.1042
open_fdi		7.0358	5.3793	3.9258
		−5.5205	−5.7761	−5.6716
ln_pop		0.7575 ***	0.7703 ***	0.8176 ***
		−0.1792	−0.1717	−0.1976
ln_internet		0.1292 **	0.1359 **	0.0814
		−0.0638	−0.0667	−0.058
ind_upg		0.057	0.0592	0.0587
		−0.0799	−0.0822	−0.0743
ln_so2		0.0258	0.0305	−0.0008
		−0.0411	−0.0416	−0.0391
City FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
N	2000	2000	2474	2000
Adj. R2	0.9495	0.9541	0.9506	0.9518

Note: Robust standard errors in parentheses; **, *** denote significance at the 5% and 1% statistical levels, respectively.

shows that when only two-way fixed effects are controlled, the coefficient of the core explanatory variable (did) is 0.3630 (p < 0.01). After further introducing the full set of control variables in Column 2, the estimated coefficient of did stably increases to 0.3895, still maintaining a high significance at the 1% level. This core result intuitively indicates that the NICP has a substantive positive driving effect on the output of green invention patents. This finding not only verifies the baseline hypothesis (H1) of this paper but also directly corroborates the empirical findings of scholars such as Zhang et al. [6] and Cui et al. [7]—that is, this comprehensive pilot policy can effectively improve the urban innovation environment and promote the spatial agglomeration of innovation factors. Meanwhile, the net effect of policy-driven high-quality green innovation also provides the latest support from China’s macro-urban level for the classic “Porter Hypothesis” and its derivative studies [23,24,25], proving that a rationally designed urban innovation and environmental governance system can effectively stimulate the compensation effect of technological innovation.

To alleviate sample self-selection bias caused by the non-random selection of policy pilot cities, Column 3 reports the regression results based on the year-by-year Propensity Score Matching (PSM-DID) method. The test results effectively exclude the interference of initial endowment differences and macro random noise, powerfully confirming from the reverse angle that the significant positive effect in the baseline regression indeed stems from the causal shock of the policy itself, ensuring the absolute reliability of the policy evaluation conclusions in this paper.

4.2. Parallel Trend Test

The prerequisite for the validity of the multi-period DID model is that the treatment group and the control group satisfy the parallel trend assumption. Drawing on the latest standards in econometrics regarding multi-period DID event study methodology, this paper rigorously tests the dynamic effects of the policy [12,13].

Specifically, the event-study plot presented in Figure 4 visually reports the dynamic treatment effects (annual ATTs) alongside their 95% confidence intervals. The plot clearly shows that before the policy implementation (periods −5 to −2), the estimated coefficients of the dummy variables in each period fluctuate around the 0 axis and lack statistical significance (e.g., the coefficient for period −2 is only −0.0366, p = 0.404). This indicates that there was no systematic difference in the evolutionary trend of green technology innovation between the two groups of cities prior to the intervention, strictly satisfying the parallel trend assumption.

Furthermore, the coefficient leaps significantly to 0.3778 in the year of policy implementation (period 0) and remains significantly positive at the 1% level in all subsequent periods (periods 1 to 5), preliminarily confirming the positive incentive effect of the policy. To further break through traditional linear parameter constraints and more accurately characterize the long-term dynamic evolutionary law of policy dividends, this paper conducts an in-depth cross-comparison combining the deep learning counterfactual prediction model in Section 4.3.

4.3. Cross-Comparison of Dynamic Treatment Effect Trajectories

To verify the reliability of the traditional multi-period DID estimation results, mine non-linear dynamic characteristics, and further validate the post-treatment dynamic estimates, after dropping period −1, this paper conducts an intertemporal cross-comparison between the annual Average Treatment Effect (ATT) predicted by LSTM and GRU and the DID dynamic coefficients.

As shown in Figure 5, before the policy implementation (periods −5 to −2), the ATTs calculated by LSTM and GRU both fluctuate slightly around the 0 axis (e.g., 0.0369 and 0.0534, respectively, in period −2), highly overlapping with the DID pre-treatment coefficient (−0.0366) and entirely falling within the 95% confidence interval of DID. This rigorously confirms the parallel trend assumption once again from a non-linear machine learning perspective.

