New energy demonstration urban construction enhances ecological environmental resilience: evidence from China Authors: Lei Li, Xinshuai Xue, Zengkun Yang, Jiamin Liu, Zedong Yang, Yaqian Ruan Published: 2026 ABSTRACT Renewable energy is a cornerstone for addressing ecological degradation, resource depletion and climate risk. Can the new energy demonstration urban construction (NEDC) as an important policy tool of renewable energy development improve ecological environmental resilience (ENR)? Based on a panel data of 283 prefecture level and above cities from 2011 to 2022, the dual machine learning model is constructed to explore the impact of NEDC on ENR. The study finds that the NEDC is an important policy driver for enhancing ENR. Industrial structure upgrading, green technology innovation and energy consumption are the crucial influence pathways. There is significant heterogeneity of geographic location and resource abundance in the impact of NEDC on ENR. The findings of the study provide policy references for relevant departments to continuously optimize and promote the NEDC. Through the rising of ENR, the beautiful modernization construction of harmonious coexistence between human and nature is being comprehensively promoted. Introduction Human society enters a new phase of intense urbanization and industrialization at the historical juncture of the twenty-first century. The vulnerability and uncertainty of eco-environmental systems are becoming increasingly evident. Global climate change, biodiversity degradation and frequent environmental risks are also intertwined. This not only continues to erode the carrying capacity of natural resources, but also profoundly reshapes the intrinsic relationship among national security, economic stability and social well-being1–3. The global eco-environmental system is gradually slipping into a state of “tipping point”. The ENR has become a key indicator of eco-environmental security and sustainable development. There has been continuous attention in the fields of ecology, environmental economics and policy sciences in recent years4–6. Especially in the context of systemic global environmental risks, ENR is not only the core of maintaining ecosystem homeostasis, but also an important fulcrum of the national environmental governance systems and capacities7. In the face of uncertainty and the overlapping impact of multiple factors, a single path of pollution control or resource conservation is no longer sustainable. A new paradigm of dynamic regulation, systemic response, and synergistic ecosystem governance is constructed8,9. As the world’s largest developing country, China faces more complex ecological pressures and governance challenges. On the one hand, rapid industrialization and urban expansion have damaged the resource environment, and the vulnerability of ecosystems has remarkably risen10,11. On the other hand, it is important to realize the synergy between green transformation and eco-environmental security in high-quality economic development12,13. Under this background, a series of policies, such as the National Climate Change Adaptation Strategy 2035 and the 14th Five-Year Plan for Ecological Protection and Supervision, have given clear instructions. The regulatory buffering capacity of eco-environmental systems is conducive to improve the comprehensive stress-bearing level to cope with sudden environmental events and long-term change trends. However, how to substantively enhance the ENR through quantifiable and traceable policy instruments. This is still a core proposition that needs to be broken through in theoretical research and policy practice. The new energy demonstration urban construction (NEDC), an important experimental policy to promote green development in China, provides a new opportunity for enhancing ENR. The policy aims to systematically develop clean energy, optimize energy infrastructure, promote green industry clustering and eco-institutional innovation. The structure of the energy system is reshaped, affecting the eco-environmental situation. Unlike traditional environmental policies that focus on end-of-pipe management, the NEDC promotes decarbonization and greening of the energy system at the source. This not only relieves pollution loads and eco-environmental pressures, but also improves the governance capacity of local governments. Theoretically, ENR can be enhanced, if the NEDC strengthens green factor allocation and environmental risk response capacity by reshaping the city’s energy structure. However, existing studies have mostly focused on the impact of NEDC on carbon emissions, green economy output or technological innovation14–17. There is no literature that systematically explores the impact of NEDC on ENR. Are the mechanisms of action and causal effects valid? There is a lack of rigorous theoretical and empirical evidence. Therefore, this study discusses the intrinsic causality of NEDC and ENR. The conclusions can provide realistic and guiding policy recommendations for the future green transformation and ecological civilization construction in China and other relevant countries. The marginal contributions of this paper primarily include the following aspects: (1) The positive impacts of optimizing the energy mix and investing in renewable energy on environmental sustainability18, reducing the ecological footprint19, and lowering carbon emissions have been extensively studied20. As a key regulator of the market economy, the government has promoted the development of the new energy industry. Environmental quality has also improved. However, the relationship between NEDC and ENR has not been explored. Based on a dual machine learning model, this study examines the impact of NEDC on ENR. The findings clarify the actual impact of the NEDC on ENR21, thereby enriching the empirical evidence for new energy policies in the field of comprehensive urban ecological management. (2) This study empirically identifies the intrinsic transmission mechanism through which the NEDC influences ENR, based on three practical transmission channels: industrial structural upgrading, green technological innovation, and optimization of the energy consumption structure. The findings provide valuable insights for relevant government departments to optimize energy system mechanisms and help enterprises improve the efficiency of resource allocation. Synergy among ecological, environmental, and socio-economic systems has also been enhanced. (3) The heterogeneity of the impact of NEDC on ENR is examined in terms of geographic location, city tier, and resource conditions. This provides theoretical support for governments and enterprises to refine energy industry policies tailored to local conditions, thereby enhancing ENR more efficiently. Furthermore, new energy sources improve ecological and environmental quality by reducing pollutants and greenhouse gases. The study’s conclusions expand the research boundaries in the field of energy economics22. Literature review In the face of the double challenges of global climate change and environmental degradation, how to enhance the ENR has become an important issue in realizing sustainable development strategies. Under the backdrop of the evolving concept of global eco-environmental governance, ENR has gradually become a criterion for measuring each subject’s ability to cope with external shocks and adapt to green transformation23–25 It incorporates multidimensional connotations of eco-environmental system services, environmental carrying capacity and dynamic regulatory mechanisms26,27. Scholars explore the resilient performance of eco-environmental systems under the natural disasters, industry shocks, and policy interventions28–30. However, the quantitative identification and enhancement path of ENR based on the NEDC perspective is still in the stage of theoretical exploration and empirical validation31. New energy technologies provide new policy grips and institutional tools for eco-environmental governance32. NEDC is an important institutional arrangement for the State to promote green and low-carbon development. Through the renewable energy projects, such as wind, solar and biomass, the traditional high-carbon energy system is replaced by green energy, realizing a synergistic improvement of the energy structure transformation and the ecological environment quality. The impact of NEDC on carbon intensity, energy efficiency and industrial structure upgrading has been widely explored. It has a significant role in reducing energy consumption, promoting technological innovation and improving air quality33–35. However, whether the NEDC can enhance the stability of eco-environmental systems. Relevant