In the year of policy implementation (period 0), the core estimated value of DID is 0.3778, while the ATTs based on LSTM and GRU predictions are 0.3559 and 0.4176, respectively. The three converge highly in numerical magnitude and direction, confirming the objective existence of immediate policy dividends and excluding the possibility of “statistical illusions” in linear models.

Over the long cycle after policy implementation (periods 1 to 5), the ATT estimated by DID shows a steady positive impact, reaching 0.3613 in period 5. In contrast, deep learning models demonstrate outstanding long-memory capture capabilities, with the ATTs of LSTM and GRU climbing to 0.5174 and 0.7066, respectively, in period 5. This indicates that traditional DID, limited by linear assumptions, may somewhat underestimate the non-linear green innovation growth potential stimulated by this policy over a long cycle.

5. Robustness Checks

5.1. Placebo Test

To eliminate the interference of omitted variables and accidental factors, this paper conducts a placebo test by randomly drawing a “pseudo-treatment group” matching the number of actual treated units, assigning pseudo-policy years (i.e., fake treatment years) to construct a fake difference-in-differences variable (fake_did), and repeating the regression 500 times [53]. As shown in Figure 6, the 500 randomly sampled fake estimated coefficients overall present a normal distribution with a mean around 0, and the vast majority of p-values are greater than 0.1, indicating that the randomly assigned pseudo-policy did not produce a significant impact. The result from one specific random seed has been reported in Table 2. Meanwhile, the actual baseline policy estimated coefficient (0.3895) acts as an extreme value, significantly lying outside the distribution range of the fake coefficients. This test result effectively rules out the possibility that the baseline conclusions are driven by accidental factors or omitted variables, corroborating the authenticity and high robustness of the innovative city pilot policy’s promotional effect on green technology innovation [6,7].

5.2. Propensity Score Matching DID (PSM-DID)

To ensure the validity of the DID model estimation and alleviate sample self-selection bias, this paper utilizes the year-by-year Propensity Score Matching (PSM) method to conduct a robustness check, with the results reported in Table 2. Taking the balance test results from the year before the policy implementation as an example, as shown in

Table 3. Sample Balance Test Results Before and After Propensity Score Matching (PSM).

Variable/Overall Indicator	p-Value Before Matching	p-Value After Matching
ln_pgdp	0.000 ***	0.508
fin_deep	0.046 **	0.678
ln_students	0.000 ***	0.775
open_fdi	0.001 ***	0.634
ln_pop	0.000 ***	0.643
ln_internet	0.000 ***	0.499
ind_upg	0.184	0.816
ln_so2	0.002 ***	0.711
Overall Test (p > chi2)	0.000 ***	0.783

Note: **, *** denote significance at the 5% and 1% confidence levels, respectively.

, prior to matching, the treatment and control groups exhibited significant systematic differences across most control variables, and the overall joint test was highly significant (p = 0.000). However, after PSM, the mean differences for all control variables were no longer significant (p-values all far greater than 0.1). Simultaneously, the p-value of the overall joint test jumped to 0.783, indicating that the covariates no longer possessed joint significance after matching [54]. The above results demonstrate that the matched sample achieved good characteristic balance, satisfying the comparability premise, thus providing a reliable data foundation for further verifying the robustness of the baseline regression results.

5.3. Alternative Dependent Variables

As shown in

Table 4. Robustness Checks for the Baseline Empirical Results.

Variable	(1) Total Applications	(2) Total Grants	(3) Invention Grants	(4) Unlogged	(5) Excluding Pandemic	(6) Lagged One Period
did	0.2555 *** (0.0604)	0.2795 *** (0.0631)	0.7206 *** (0.0877)	800.2473 *** (153.8474)	0.4599 *** (0.0847)
did_l1						0.2957 *** (0.0677)
ln_pgdp	0.6268 *** (0.1329)		0.6102 *** (0.1814)	1022.9042 ** (425.6994)	0.5237 *** (0.1530)	0.7328 *** (0.1400)
fin_deep	0.1266 *** (0.0481)		0.6569 *** (0.1515)	417.3886 (271.6065)	0.1678 ** (0.0782)	0.1376 ** (0.0566)
ln_students	0.1103 (0.0836)	0.0902 (0.0974)	0.1568 *** (0.0539)	−799.5058 *** (243.1323)	0.1989 (0.1380)	0.1230 (0.0989)
open_fdi	3.5458 (4.7634)	2.2558 (4.6312)	7.2021 (5.9444)	50,600.00 * (27,800.00)	−2.7275 (7.0102)	6.6797 (5.4778)
ln_pop	0.6666 *** (0.1402)	0.8389 *** (0.1623)	1.1472 *** (0.2396)	4364.5510 ** (1817.7134)	1.3822 ** (0.6726)	0.6309 *** (0.1783)
ln_internet	0.1680 *** (0.0589)	0.1677 *** (0.0588)	0.1314 * (0.0726)	−792.2453 ** (307.3246)	0.1226 * (0.0714)	0.1070 * (0.0629)
ind_upg	0.1035 * (0.0605)	0.0522 (0.0812)	−0.2057 ** (0.1004)	1412.4220 ** (583.8192)	−0.1108 (0.1063)	0.0973 (0.0800)
ln_so2	0.0306 (0.0337)	0.0195 (0.0375)	−0.0278 (0.0449)	−423.9329 *** (153.2939)	−0.0259 (0.0487)	0.0522 (0.0421)
_cons	−0.8618 (0.9939)	−2.3546 ** (1.1659)	−5.8393 *** (1.7068)	−24,600.00 ** (11,100.00)	−5.9096 (4.0654)	−1.5032 (1.1929)
Observations N	2000	2000	2000	2000	1600	1900
Adj. R-squared	0.9675	0.9642	0.9344	0.7598	0.9485	0.9552
City FE	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes

Note: Robust standard errors clustered at the city level in parentheses; ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.

, this paper uses “total green patent applications” (including inventions and utility models, Column 1), “total green patent grants” (Column 2), and “green invention patent grants”—which most represent substantive technological breakthroughs (Column 3)—as alternative variables for robustness checks. The coefficients for the core policy explanatory variable did are 0.2555, 0.2795, and 0.7206, respectively, all of which are significantly positive at the 1% level. This indicates that regardless of whether viewed from the “activeness” of innovation (total applications) or the “substantive quality” of innovation (grants), the innovative city pilot policy has played a significant promotional role. It further proves that the policy successfully avoided the trap of prioritizing “quantity over quality,” substantively driving high-quality, hard-core green technological breakthroughs in cities [46].

To eliminate potential measurement errors caused by logarithmic processing, Column 4 of Table 4 directly uses the non-logarithmic level value of “total green invention patent applications” (green_invention_total) for regression. The results show that the coefficient of did is 800.2473 (p < 0.01), meaning that on average, policy implementation brought an additional increment of approximately 800 green invention patents per year to the pilot cities. This result once again corroborates the high robustness of the baseline conclusions.

Excluding Exogenous Macroeconomic Shocks: In panel data for policy evaluation, macroeconomic extreme events often trigger structural breaks [55]. Considering that the outbreak of the COVID-19 pandemic in 2020 comprehensively shocked the real economy operations, enterprise R&D investments, and administrative approval efficiency in eastern Chinese cities, Column 5 excludes the sample data from 2020 and beyond. The regression results show that after excluding the pandemic years, the estimated coefficient of did is 0.4599, remaining significantly positive at the 1% level, proving that the innovation incentive effect is caused by the policy’s endogenous mechanism rather than being an accidental result of external sudden shocks.

Although policy evaluation possesses certain quasi-natural experiment characteristics, when establishing innovative pilot cities, the state might be inclined to select eastern cities that already have a strong innovation foundation, raising endogeneity concerns caused by reverse causality [14]. Therefore, this paper re-runs the regression using a one-period lagged policy dummy variable (did_l1). The results in Column 6 show that the coefficient of the lagged core explanatory variable is 0.2957 (p < 0.01), which is highly consistent with the baseline regression conclusion, effectively alleviating the interference of reverse causality on this study.