research is still relatively scarce. On the one hand, most of the existing studies focus on the reduction effect of static environmental emission of NEDC, ignoring the dynamic attribute dimension of resilience36,37. ENR requires not only a short-term decline of pollution emissions, but more systematically emphasizes continuous stability under perturbations such as extreme climate, resource constraints, and population density growth38–40. The absence of this perspective leads to a one-sided assessment of policy effects, making it difficult to fully reflect the deep impacts of NEDC on ecosystem structure and function. On the other hand, some studies discuss the effect of NEDC on pollution emissions, traditional difference-in-difference models are mostly used in the studies41. Non-parametric causal inference methods represented by machine learning have gradually emerged in economic research, providing technical support for complex policy evaluation. Dual machine learning has significant advantages in overcoming high-dimensional covariate interference, mitigating model-setting bias42–44. Especially, it is applied to assess the causal relationship between non-randomized policy pilots and multiple environmental responses. Studies have been examined around the topics of green finance, carbon markets and pollution management by using machine learning methods45–47. No scholars have used the dual machine learning method to assess the impact of NEDC on ENR. In addition, the measurement of ENR constitutes a key bottleneck in current research. Some scholars have tried to construct a resilience indicator system from the dimensions of ecosystem service redundancy, pollution recovery cycle, and land cover change48–50. How to construct systematic and comparable indicators of ENR from the multi-source integration of state resilience, pressure resilience and response resilience, remaining the groundwork for realizing relevant policy identification. In summary, it can be seen that the potential enhancement impact of NEDC on ENR urgently should be systematically tested through theory and empirical evidence. Meanwhile, the new causal inference methods, such as dual machine learning, provides theoretical and technical paths to solve the traditional recognition challenges. Therefore, the dual machine learning model is used to discuss the impact of NEDC on ENR. This fills in the intrinsic causal relationship between NEDC and ENR at the empirical level, and provides theoretical logic to support the understanding of the complex relationship among energy transition, policy interventions, and ecological resilience. The research boundaries of eco-environmental policy assessment are further expanded. The Chinese experience of the global green transition and eco-environmental governance in terms of relevance and policy implications is also provided. Theoretical mechanisms and research hypotheses Direct effect of NEDC on ENR The enhancement effect of NEDC on ENR is rooted in the structural change of resource allocation and functional reconfiguration of the environmental system51. In theory, ENR emphasizes the ability to maintain its core functions and structural stability under the external risk shocks, which is reflected in the eco-environmental resistance to perturbation, adaptability and resilience. The essence is the capacity of eco-environmental systems to address dynamic uncertainty. NEDC are policy experiments to promote the clean energy transition. Its institutional design is deeply embedded in the urban energy consumption structure and eco-environmental governance system, enhancing ENR through the following effects. Firstly, under the factor substitution mechanism, NEDC achieves effective substitution of traditional high-carbon energy sources through large-scale deployment of renewable energy such as wind, photovoltaic and biomass52. This reduces the pollution emissions in terms of air, water and soil and alleviates the sensitivity of eco-environmental systems to pollution shocks at source. The stability and marginal adaptive capacity of eco-environmental systems are enhanced53. Secondly, under the eco-environmental pressure mitigation mechanism, polit cities are often accompanied by a double constraint regime of energy intensity and emission intensity. This guides local governments to optimize the resource investment of eco-environmental management, forming a closed loop of governance that integrates pollution control, ecological restoration and risk early warning. The disturbance and restoration cycle of the eco-environmental system is effectively shortened. The self-organization and rapid response capacity of the eco-environmental system is improved54. Furthermore, at the level of institutional coordination mechanisms, the NEDC strengthens the forward-looking preventive capacity of local governments with regard to unexpected environmental risks through institutional tools such as policy support, financial support and performance appraisal. The eco-environmental governance is transforming from reactive response to proactive governance, enhancing the early warning capacity and inter-temporal resilience of the eco-environmental system55. Finally, under the incentive mechanism of eco-environmental behavior, the NEDC promotes the participation of enterprises, residents and other multiple actors in eco-environmental protection and green production. The micro subjects form the behavioral inertia of flexible regulation, positive feedback and synergistic co-management in the way of resource use, emission behavior and eco-environmental restoration. The stability and synergy of the overall eco-environmental system is enhanced56. Thus, the research hypothesis H1 is formulated: H1 NEDC can enhance ENR. Indirect effect of NEDC on ENR Mechanisms of the industrial structure upgrading The NEDC effectively promotes the optimize and upgrading of the regional industrial structure by reshaping the resource allocation pattern and institutional guidance system. The structural stability and functional resilience of eco-environmental systems can be thus enhanced. This constitutes an important pathway for enhancing ENR. According to the theory of structural change, the industrial structure upgrading is not only manifested in the substitution of the proportion of the secondary industry to the tertiary industry, but also in the process of the gradual substitution of high-pollution and high-emission industries by low-pollution and low-carbon emission industries57. The NEDC guides the agglomeration of key factors, such as capital, land and labor, towards environmentally friendly industries through clear green-oriented policies and institutional arrangements. The living space of high-energy-consuming and high-emission industries would be effectively compressed, thereby reducing the systemic pressure on the ecological environment. The continuous evolution of the internal structure of the industry is accompanied by a significant decline in energy consumption and pollutant emissions per unit of output. The load on the eco-environmental system is alleviated and the resilience and regulation of this system is subsequently increased58. In addition, the industrial structure upgrading further changes the spatial distribution of regional economic activities and the use of eco-environmental resources59. Green services and circular economy relative industries are usually characterized by intensive land use and low disturbance. Such industrial expansion not only avoids over-occupation of eco-environmental core areas, but also prompts local governments to optimize the eco-environmental spatial layout. This promotes the protection of ecologically sensitive areas and the function restorations, thereby enhancing the spatial connectivity and structural integrity of eco-environmental systems. Meanwhile, under the double role of institutional pressure and environmental regulation of the NEDC, companies face higher environmental compliance standards. The original high-polluting enterprises gradually withdraw or shift to the green sector, and emerging green enterprises become the mainstay of the regional economy. Business behavior is generally characterized by resource conservation, pollution control and eco-environmental synergy60. The change of producer behavior not only reduces the total load of the eco-environmental system, but also demonstrates the system coordination in terms of resource use patterns and emission pathways. The ability of eco-environmental systems to self-regulate and recover is thus significantly improved61. Thus, the research hypothesis H2 is formulated: H2 NEDC can enhance ENR by promoting industrial structure upgrading. Mechanisms of the green technology innovation The NEDC builds an important path to enhance ENR by prompting green technology innovation. According to environmental economics and the theory of technological change, green technological innovation is seen as a key endogenous variable for environmental quality improvement. It has a direct contribution to the enhancement of eco-environmental system resistance, resilience and adaptive capacity of eco-environmental systems62. The NEDC is institutional experiments. By constructing a high-intensity incentive environment, improving the green innovation ecosystem, and optimizing the technology diffusion path, the generation and application of green technologies are provided with superior soil63. On the one hand, the NEDC significantly lowers the threshold and trial-and-error cost of green technology development. Through financial subsidies, tax breaks, green credits and other means, the benefits expected from green technological innovation by enterprises and research institutions are significantly enhanced. The dilemma of “high cost, long cycle and uncertainty” in traditional green technology innovation is effectively overcome. On the other hand, NEDC focus on key technological areas such as green energy, efficient energy storage and smart grids. Cross-industry and cross-sectoral collaborative innovation mechanisms are created to strengthen the transformation efficiency and diffusion effect of innovation results. Green technologies are more widely embedded in urban energy systems, industrial systems and governance systems. Green technology innovation effectively enhances the precision and systematicity of pollution management through process substitution, equipment renewal and workflow optimization. Resource-environmental pressures are reduced, enhancing ecosystem self-repair in external risk shocks64. The functional resilience and structural stability of eco-environmental systems are improved. In addition, green technological innovation promotes green transformation of the industrial structure. Environmental governance is transformed from end-of-pipe to whole-process control and evolved from static compliance to dynamic optimization, achieving a fundamental reconstruction of the governance paradigm. Furthermore, the knowledge accumulation and spillover effects from the continuous iterative upgrading of green technologies are increased65. This not only enhances eco-environmental governance capacity, but also leads to an overall leap of green technology in non-demonstration regions through technology diffusion and institutional imitation. ENR is empowered. Thus, the research hypothesis H3 is formulated: H3 NEDC can improve ENR by prompting green technology innovation. Mechanisms of the energy consumption The NEDC is an important transmission mechanism for improving ENR by effectively reducing fossil energy consumption. According to resource and environmental economics and sustainable development theory, energy consumption is not only related to the boundary of resource carrying capacity, but also directly affects the eco-environmental load intensity and recovery rhythm66. Under the traditional path of industrialization, high-intensity energy consumption often releases large amounts of carbon emissions and pollutants. This weakens the buffering capacity and stability of eco-environmental systems67. The NEDC significantly reduces the energy input intensity per unit of economic output by reshaping the energy structure and improving end-use energy efficiency. Physical disturbances and pollution coercion of the eco-environmental system are weakened at the source68, realizing a fundamental shift from “emission control” to “load reduction”. On the one hand, NEDC promotes the substitution of traditional fossil energy sources by clean energy sources, reducing the carbon intensity and total pollutant emissions. On the other hand, energy use in buildings, transportation, manufacturing and other areas is systematically optimized, based on the development of green infrastructure and the promotion of energy-efficient technologies. More critically, the continuous decline of energy consumption creates more time and space for adaption of eco-environmental systems. This enhances the resilient responsiveness of eco-environmental systems to sudden risk shocks26. As energy consumption moves away from the traditional crude growth path, eco-environmental systems are able to repair and rebuild their structure and function with low intervention. The system’s ability to maintain steady state and to recover quickly from nonlinear shocks is strengthened. In addition, the suppression mechanism of energy consumption triggers the transformation of production and lifestyle. This promotes low-carbon behaviors as a mainstream choice further consolidates ENR at the societal level. Thus, the research hypothesis H4 is formulated: H4 NEDC can improve ENR by reducing energy consumption. Figure 1 is a diagram of the theoretical mechanisms of NEDC affecting ENR. Fig. 1Theoretical mechanisms of NEDC impacting ENR. Notes Fig. 1 is drawn by the research team. Theoretical mechanisms of NEDC impacting ENR. Notes Fig. 1 is drawn by the research team. Research design Model construction Based on the theoretical mechanism analysis, the impact of NEDC on ENR is empirically discussed. Existing studies that assess policy effects often employ methods such as difference-in-differences (DID) and synthetic controls. While these methods are well-established, they encounter difficulties when numerous control variables are required to satisfy the conditional unconfoundedness assumption. Traditional regression approaches with high-dimensional controls are prone to the curse of dimensionality and model misspecification, which may result in biased estimation results. To address these econometric problems arising from high-dimensional data, this paper adopts the double machine learning (DML) framework proposed by Chernozhukov et al.69. DML is highly applicable to this study for two main reasons. First, ENR is a comprehensive indicator affected by a wide range of socioeconomic factors. DML can automatically screen prespecified high-dimensional control variables and identify core influencing factors. This approach mitigates the curse of dimensionality and alleviates omitted variable bias caused by subjective selection of limited key control variables. Second, DML is robust to potential nonlinear relationships among variables. Combined with machine learning algorithms and cross-fitting strategy, the DML method can effectively cope with complex data structures and reduce regularization bias, and thus ensures reliable statistical inference of treatment effect estimates. The method has been widely used in studies of policy assessment43,44,46,47,70,71. Referring to Ling et al.43, Zhang et al.47, and Hu and Xu44, the following DML models are constructed:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$EN{R_{{\mathrm{it}}}}={\alpha _0}NED{C_{it}}+g\left( {{X_{it}}} \right)+{V_{it}}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E\left( {{V_{{\mathrm{it}}}}|NED{C_{{\mathrm{it}}}},{X_{{\mathrm{it}}}}} \right)=0$$\end{document} In Eqs. (1) and (2), i denotes individual city. t denotes year. ENR denotes ecological environmental resilience. NEDC denotes the dummy variable of new energy demonstration urban construction. α0 is the disposition coefficient corresponding to the policy. X represents a set of control variables and their various functional transformation forms. g(·) is the machine learning model. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{g}}\mathrm{(}{\mathrm{X}}_{\mathrm{it}}\mathrm{)}$$\end{document} is estimated by using the machine learning algorithm. Vit represents the error term with a conditional mean of zero. Performing a regression on Eqs. (1) and (2) yields the following estimate of NEDC:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat {\alpha }_0}={\left( {\frac{1}{{\mathrm{n}}}\sum\limits_{{i \in I,t \in T}} {NEDC_{{it}}^{2}} } \right)^{ - 1}}\frac{1}{{\mathrm{n}}}\sum\limits_{{i \in I,t \in T}} {NED{C_{it}}} \left( {EN{R_{{\mathrm{it}}}} - \widehat {g}\left( {{X_{it}}} \right)} \right)$$\end{document} Here, n represents the sample size. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{ \alpha }}_{\mathrm{0}}$$\end{document} does not converge to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{ \alpha }_{\mathrm{0}}$$\end{document}, this results in estimation bias. Further investigation of the bias is required, followed by corrections. The procedure is as follows:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \sqrt {\mathrm{n}} \left( {{{\widehat {\alpha }}_0} - {\alpha _0}} \right) & ={\left( {\frac{1}{{\mathrm{n}}}\sum\limits_{{i \in I,t \in T}} {NEDC_{{it}}^{2}} } \right)^{ - 1}}\frac{1}{{\sqrt {\mathrm{n}} }}\sum\limits_{{i \in I,t \in T}} {NED{C_{it}}} {V_{it}} \\ & \quad +{\left( {\frac{1}{{\mathrm{n}}}\sum\limits_{{i \in I,t \in T}} {NEDC_{{it}}^{2}} } \right)^{ - 1}}\frac{1}{{\sqrt {\mathrm{n}} }}\sum\limits_{{i \in I,t \in T}} {NED{C_{it}}} \left[ {g\left( {{X_{it}}} \right) - \widehat {g}\left( {{X_{it}}} \right)} \right] \\ \end{aligned}$$\end{document} Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{a = }}\left( {\frac{{\mathrm{1}}}{{\mathrm{n}}}\sum\limits_{{{\mathrm{i}} \in {\mathrm{I,t}} \in {\mathrm{T}}}} {{\mathrm{NEDC}}_{{{\mathrm{it}}}}^{{\mathrm{2}}} } } \right)^{{{\text{ - 1}}}} \frac{{\mathrm{1}}}{{\sqrt {\mathrm{n}} }}\sum\limits_{{{\mathrm{i}} \in {\mathrm{I,t}} \in {\mathrm{T}}}} {{\mathrm{NEDC}}_{{{\mathrm{it}}}} {\mathrm{V}}_{{{\mathrm{it}}}} }$$\end{document}, characterized by a normal distribution with mean zero, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{b = }}\left( {\frac{{\mathrm{1}}}{{\mathrm{n}}}\sum\nolimits_{{{\mathrm{i}} \in {\mathrm{I,t}} \in {\mathrm{T}}}} {{\mathrm{NEDC}}_{{{\mathrm{it}}}}^{{\mathrm{2}}} } } \right)^{{ - 1}} \frac{{\mathrm{1}}}{{\sqrt {\mathrm{n}} }}\sum\nolimits_{{{\mathrm{i}} \in {\mathrm{I,t}} \in {\mathrm{T}}}} {{\mathrm{NEDC}}_{{{\mathrm{it}}}} \left[ {{\mathrm{g}}\left( {{\mathrm{X}}_{{{\mathrm{it}}}} } \right) - \widehat{{\mathrm{g}}}\left( {{\mathrm{X}}_{{{\mathrm{it}}}} } \right)} \right]}$$\end{document}. DML uses machine learning algorithms and regularization methods to estimate the functional form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{g}}\left({\mathrm{X}}_{\mathrm{it}}\right)$$\end{document}, but the introduction of regularization terms introduces estimation bias. This results in a slower convergence rate from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{g}}\left({\mathrm{X}}_{\mathrm{it}}\right)$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{g}\left({\mathrm{X}}_{\mathrm{it}}\right)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{{ - \phi s}}$$\end{document}>\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{n}}^{\mathrm{-1/2}}$$\end{document}. As n approaches infinity, b also approaches infinity, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{ \alpha }}_{\mathrm{0}}$$\end{document} does not converge to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{ \alpha }_{\mathrm{0}}$$\end{document}. Therefore, to improve the accuracy and robustness of the estimates, the orthogonal method is used to correct this bias, and the K-fold cross-validation method is employed to train the prediction function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{g}}\left({\mathrm{X}}_{\mathrm{i}\mathrm{t}}\right).$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$NED{C_{{\mathrm{it}}}}=m\left( {{X_{it}}} \right)+{U_{it}}$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E\pi \left( {{U_{{\mathrm{it}}}}|{X_{it}}} \right)=0$$\end{document} Performing a regression on Eqs. (5) and (6) yields an estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{\mathrm{U}}}_{\mathrm{it}}\mathrm{=}{\mathrm{N}\mathrm{E}\mathrm{D}\mathrm{C}}_{\mathrm{it}}\mathrm{-m}\left({\mathrm{X}}_{\mathrm{it}}\right)$$\end{document} for the residual term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{U}}_{\mathrm{it}}$$\end{document}. Treating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{U}}_{\mathrm{it}}$$\end{document} as an instrumental variable for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{N}\mathrm{E}\mathrm{D}\mathrm{C}}_{\mathrm{it}}$$\end{document}, we can perform a regression to obtain an unbiased estimate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\alpha}}_{\mathrm{0}}$$\end{document}:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat {\alpha }_0}={\left( {\frac{1}{{\mathrm{n}}}\sum\limits_{{i \in I,t \in T}} {NEDC_{{it}}^{2}} } \right)^{ - 1}}\frac{1}{{\sqrt {\mathrm{n}} }}\sum\limits_{{i \in I,t \in T}} {{{\widehat {U}}_{it}}} \left( {EN{R_{{\mathrm{it}}}} - \widehat {g}\left( {{X_{{\mathrm{it}}}}} \right)} \right)$$\end{document} Further examination of estimation errors:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{gathered} \sqrt {\mathrm{n}} \left( {{{\widehat {\alpha }}_0} - {\alpha _0}} \right)=\left[ {E{{\left( {U_{{{\mathrm{it}}}}^{2}} \right)}^{ - 1}}} \right]\frac{1}{{\sqrt n }}\sum\limits_{{i \in I,t \in T}} {{U_{it}}{V_{it}}} + \\ {\left[ {E\left( {U_{{{\mathrm{it}}}}^{2}} \right)} \right]^{ - 1}}\frac{1}{{\sqrt n }}\sum\limits_{{i \in I,t \in T}} {\left[ {m\left( {{X_{it}}} \right) - \widehat {m}\left( {{X_{it}}} \right)} \right]} \left( {g\left( {{X_{{\mathrm{it}}}}} \right) - \widehat {g}\left( {{X_{{\mathrm{it}}}}} \right)} \right) \\ \end{gathered}$$\end{document} In particular, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {{\mathrm{E}}\left( {{\mathrm{U}}_{{{\mathrm{it}}}}^{{\mathrm{2}}} } \right)} \right]^{{ - 1}} \frac{{\mathrm{1}}}{{\sqrt {\mathrm{n}} }}\sum\nolimits_{{{\mathrm{i}} \in {\mathrm{I,t}} \in {\mathrm{T}}}} {{\mathrm{U}}_{{{\mathrm{it}}}} {\mathrm{V}}_{{{\mathrm{it}}}} }$$\end{document} follows a normal distribution with a mean of zero. After applying machine learning twice, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {{\mathrm{E}}\left( {{\mathrm{U}}_{{{\mathrm{it}}}}^{{\mathrm{2}}} } \right)} \right]^{{ - 1}} \frac{{\mathrm{1}}}{{\sqrt {\mathrm{n}} }}\sum _{{{\mathrm{i}} \in {\mathrm{I,t}} \in {\mathrm{T}}}} \left[ {{\mathrm{m}}\left( {{\mathrm{X}}_{{{\mathrm{it}}}} } \right) - {\hat{\mathrm{m}}}\left( {{\mathrm{X}}_{{{\mathrm{it}}}} } \right)} \right]\left[ {{\mathrm{g}}\left( {{\mathrm{X}}_{{{\mathrm{it}}}} } \right){\hat{ - \mathrm{g}}}\left( {{\mathrm{X}}_{{{\mathrm{it}}}} } \right)} \right]$$\end{document} converges at a rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{n}}^{-({{\upphi\:}}_{\mathrm{g}}+{{\upphi\:}}_{\mathrm{m}})}$$\end{document}. Compared to Eq. (4), the convergence rate toward zero in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sqrt{\mathrm{n}}\left({\widehat{{\upalpha\:}}}_{0}-{{\upalpha\:}}_{0}\right)$$\end{document} direction is faster; by using this to calculate the unbiased estimator, we obtain a more accurate estimate. In addition, to account for potential unobservable factors that may simultaneously affect NEDC pilot selection and ENR development, including local willingness for green transition, long-term ecological governance performance, expected regional energy demand growth, central policy priority arrangement and local officials’ governance characteristics, this study adopts city fixed effects and year fixed effects together with ten control variables screened by DML to effectively isolate and control various confounding factors. City fixed effects capture time-invariant urban heterogeneity and absorb non-time-varying unobservable characteristics such as geographical endowments, long-term governance concepts, regional governance culture and stable policy inclinations of local officials. Year fixed effects identify nationwide common macro shocks and eliminate systematic temporal disturbances including adjustments in central policy support, shifts in national energy strategies and macroeconomic cycle fluctuations. As for time-varying latent unobservable factors that cannot be fully eliminated by two-way fixed effects, the ten control variables act as valid proxies. Government intervention and comprehensive pollution governance capacity directly reflect local green development willingness and actual environmental governance outcomes; R&D investment intensity, tertiary industry share and financial development level well reflect regional energy structure upgrading potential and future energy demand trends. Population agglomeration, logarithm of per capita GDP, urbanization rate, internet infrastructure and the number of universities further depict urban economic foundation, policy implementation conditions as well as scientific, educational and digital development environment, which indirectly reflect local officials’ governance preferences and development orientations. In summary, under the empirical framework of double machine learning, the design combining two-way fixed effects and multi-dimensional control variables can fully satisfy the conditional unconfoundedness assumption, greatly alleviate endogeneity problems arising from omitted variable bias, and guarantee the scientificity, rigor and reliability of policy effect identification results. Furthermore, regarding machine learning algorithms, the gradient boosting algorithm was selected for result estimation. In subsequent robustness tests, we