Leave-one-out Sensitivity Test: To rule out the possibility that the baseline results are driven by extreme outlier cities, we systematically excluded one city at a time from the sample and re-estimated the baseline model. The estimated coefficients of the core policy variable across the 100 iterations remain highly stable, ranging from 0.3392 to 0.4583, with a mean value of 0.3895. All estimates tightly cluster around our original baseline coefficient (0.3895) and remain statistically significant, as visually demonstrated in Appendix A Figure A1. This confirms that our causal estimates are highly robust and not overly reliant on any specific localized sample or accidental extreme values.

6. Mechanism Analysis

6.1. The Path of Government Support

Green technology innovation generally faces issues of long R&D cycles, high risks, and “market failures” caused by positive externalities, often requiring government fiscal intervention to compensate for corporate R&D costs. To test the resource compensation path of “government sci-tech support,” this paper constructs the following three-step mediation effect testing model:

First, testing the baseline total effect of the policy on the dependent variable (i.e., the baseline regression in Table 2):

$$\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+\sum {\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(15)

Second, testing the impact of the policy shock on the mediating variable “sci-tech expenditure” ($\mathrm{l}\mathrm{n}\_\mathrm{s}\mathrm{c}\mathrm{i}\_\mathrm{e}\mathrm{x}{\mathrm{p}}_{\mathrm{i}\mathrm{t}}$) (corresponding to Column 1 in

Table 5. Empirical Test Results of the Dual-track Synergistic Mechanism.

Variable	(1) Sci-Tech Exp (M1)	(2) Adding M1	(3) Env. Reg. (M2)	(4) Adding M2	(5) Comprehensive Model
did	0.5895 *** (0.0789)	0.2320 *** (0.0608)	0.1107 *** (0.0163)	0.3580 *** (0.0696)	0.1895 *** (0.0609)
ln_sci_exp		0.2671 *** (0.0321)			0.2715 *** (0.0314)
ln_er				0.2840 ** (0.1241)	0.3602 *** (0.1196)
ln_pgdp	1.5759 *** (0.1876)	0.2847 ** (0.1416)	0.0644 ** (0.0270)	0.6873 *** (0.1367)	0.2545 * (0.1399)
fin_deep	0.4153 *** (0.0749)	0.0753 (0.0521)	−0.0272 ** (0.0111)	0.1939 *** (0.0571)	0.0832 (0.0520)
ln_students	0.0935 (0.0978)	0.1373 (0.0898)	−0.0028 (0.0125)	0.1630 (0.1003)	0.1379 (0.0891)
open_fdi	15.5747 *** (4.8891)	2.8762 (5.0120)	−0.6609 (0.6538)	7.2234 (5.5158)	3.0447 (5.0000)
ln_pop	1.4415 *** (0.2946)	0.3726 ** (0.1479)	−0.1456 *** (0.0331)	0.7989 *** (0.1842)	0.4186 *** (0.1525)
ln_internet	0.1608 ** (0.0701)	0.0862 (0.0568)	0.0318 *** (0.0093)	0.1201 * (0.0628)	0.0740 (0.0556)
ind_upg	0.0909 (0.1576)	0.0327 (0.0697)	0.0277 (0.0170)	0.0491 (0.0792)	0.0223 (0.0679)
ln_so2	−0.0524 (0.0476)	0.0398 (0.0351)	−0.0128 ** (0.0065)	0.0294 (0.0411)	0.0446 (0.0350)
Observations N	2000	2000	2000	2000	2000
Adj. R-squared	0.9438	0.9575	0.6733	0.9542	0.9578
City FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes

Note: Robust standard errors clustered at the city level in parentheses; ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.

):

$$\mathrm{l}\mathrm{n}\_\mathrm{s}\mathrm{c}\mathrm{i}\_\mathrm{e}\mathrm{x}{\mathrm{p}}_{\mathrm{i}\mathrm{t}}={\mathsf{\theta }}_0+{\mathsf{\theta }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+\sum {\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(16)

Third, incorporating both the policy variable and sci-tech expenditure into the model to test their joint impact on the dependent variable (corresponding to Column 2 in Table 5):

$$\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathrm{\lambda }}_0+{\mathrm{\lambda }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+{\mathrm{\lambda }}_2\mathrm{l}\mathrm{n}\_\mathrm{s}\mathrm{c}\mathrm{i}\_\mathrm{e}\mathrm{x}{\mathrm{p}}_{\mathrm{i}\mathrm{t}}+\sum {\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(17)