re-estimated the model using four alternative machine learning algorithms: LassoCV (Lasso regression with cross-validation), ElasticNetCV (Elastic Net with cross-validation), and RidgeCV (Ridge regression with cross-validation). Set K-fold to 5, with a sample split ratio of 1:4. Variable description Explained variable: ecological environmental resilience (ENR) Referring to Li et al.23, Xie et al.72, Lee et al.73, this study constructs an ecological environmental resilience index from the three dimensions of state resilience, pressure resilience, and response resilience. The construction of the Ecological and Environmental Resilience Index (ENR) follows the classic “Pressure-State-Response” (PSR) theoretical framework. Incorporating ecological state indicators, pollution emission pressure indicators, and governance response indicators into a single comprehensive system provides a solid theoretical foundation in terms of conceptual logic. From the perspective of the core essence of ecological resilience, it is fundamentally a dynamic process through which ecosystems maintain stability, achieve recovery, and enhance their adaptive capacity in the face of external disturbances. State-related indicators characterize the baseline endowment and fundamental steady-state conditions of an ecosystem, serving as the material foundation for resilience; pressure-related indicators quantify the intensity of disturbances caused by human activities to the ecosystem, representing the core challenges that resilience must address; response-related indicators characterize the system’s proactive adaptive capacity to mitigate pressures and restore balance through governance and regulation, serving as the key safeguard for resilience. These three categories correspond respectively to the complete closed loop of ecological resilience including shock resistance, disturbance tolerance, and recovery and adaptation, and together constitute a systematic characterization of the multidimensional attributes of ecological resilience. Therefore, the integration of these indicators to construct a comprehensive index provides a complete coverage and concrete expression of the theoretical essence of ecological resilience. For state resilience, scholars consider factors such as water resources, green coverage, and built-up area. In particular, when it comes to state resilience, most existing scholars focus on water resources, green coverage, and the area of built-up zones. These factors are used to provide a visual representation of the external state of the urban ecosystem. However, ecological resilience is not merely an external characteristic. Reciprocal cycling processes among systems such as the atmosphere, surface and subsurface together form complex ecological adaptive systems74,75. The core of this should be the ecosystem’s ability to regulate itself. Therefore, building on existing research, factors that support ecosystem stability—such as forested areas, grasslands, shrublands, and wetlands—have been incorporated into the state-resilience indicator system. For pressure resilience, the existing literature focuses on wastewater, exhaust gas, and carbon emission. These traditional pollutants place additional strain on the urban ecological environment. Large external risk shocks caused by warming temperatures and frequent climate extremes are an important component of pressure resilience The two greenhouse gases, CH4 and N2O, directly exacerbate external pressures on the urban ecological environment. Accordingly, these two greenhouse gas factors (CH4 and N2O) are added. For response resilience, factors such as wastewater, exhaust and non-hazardous waste treatment rates are commonly used for characterization. These factors reflect the capacity for end-of-pipe treatment of environmental issues. However, focusing solely on end-stage governance while neglecting front-end oversight fails to fully reflect the responsiveness of the urban ecosystem. Energy-saving and environmental protection expenditures and green innovation can enhance factor utilization and productivity, reducing pollution emissions at the source15. Therefore, the two factors are included in the indicator system of response resilience. The indicator system of ecological environmental resilience (ENR) is presented in Table 1. The entropy weight method is used to calculate the composite value of the ENR, referring to Lee et al.73. The specific calculations are as follows: ①Standardization of sub-indicators: Convert the third-level indicators to dimensionless form.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R_{{\mathrm{ij}}}}=\left\{ \begin{gathered} \frac{{{X_{{\mathrm{ij}}}} - \hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)-}}\hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}~~~~{X_{{\mathrm{ij}}}}{\text{is a positive indicator}} \hfill \\ \frac{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}} - {X_{{\mathrm{ij}}}}}}{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)-}}\hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}~~~~{X_{{\mathrm{ij}}}}{\text{is a negative indicator}} \hfill \\ \end{gathered} \right.$$\end{document} ①Standardization of sub-indicators: Convert the third-level indicators to dimensionless form.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R_{{\mathrm{ij}}}}=\left\{ \begin{gathered} \frac{{{X_{{\mathrm{ij}}}} - \hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)-}}\hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}~~~~{X_{{\mathrm{ij}}}}{\text{is a positive indicator}} \hfill \\ \frac{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}} - {X_{{\mathrm{ij}}}}}}{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)-}}\hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}~~~~{X_{{\mathrm{ij}}}}{\text{is a negative indicator}} \hfill \\ \end{gathered} \right.$$\end{document} Standardization of sub-indicators: Convert the third-level indicators to dimensionless form.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R_{{\mathrm{ij}}}}=\left\{ \begin{gathered} \frac{{{X_{{\mathrm{ij}}}} - \hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)-}}\hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}~~~~{X_{{\mathrm{ij}}}}{\text{is a positive indicator}} \hfill \\ \frac{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}} - {X_{{\mathrm{ij}}}}}}{{\hbox{max} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)-}}\hbox{min} ({X_{\mathrm{i}}}_{{\mathrm{j}}}{\mathrm{)}}}}~~~~{X_{{\mathrm{ij}}}}{\text{is a negative indicator}} \hfill \\ \end{gathered} \right.$$\end{document} In Eq. (9), Xij is the original value. Rij is the normalized value. i represents the sample city. j represents the sample indicator. min (Xij) is the minimum value of the observation. max (Xij) is the maximum value of the observation. ②The information entropy Ej and weights Wj of Rij: Calculate information entropy and indicator weights using the entropy method.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{j} = - \ln \frac{1}{n}\sum\limits_{{i = 1}}^{j} {\left[ {\left( {R_{{{\mathrm{ij}}}} /\sum\limits_{{i = 1}}^{n} {R_{{{\mathrm{ij}}}} } } \right) \times \ln \left( {R_{{{\mathrm{ij}}}} /\sum\limits_{{i = 1}}^{n} {R_{{{\mathrm{ij}}}} } } \right)} \right]}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{{\mathrm{j}}} = \left( {1 - E_{{\mathrm{j}}} } \right)/\sum\limits_{{j = 1}}^{m} {\left( {1 - E_{{\mathrm{j}}} } \right)}$$\end{document} ②The information entropy Ej and weights Wj of Rij: Calculate information entropy and indicator weights using the entropy method.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{j} = - \ln \frac{1}{n}\sum\limits_{{i = 1}}^{j} {\left[ {\left( {R_{{{\mathrm{ij}}}} /\sum\limits_{{i = 1}}^{n} {R_{{{\mathrm{ij}}}} } } \right) \times \ln \left( {R_{{{\mathrm{ij}}}} /\sum\limits_{{i = 1}}^{n} {R_{{{\mathrm{ij}}}} } } \right)} \right]}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{{\mathrm{j}}} = \left( {1 - E_{{\mathrm{j}}} } \right)/\sum\limits_{{j = 1}}^{m} {\left( {1 - E_{{\mathrm{j}}} } \right)}$$\end{document} The information entropy Ej and weights Wj of Rij: Calculate information entropy and indicator weights using the entropy method.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{j} = - \ln \frac{1}{n}\sum\limits_{{i = 1}}^{j} {\left[ {\left( {R_{{{\mathrm{ij}}}} /\sum\limits_{{i = 1}}^{n} {R_{{{\mathrm{ij}}}} } } \right) \times \ln \left( {R_{{{\mathrm{ij}}}} /\sum\limits_{{i = 1}}^{n} {R_{{{\mathrm{ij}}}} } } \right)} \right]}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{{\mathrm{j}}} = \left( {1 - E_{{\mathrm{j}}} } \right)/\sum\limits_{{j = 1}}^{m} {\left( {1 - E_{{\mathrm{j}}} } \right)}$$\end{document} ③Ecological environmental resilience (ENR): The ENR index is obtained by multiplying the standardized indicators by the weights calculated using the entropy method and then summing the results.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ENR=\sum\limits_{{j=1}}^{m} {{w_j} \times {R_{ij}}}$$\end{document} ③Ecological environmental resilience (ENR): The ENR index is obtained by multiplying the standardized indicators by the weights calculated using the entropy method and then summing the results.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ENR=\sum\limits_{{j=1}}^{m} {{w_j} \times {R_{ij}}}$$\end{document} Ecological environmental resilience (ENR): The ENR index is obtained by multiplying the standardized indicators by the weights calculated using the entropy method and then summing the results.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ENR=\sum\limits_{{j=1}}^{m} {{w_j} \times {R_{ij}}}$$\end{document} Table 1Indicator system of ENR.General layerGoal layerIndicator layerUnitDirectionIndicator weighting(%)Ecological environmental resilience (ENR)State resiliencePer capita water resourcesCubic meters/person+0.18Greening coverage of built-up areas%+0.10Per capita area of parklandHectares/ten thousand people+4.82Per capita built-up areaSquare kilometers/ten thousand people+4.24Area of woodland coverKM2+3.81Grass-covered areaKM2+13.82Area of bush coverKM2+18.43Wetland coverageKM2+34.64Pressure resiliencePer capita industrial wastewater dischargeTons/person−0.00Per capita industrial soot emissionsTons/person−0.02Per capita SO2 emissionsTons/person−0.03Per capita CO2 emissionsTons/person−0.03Average PM2.5 concentrationug/m3−0.42CH4 emissionston−0.05N2O emissionston−0.09Response resilienceIndustrial sulfur dioxide removalton+0.08Industrial soot removalton+0.08Non-hazardous disposal rate of domestic waste%+0.13Comprehensive utilization rate of industrial solid waste%+0.06Centralized treatment rate of sewage treatment plants%+0.16Expenditures on energy conservation and environmental protectionten thousand yuan+6.01Green innovation capacityPieces+12.78 Table 1Indicator system of ENR.General layerGoal layerIndicator layerUnitDirectionIndicator weighting(%)Ecological environmental resilience (ENR)State resiliencePer capita water resourcesCubic meters/person+0.18Greening coverage of built-up areas%+0.10Per capita area of parklandHectares/ten thousand people+4.82Per capita built-up areaSquare kilometers/ten thousand people+4.24Area of woodland coverKM2+3.81Grass-covered areaKM2+13.82Area of bush coverKM2+18.43Wetland coverageKM2+34.64Pressure resiliencePer capita industrial wastewater dischargeTons/person−0.00Per capita industrial soot emissionsTons/person−0.02Per capita SO2 emissionsTons/person−0.03Per capita CO2 emissionsTons/person−0.03Average PM2.5 concentrationug/m3−0.42CH4 emissionston−0.05N2O emissionston−0.09Response resilienceIndustrial sulfur dioxide removalton+0.08Industrial soot removalton+0.08Non-hazardous disposal rate of domestic waste%+0.13Comprehensive utilization rate of industrial solid waste%+0.06Centralized treatment rate of sewage treatment plants%+0.16Expenditures on energy conservation and environmental protectionten thousand yuan+6.01Green innovation capacityPieces+12.78 Indicator system of ENR. Core explanatory variables: new energy demonstration urban construction (NEDC) NEDC is a dummy variable for new energy demonstration urban construction. The interaction term between the city dummy variable (Treated) and the time dummy variable (Time) is NEDC. Referring to Liu et al.14, the city dummy variable Treated is set to 1 if a city is located in a pilot area and 0 otherwise. The time dummy variable (Time) is assigned a value of 1 if the year is 2014 and later, otherwise it is assigned a value of 0. Treated×Time=NEDC. Control variables Referring to Liu et al.14, the interference of urban characteristics on the model estimation results is eliminated. A set of candidate control variables was constructed, comprising financial conditions; education expenditure; the logarithm of per capita GDP; the share of the secondary sector; the share of the tertiary sector; government intervention; the scale of local fiscal revenue; R&D intensity; foreign direct investment; urban employment levels; passenger transport infrastructure levels; internet infrastructure levels; comprehensive pollution control capacity; population agglomeration; economic agglomeration; the number of universities; and the urbanization rate. The final control variables are then selected using the PDS-Lasso method. The following 10 control variables were ultimately selected: financial conditions (FIC); population agglomeration (PAG); logarithm of per capita GDP (LPGDP); share of the tertiary sector (TES); government intervention (GVI); R&D intensity (RDI); level of internet infrastructure (INI); number of universities (NOU); comprehensive pollution control capacity (PCC); and urbanization rate (URR). FIC is measured by the loan-to-deposit ratio of financial institutions; PAG is calculated using the location entropy method, with the specific formula as follows; LPGDP is measured by the logarithm of GDP per capita; TES is measured by the share of tertiary sector value added in GDP; GVI is measured by the share of local government expenditure in regional gross domestic product; RDI is measured by the ratio of city-wide science and technology expenditure to regional gross domestic product; INI is measured by the ratio of Internet users to the registered population; NOU is measured by the number of higher education institutions in the city; PCC is measured by the average of the household waste harmless treatment rate and the urban sewage treatment rate;. URR is measured by the proportion of the urban permanent resident population to the total permanent resident population.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PA{G_{{\mathrm{it}}}}=\frac{{{{po{p_{it}}} \mathord{\left/ {\vphantom {{po{p_{it}}} {Are{a_{{\mathrm{it}}}}}}} \right. \kern-0pt} {Are{a_{{\mathrm{it}}}}}}}}{{{{\sum\nolimits_{{i=1}}^{n} {po{p_t}} } \mathord{\left/ {\vphantom {{\sum\nolimits_{{i=1}}^{n} {po{p_t}} } {\sum\nolimits_{{i=1}}^{n} {Are{a_t}} }}} \right. \kern-0pt} {\sum\nolimits_{{i=1}}^{n} {Are{a_t}} }}}}$$\end{document} Data sources and descriptive statistics of variables Based on data availability, panel data of 283 cities at prefecture level and above from 2011 to 2022 are selected for the empirical study. The study sample does not include data from Hong Kong, Macau, Taiwan and Tibet. Prefectural-level cities such as Sansha, Danzhou, Bijie, Tongren, Haidong, Turpan, and Hami, which had significant data gaps, were excluded. The data sources for constructing variables of ENR are as follows: The data source for the area covered by grassland, woodland, shrubland and wetland comes from “The 30 m annual land cover datasets and its dynamics in China from 1985 to 2023”76. We use the terra package in R to crop the downloaded raster data and calculate the areas of different land cover types for each city in each year. CO2, CH4, N2O, PM2.5 are from EDGAR and ACAG. Using the same method, the annual average values for CO2, CH4, N2O, and PM2.5 emissions for each city over the years were obtained. Other data is from China Urban Statistical Yearbook, China Urban Construction Statistical Yearbook, China Financial Yearbook, public reports on annual accounts of finance bureaus or people’s governments of prefectural-level cities, and the State Intellectual Property Office. The results of descriptive statistics for each variable are displayed in Table 2. Table 2Results of statistical description of variables.NameVariablesObsMeanS.D.Min.P25P50P75Max.VIFEcological environmental resilienceENR33960.05070.03920.00750.02450.03940.06040.1931—New energy demonstration urban constructionNEDC33960.16170.36820.00000.00000.00000.00001.00001.06Financial conditionsFIC33961.52780.43050.77611.21241.45511.76953.06511.42Population agglomerationPAG33961.62011.27230.06780.66071.29902.32858.22851.73Logarithm of per capita GDPLGGDP339610.78680.56149.454710.390710.771411.180912.07564.70Share of the tertiary sectorTES33960.42850.09890.2203035950.42540.48790.72131.96Government interventionGVI33960.20070.09600.07860.13550.17560.23760.60682.69R&D intensityRDI33960.00280.00240.00020.00120.00200.00360.013851.51Level of internet infrastructureINI33965.27210.75013.63864.75585.22575.70267.70061.68Number of universitiesNOU33960.42401.32550.00000.00000.00000.00009.00001.51Comprehensive pollution control capacityPCC33960.91690.10120.51330.89570.95710.98141.00001.37Urbanization rateURR33960.56880.14810.26680.46430.54970.66250.94983.03 Table 2Results of statistical description of