Column 1 of Table 5 explores the policy’s impact on local government sci-tech expenditure; the result shows the coefficient ${\mathsf{\theta }}_1$ for the core explanatory variable did is 0.5895 and significant at the 1% level. This indicates that the NICP effectively guided local governments to increase fiscal appropriations and tilt resources towards sci-tech innovation. In Column 2, the coefficient ${\mathsf{\lambda }}_2$ for sci-tech expenditure (ln_sci_exp) is 0.2671 and significantly positive at the 1% level. More importantly, the coefficient ${\mathsf{\lambda }}_1$ for the core explanatory variable did dropped from 0.3895 in the baseline regression to 0.2320, while still maintaining significance. This series of results verifies the partial mediating effect of “government sci-tech support”. This confirms hypothesis H2: the NICP effectively alleviates the financial constraints of green innovation by increasing local sci-tech expenditure, thereby promoting the development of green technology.

6.2. The Path of Environmental Regulation

In addition to financial support, environmental regulation is a vital means to force green economic transition. According to the “Porter Hypothesis,” reasonable and strict environmental regulations can stimulate the “innovation compensation” effect of enterprises. Similarly, to test the path of environmental institutional forcing, this paper constructs the following three testing equations:

First, baseline total effect test (same as above):

$$\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+\sum {\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(18)

Second, testing the impact of the policy shock on the mediating variable “environmental regulation” ($\mathrm{l}\mathrm{n}\_\mathrm{e}{\mathrm{r}}_{\mathrm{i}\mathrm{t}}$) (corresponding to Column 3 in Table 5):

$$\mathrm{l}\mathrm{n}\_\mathrm{e}{\mathrm{r}}_{\mathrm{i}\mathrm{t}}={\mathsf{\varphi }}_0+{\mathsf{\varphi }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+\sum {\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(19)

Third, incorporating both the policy variable and environmental regulation into the model (corresponding to Column 4 in Table 5):

$$\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathsf{\delta }}_0+{\mathsf{\delta }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\delta }}_2\mathrm{l}\mathrm{n}\_\mathrm{e}{\mathrm{r}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(20)

$\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}$: The regression results in Column 3 of Table 5 show that the estimated coefficient ${\mathsf{\varphi }}_1$ of did is 0.1107 (p < 0.01), indicating that the innovative city pilot not only focuses on economic indicators but also significantly enhances the local environmental regulation intensity of pilot cities. In Column 4, when environmental regulation (ln_er) is introduced into the main regression model, its coefficient ${\mathsf{\delta }}_2$ is significantly positive at the 5% level (0.2840). Simultaneously, the coefficient ${\mathsf{\delta }}_1$ of the policy variable did remains highly significant (0.3580). This confirms hypothesis H3, illustrating the existence of the “government environmental regulation” path. The environmental assessment pressure of innovative cities is effectively transformed into the motivation for enterprises to conduct green process improvements and clean technology R&D, forcing micro-entities to evade environmental compliance costs through green innovation.

6.3. Comprehensive Analysis of Dual-Track Synergistic-Driven Network

To test whether the dual mechanisms can function in parallel without conflict, this paper simultaneously incorporates the two major mechanism variables, “sci-tech expenditure” and “environmental regulation,” into the model in Column 5 of Table 5 for comprehensive testing. The corresponding comprehensive testing equation is as follows:

$$\mathrm{l}\mathrm{n}\_\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\_\mathrm{i}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}\mathrm{t}}={\mathsf{\rho }}_0+{\mathsf{\rho }}_1\mathrm{d}\mathrm{i}{\mathrm{d}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\rho }}_2\mathrm{l}\mathrm{n}\_\mathrm{s}\mathrm{c}\mathrm{i}\_\mathrm{e}\mathrm{x}{\mathrm{p}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\rho }}_3\mathrm{l}\mathrm{n}\_\mathrm{e}{\mathrm{r}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\beta }}_{\mathrm{k}}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}{\mathrm{l}}_{\mathrm{k},\mathrm{i}\mathrm{t}}+{\mathsf{\mu }}_{\mathrm{i}}+{\mathrm{\gamma }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(21)

The regression results with comprehensive inclusion are highly explanatory: the coefficient ${\mathsf{\rho }}_2$ for ln_sci_exp is 0.2715 (p < 0.01), and the coefficient ${\mathsf{\rho }}_3$ for ln_er is 0.3602 (p < 0.01), both maintaining extremely high statistical significance. At the same time, the coefficient ${\mathsf{\rho }}_1$ of the core policy dummy variable did is further absorbed, decreasing to 0.1895 (p < 0.01). This evidence perfectly delineates the complete logical map of innovative city construction: the policy does not rely on a single dimension of stimulus but cleverly constructs a dual-track synergistic mechanism featuring both “government sci-tech support (the carrot)” and “government environmental regulation (the stick)” in parallel. The two mechanisms complement each other, jointly promoting the leap in green technology innovation levels in eastern cities.

Furthermore, a comparative analysis of the marginal effects provides specific references for practical implementation. In the comprehensive model, the marginal elasticity of environmental regulation (0.3602) is slightly higher than that of government sci-tech support (0.2715). This indicates that in the current implementation of the dual-track mechanism, the “stick” of environmental compliance pressure yields a slightly stronger marginal return in forcing green technological breakthroughs than the “carrot” of financial subsidies. Therefore, finding an optimal dynamic balance that slightly leans toward strict environmental constraints while ensuring sufficient financial compensation is crucial for maximizing the policy’s effectiveness.

7. Heterogeneity Analysis

To further explore the driving effects of the innovative city pilot policy on green technology innovation across cities with distinct characteristics, this study conducts a heterogeneity analysis based on city size and industrial structure. The detailed regression results are presented in

Table 6. Heterogeneity Analysis of City Size and Industrial Structure.

Variable	(1) Large Cities	(2) Small Cities	(3) High Ind Upg	(4) Low Ind Upg
did	0.3491 ***(0.0886)	0.4635 *** (0.1226)	0.3110 *** (0.0910)	0.4491 *** (0.0963)
ln_pgdp	0.4578 * (0.2446)	1.0188 *** (0.2520)	1.2599 *** (0.2366)	0.3238 (0.2081)
fin_deep	0.2831 ** (0.1132)	0.1169 ** (0.0543)	0.2858 * (0.1444)	0.1705 ** (0.0691)
ln_students	0.3546 * (0.1873)	−0.1105 (0.1068)	−0.0631 (0.1106)	0.2078 * (0.1161)
open_fdi	10.3688 (7.3698)	−13.0641 ** (6.1360)	−0.1982 (7.6200)	9.0717 (6.5413)
ln_internet	0.1417 (0.0927)	0.1361 * (0.0768)	0.1583 (0.1017)	0.1147 * (0.0632)
ind_upg	0.1186 (0.1042)	0.0401 (0.1049)
ln_pop			0.8174 *** (0.2248)	0.7601 *** (0.1953)
ln_so2	0.0805 (0.0526)	0.0248 (0.0428)	0.0764 * (0.0436)	0.0075 (0.0475)
_cons	−9.9482 ** (3.8631)	−1.1687 (2.4283)	−11.0558 *** (3.1238)	−8.1009 *** (2.6074)
N	1000	1000	1000	1000

Note: Robust standard errors clustered at the city level in parentheses; ***, **, * denote significance at the 1%, 5%, and 10% levels, respectively.

.

7.1. Heterogeneity of City Size

Following common empirical practices, the sample is divided into “large cities” and “small-to-medium cities” based on the median of the urban population logarithm (ln_pop). As reported in Columns (1) and (2) of Table 6, the coefficient of the core explanatory variable did is significantly positive in both sub-samples. Notably, the marginal incentive effect in small-to-medium cities (0.4635) is substantially greater than that in large cities (0.3491). A plausible explanation is that large cities generally possess an already mature innovation environment and abundant R&D resources; thus, the policy functions merely as the “icing on the cake.” Conversely, for small-to-medium cities facing stronger constraints in innovation elements, the policy acts as “snow in summer,” effectively bridging the resource gap and triggering a higher marginal elasticity of green innovation growth.