variables.NameVariablesObsMeanS.D.Min.P25P50P75Max.VIFEcological environmental resilienceENR33960.05070.03920.00750.02450.03940.06040.1931—New energy demonstration urban constructionNEDC33960.16170.36820.00000.00000.00000.00001.00001.06Financial conditionsFIC33961.52780.43050.77611.21241.45511.76953.06511.42Population agglomerationPAG33961.62011.27230.06780.66071.29902.32858.22851.73Logarithm of per capita GDPLGGDP339610.78680.56149.454710.390710.771411.180912.07564.70Share of the tertiary sectorTES33960.42850.09890.2203035950.42540.48790.72131.96Government interventionGVI33960.20070.09600.07860.13550.17560.23760.60682.69R&D intensityRDI33960.00280.00240.00020.00120.00200.00360.013851.51Level of internet infrastructureINI33965.27210.75013.63864.75585.22575.70267.70061.68Number of universitiesNOU33960.42401.32550.00000.00000.00000.00009.00001.51Comprehensive pollution control capacityPCC33960.91690.10120.51330.89570.95710.98141.00001.37Urbanization rateURR33960.56880.14810.26680.46430.54970.66250.94983.03 Results of statistical description of variables. Empirical results and analysis Benchmark regression results and discussion The gradient boosting algorithm and a sample split ratio of 1:4 are used to predict the impact of NEDC on ENR. The regression results of the cross-introducing control variables, time fixed effects, and city fixed effects are presented in columns (1)–(3) of Table 3. The regression coefficients of NEDC are all positive and significant at the 1% level. The significant regression coefficient for NEDC is 0.0024 in column (3) of Table 3. This coefficient corresponds to 4.73% of the mean and 0.061 times the standard deviation of the ENR sample, representing a robust marginal improvement effect. In practical terms, this increase in the index corresponds to an average increase of approximately 1.75 square kilometers in the city’s wetland area, an average annual decrease of approximately 0.10 µg/m³ in PM2.5 concentration of approximately 0.10 µg/m³, and an average increase of about 0.3 urban green patent applications. This indicates that the policy has effectively promoted a gradual improvement in regional ecological resilience by optimizing ecological conditions, alleviating environmental pressures, and strengthening response capabilities, demonstrating clear economic value and practical significance. Its economic significance is comparable to that reported by Yang et al.77 and Wang and Yi78. This suggests that the NEDC effectively enhances ENR. Hypothesis H1 is tested. The NEDC accelerates the development of renewable resources such as wind and solar energy. New energy is widely used in various fields of production and life such as heating and power supply, prompting the transformation of the energy structure79. Fossil energy use is minimized, and pollution (carbon) emissions are reduced14,77. The NEDC also promotes the development of a green economy78 and eases the pressure of ecological restoration. With the promotion of NEDC, the self-healing capacity of the eco-environmental system is improved, and the threshold of resistance to external risk shocks is raised. ENR is strengthened. Scholars have examined the environmental governance of policies for building new energy demonstration cities77, urban green economic growth78, green technological innovation14, and industrial transformation and upgrading14. This study found that the NEDC effectively enhances ENR. The findings not only align with existing literature but also complement previous research conclusions. Table 3Benchmark regression results.Variables(1)(2)(3)NEDC0.0060***0.0027**0.0024**(0.0019)(0.0012)(0.0011)Control variablesNoYesYesUrban fixed effectYesYesYesTime fixed effectYesNoYesObs339633963396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Table 3Benchmark regression results.Variables(1)(2)(3)NEDC0.0060***0.0027**0.0024**(0.0019)(0.0012)(0.0011)Control variablesNoYesYesUrban fixed effectYesYesYesTime fixed effectYesNoYesObs339633963396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Benchmark regression results. **, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Results of the robustness test Placebo test A placebo test is used to rule out the possibility that the conclusions of this paper are due to random variation rather than causal relationships. First, events are randomly assigned to sample observations, and virtual policy variables are regenerated, followed by regression analysis. This process is repeated 1000 times to obtain 1000 regression coefficients, and their distribution is plotted in Fig. 2. The regression results show that most placebo test coefficients are concentrated around zero, and the p-values of the estimates are greater than 0.1. This indicates that the findings of this study are more likely to represent a causal relationship rather than a random phenomenon. Fig. 2Placebo test results. Placebo test results. Test for parallel trends The use of the difference-in-differences (DID) model requires that the parallel trends assumption be satisfied. To verify whether the trends in the outcome variable were essentially consistent between the treatment and control groups prior to policy implementation—thereby ensuring the validity of the causal identification in the DID estimates—a parallel trends test was conducted. Figure 3 reports the results of the parallel trends test; the coefficients for all periods prior to policy implementation were insignificant, indicating that the parallel trends assumption holds. Fig. 3Results of the Parallel Trends Test. Results of the Parallel Trends Test. Abbreviation test To further avoid estimation bias caused by outliers in the sample, the explanatory variable and all control variables were truncated at the 5% level. Column (1) in Table 4 shows the results of regression where the sample is shrunk. The results show that the effect of NEDC on ENR is 0.0035 and significant at the 1% level. This suggests that the NEDC significantly improves ENR. The robustness of the benchmark regression results is verified. Province-time interaction effect Under the influence of China’s administrative mechanism, the policy regimes among regions show large differences. However, there is a high degree of similarity in the economic base, resource endowment and industrial structure of cities within the same province. Provincial characteristics also change dynamically over time. This suggests that urban characteristics can be influenced by provincial differences. Therefore, province-time interaction effects are added in the model. The estimation results are displayed in column (2) of Table 4. The effect of NEDC on ENR is 0.0024 and significant at 5% level. The robustness of the benchmark regression results is verified. Test results with the addition of the quadratic term of the control variable Some of the city characteristics are considered in the baseline regression model. In fact, there are complex interactions between systems such as social systems and ecosystems. This makes the effect of urban characterization variables on ENR likely exhibit nonlinearity. Therefore, the quadratic terms of the control variables are included in the baseline regression model to enhance the accuracy of the model fit. The regression results are displayed in column (3) of Table 4. The effect of NEDC on ENR is 0.0026 and significant at 5% level. The robustness of the benchmark regression results is verified. Results of adjusting the sample split ratio The 1:4 sample split ratio is used for estimation in Table 3. The 1:4 sample split ratio is adjusted to 1:3 to avoid estimation bias caused by the sample split ratio setting. The estimation results are displayed in column (4) of Table 4. When the sample split ratio is adjusted, the effect of NEDC on ENR is 0.0032 and significant at the 1% level. This indicates that the benchmark regression results are robust. Table 4Robustness test results: Ⅰ.Variables(1)(2)(3)(4)NEDC0.0035***0.0024**0.0026**0.0032***(0.0013)(0.0012)(0.0011)(0.0012)Control variablesYesYesYesYesQuadratic term of the control variablesYesProvince -time interaction effectYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs3396339633963396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Table 4Robustness test results: Ⅰ.Variables(1)(2)(3)(4)NEDC0.0035***0.0024**0.0026**0.0032***(0.0013)(0.0012)(0.0011)(0.0012)Control variablesYesYesYesYesQuadratic term of the control variablesYesProvince -time interaction effectYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs3396339633963396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Robustness test results: Ⅰ. **, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Replacement of the prediction model The gradient boosting algorithm is used