7.2. Heterogeneity of Industrial Structure

Furthermore, the sample is split into “high industrial upgrade” and “low industrial upgrade” groups based on the median of the industrial structure upgrading index (ind_upg). The results in Columns (3) and (4) of Table 6 indicate that the policy’s driving effect is significantly more pronounced in cities with a lower degree of industrial upgrading (coefficient of 0.4491 vs. 0.3110). This finding perfectly corroborates the “Porter Hypothesis” discussed in our earlier mechanism analysis. Cities with a lower degree of industrial upgrading (i.e., those heavily reliant on secondary industries and traditional manufacturing) face the brunt of environmental assessment pressures under the comprehensive pilot policy. This stringent environmental regulation framework forces highly polluting or energy-intensive local enterprises to actively seek green technological breakthroughs to offset compliance costs, thereby generating a stronger and more immediate “innovation compensation effect”.

8. Deep Learning Counterfactual Analysis and Robustness

8.1. Trajectory Validation of Cumulative Innovation Effect

More importantly, Figure 7 clearly demonstrates that after the policy intervention (relative years 0 to 15), the predictive deviation (ATT) between the actual innovation output and the deep learning counterfactual baseline did not maintain a fixed constant level but exhibited a monotonically increasing divergence. As quantitatively demonstrated in our cross-comparison of dynamic treatment trajectories (see Section 4.3 and Figure 5), the initial policy shock in period 0 yielded an ATT of 0.3778 under the dynamic DID framework, closely matching the LSTM (0.3559) and GRU (0.4176) estimates. However, over the long cycle (by period 5), this quantitative gap expanded significantly. While the dynamic DID estimated an ATT of 0.3613, the deep learning models, which are free from linear decay constraints, quantified the cumulative ATT at 0.5174 (LSTM) and 0.7066 (GRU). This explicitly cross-validated, expanding quantitative divergence provides rigorous statistical evidence—beyond mere visual trends—that the national innovative city pilot policy is not a short-term stimulus that quickly exhausts its potential. Instead, it fosters the accumulation of knowledge stock and the agglomeration of factors over a long cycle, generating a statistically significant cumulative effect on green innovation.

8.2. Random Seed Robustness Check

As visually confirmed by the distributions in Figure 8, to eliminate the accidental interference of random initialization of network weights, this paper, based on relevant methods, conducted 100 independent training and counterfactual inference runs with different random seeds for both LSTM and GRU [20]. The ATT kernel density distributions are concentrated and present normal distributions, with all predicted values greater than 0 (mainly concentrated between 0.15 and 0.55), which robustly confirms that the NICP possesses a true positive promotional effect. The ATT means of LSTM (0.4094) and GRU (0.4215) are not only highly consistent with the baseline estimated value of the multi-period DID (0.3895) but are also numerically slightly higher, cross-validating the reliability of the DID causal estimates via predictive trajectory comparison, and implying that linear models may underestimate the long-term non-linear innovation dividends. Over the 100 independent runs, the average Root Mean Square Error (RMSE) for LSTM and GRU was approximately 0.43, the average Mean Absolute Percentage Error (MAPE) was maintained at around 7.8%, and the average coefficient of determination (R²) for both exceeded 0.78, indicating that the models possess extremely high and stable generalization prediction performance. The detailed evaluation metrics and output data for all 100 independent runs (LSTM_100seeds Runs and GRU_100seeds Runs) are accessible in the Supplementary Materials.

9. Conclusions and Policy Implications

9.1. Main Conclusions

This paper focuses on a refined panel dataset of cities in eastern China, comprehensively utilizing a multi-period Difference-in-Differences (DID) model and deep learning counterfactual prediction networks (LSTM/GRU) to fully deconstruct the net impact effect and internal logic of the National Innovative City Pilot Policy on green technology innovation. The core findings are as follows:

Significant baseline promotional effect: The NICP indeed significantly enhanced the level of green technology innovation in pilot cities. This core conclusion remains highly robust even after undergoing a series of stringent robustness checks, including placebo tests, PSM-DID, and replacing multi-dimensional dependent variables.

Synergistic dual-track driving mechanism: The policy does not rely on a single dimension of stimulus but cleverly constructs a dual-track synergistic transmission network. The policy not only leverages the “government sci-tech support” effect to alleviate financial constraints by increasing local government sci-tech expenditure, but also stimulates the “government environmental regulation” effect to increase compliance costs by strengthening local environmental regulation intensity.

Prominent long-cycle cumulative characteristics: The policy’s incentive for green technology innovation is not a short-term “pulse” stimulus but exhibits a long-cycle cumulative effect that gradually strengthens over time. Meanwhile, the cross-validation of dynamic trajectories via deep learning indicates that, restricted by strict linear assumptions, traditional multi-period DID models may somewhat underestimate the non-linear green innovation dividends stimulated by the policy over the long cycle.

Significant asymmetric driving effects: The policy’s incentive effect exhibits distinct heterogeneity. The marginal incentive effect is substantially greater in small-to-medium cities compared to large cities, acting as “snow in summer” for resource-constrained regions. Additionally, the policy generates a stronger “innovation compensation effect” in cities with a lower degree of industrial upgrading, forced by stricter environmental assessments.

9.2. Policy Implications

Based on the above findings, this paper proposes the following policy implications:

Deepen the “green” development orientation of innovative cities: Aimed at eastern regions and others where innovation resources are highly agglomerated, future construction and assessment of national innovative cities should further clarify and deepen their “green” orientation. The quality and quantity of green technology innovation should be established as key indicators for evaluating policy effectiveness, ensuring that macro-level policy incentives can be genuinely translated into a “win-win” situation that balances economic growth and ecological protection.

Comprehensively implement the collaborative governance system of “carrots + sticks”: When local governments execute innovation incentive policies, they must abandon single-dimensional management thinking. While providing ample financial subsidies and sci-tech expenditure support (“carrots”) to scientific research entities to offset the costs of positive innovation externalities, they must simultaneously implement and enforce high standards of environmental protection assessments and regulations (“sticks”). Based on our marginal effect analysis, while financial subsidies are foundational, the higher marginal return of environmental constraints must be fully leveraged. Local governments should continuouslfy calibrate the optimal matching ratio between the two, leaning slightly toward stringent environmental forcing while ensuring adequate financial compensation. Only through this organic integration and synergistic governance of “government sci-tech support” and “government environmental regulation” can enterprises’ path dependency be completely broken, reshaping the green and sustainable development momentum of the urban economy from its underlying logic.

Implement differentiated and targeted innovation strategies: Given the heterogeneous policy effects, policymakers should avoid a ‘one-size-fits-all’ approach. Greater policy inclination, resource allocation, and targeted financial support should be directed toward small-to-medium cities and industry-heavy cities to maximize the marginal elasticity of green innovation growth and effectively bridge their resource gaps.

9.3. Limitations and Future Research

Although this paper offers marginal contributions in evaluation methodologies and the exploration of dual mechanisms, several limitations remain to be expanded upon in the future. First, the empirical sample of this study strictly focuses on 100 prefecture-level and above cities in eastern China. Considering the heterogeneity in economic foundations and innovation environments in central and western regions, the generalizability of the conclusions nationwide requires further validation by expanding the research scope in the future. Second, regarding the measurement of green technology innovation, this paper primarily relies on various green patent data. While patent data effectively captures the output of R&D, it inherently struggles to fully evaluate the landing and transformation effects of these innovations. In other words, patent counts do not directly reflect the practical application of green technologies, their commercialization, or their ultimate actual ecological benefits (such as reductions in carbon emission intensity or energy consumption). Future research must break through the limitation of measuring innovation solely by patent data by incorporating micro-level green product sales or macro-level green total factor productivity to explore the policy’s promoting role in the practical transformation of green technologies. Finally, although this paper effectively overcame model overfitting through techniques like single-layer network simplification and L2 regularization constraints, the application of deep learning (LSTM and GRU) on macro short-panel data with limited time spans still faces inherent constraints. If future studies could introduce higher-dimensional, high-frequency micro-enterprise-level panel data, it is expected to further unleash the immense potential of the non-linear counterfactual prediction framework.