to predict the impact of NEDC on ENR in the benchmark regression. Given that the gradient boosting algorithm is set randomly, this may lead to biased estimates. Therefore, the gradient boosting algorithm is adjusted to lasso regression, neural network, support vector machine and elastic network algorithms. The regression results are displayed in Table 5. Under the LassoCV (Lasso regression with cross-validation), ElasticNetCV (Elastic Net with cross-validation), and RidgeCV (Ridge regression with cross-validation), the impacts NEDC on ENR are 0.0015, 0.0015, and 0.0013, and significant at the 10% level. This demonstrates that after the prediction algorithm is adjusted, the NEDC still effectively ENR. The robustness of the benchmark regression results is verified. Table 5Robustness test results: Ⅱ.VariablesLassoCVElasticNetCVRidgeCV(1)(2)(3)NEDC0.0015**0.0015**0.0013*(0.0007)(0.0007)(0.0007)Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Table 5Robustness test results: Ⅱ.VariablesLassoCVElasticNetCVRidgeCV(1)(2)(3)NEDC0.0015**0.0015**0.0013*(0.0007)(0.0007)(0.0007)Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Robustness test results: Ⅱ. *** indicate significance at the 1% levels. Robust standard errors are in parentheses. Revising the methodology for constructing the ENR indicator The ENR metric was constructed using an entropy-based method. To ensure that the results do not rely excessively on the method used to construct the metric, the metric construction method was changed. ENR1 was constructed using principal component analysis, while ENR2 was constructed using a combination of the entropy method and the equal-weighting method. The entropy method determines the importance of indicators based on the degree of data dispersion; principal component analysis extracts common information based on the contribution of variance; and the equal-weighting method assumes that all indicators contribute equally. These three weighting approaches have different underlying logics, which can lead to certain differences in the index calculation results. Columns (1)–(3) in Table 6 present the estimated results for indicator ENR1. The effects of NEDC on ENR1 are 0.2030, 0.0393, and 0.0354, respectively, and all are statistically significant at the 1% level. Columns (4)–(6) in Table 6 present the estimated results for indicator ENR2. The effects of NEDC on ENR2 are 0.0020, 0.0008, and 0.0008, respectively, and all are statistically significant at the 5% level. This indicates that changing the weighting scheme does not substantially alter the core regression findings, suggesting that the research results are not influenced by the method of index weighting and that the index calculation is both reasonable and reliable. Table 6Robustness test results: Ⅲ.VariablesENR1ENR2(1)(2)(3)(4)(5)(6)NEDC0.2030***0.0393***0.0354***0.0020***0.0008**0.0008**(0.0249)(0.0110)(0.0111)(0.0006)(0.0004)(0.0004)Control variablesNoYesYesNoYesYesUrban fixed effectYesYesYesYesYesYesTime fixed effectYesNoYesYesNoYesObs339633963396339633963396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Table 6Robustness test results: Ⅲ.VariablesENR1ENR2(1)(2)(3)(4)(5)(6)NEDC0.2030***0.0393***0.0354***0.0020***0.0008**0.0008**(0.0249)(0.0110)(0.0111)(0.0006)(0.0004)(0.0004)Control variablesNoYesYesNoYesYesUrban fixed effectYesYesYesYesYesYesTime fixed effectYesNoYesYesNoYesObs339633963396339633963396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Robustness test results: Ⅲ. **, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Robustness tests based on propensity score matching (PSM-DID and PSM-DML) Although the DML method mitigates omitted variable bias to some extent by screening for high-dimensional control variables, systematic selection bias may exist given that the selection of pilot cities was not entirely random. That is, pilot cities may inherently be superior to non-pilot cities in terms of certain observable characteristics. To construct a more rigorous counterfactual framework, this section further employs propensity score matching to preprocess the sample. Specifically, based on cross-sectional data from the period immediately preceding policy implementation (2013), we use a Logit model to estimate the probability of each city being selected for the pilot program. We then employ a kernel matching method to identify control cities with the most similar characteristics to the treatment group, thereby constructing a highly comparable control group. Figure 4 compares the standardized deviations of various covariates before and after propensity score matching. As shown, there were significant differences between the treatment group and the control group on multiple variables prior to matching. After matching, the standardized deviations for all variables decreased substantially, with most deviations falling within 10%. This indicates that matching effectively improved the balance between groups and satisfied the parallel assumptions of PSM. Fig. 4PSM Standardized Deviation Chart. PSM Standardized Deviation Chart. Table 7 presents the results of the PSM-DID and PSM-DML regressions based on the PSM-matched sample. As shown, the coefficient of the key explanatory variable NEDC is significantly positive under both methods, consistent with the findings of the baseline regression. This indicates that, after mitigating sample selection bias, the policies of new energy demonstration cities still have a significant positive effect on urban ecological resilience. Specifically, the coefficient of NEDC in the PSM-DID model is 0.0010, which is significant at the 5% level; whereas the coefficient in the PSM-DML model increases to 0.0023 and is highly significant at the 5% level. This suggests that the dual machine learning method can more effectively capture policy effects in matched samples and enhance the statistical significance of the estimates. Table 7Robustness Test Results for PSM-DID and PSM-DML.VariablesPSM-DIDPSM-DML(1)(2)NEDC0.0010**0.0023**(0.0005)(0.0011)Control variablesYesYesUrban fixed effectYesYesTime fixed effectYesYesObs33843396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Table 7Robustness Test Results for PSM-DID and PSM-DML.VariablesPSM-DIDPSM-DML(1)(2)NEDC0.0010**0.0023**(0.0005)(0.0011)Control variablesYesYesUrban fixed effectYesYesTime fixed effectYesYesObs33843396**, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Robustness Test Results for PSM-DID and PSM-DML. **, *** indicate significance at the 5%, 1% levels. Robust standard errors are in parentheses. Results of the mechanism test According to the theoretical mechanism, industrial structure upgrading, green technology innovation and energy consumption are the paths of NEDC influencing ENR. To reveal the intrinsic pathways of NEDC affecting ENR, the mechanisms are tested. Specifically, two methods were employed to test for mediating effects. First, a two-stage machine learning model was used to examine whether there were significant causal effects between NEDC and the three mediating pathways. After confirming the validity of the first half of the mediation path, robustness tests were conducted using hierarchical regression mediation analysis and bootstrap mediation tests. (1) Industrial structural upgrading (ISG), Referring to Jiang et al.80 and Zhao et al.81, the vector pinch-angle method is used to compute the ISG. The formula is shown in Eq. (14)–(15). (2) Green technology innovation (GTI), with reference to Liu et al.82, the total number of green patents granted is used to characterize green technology innovation. In the empirical tests, the variable is added to 1 and logarithmized. (3) Energy consumption (ENC), referring to Liu et al.82, electricity consumption, natural gas and liquefied petroleum are converted to tons of standard coal based on conversion factors. Further, referring to Ren et al.83, the ratio of tons of standard coal to GDP is used to capture energy consumption.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _{{\mathrm{j}}} = \arccos \left( {\frac{{\sum\nolimits_{{i = 1}}^{3} {\left( {X_{{i,j}} \cdot X_{{i,0}} } \right)} }}{{\left( {\sum\nolimits_{{i = 1}}^{j} {\left( {X_{{i,j}}^{2} } \right)^{{1{\text{ /}}2}} \cdot \sum\nolimits_{{i = 1}}^{j} {\left( {X_{{i,0}}^{2} } \right)^{{1{\text{ /}}2}} } } } \right)}}} \right),\quad j = 1,2,3$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ISP=\sum\limits_{{k=1}}^{3} {\sum\limits_{{j=1}}^{k} {{\theta _j}} }$$\end{document} In Eq. (15), larger values of ISP indicate a more advanced industrial structure. To clarify the transmission mechanism from NEDC to ENR, this paper conducts a mechanism test by strictly adhering to the causal mediation analysis framework proposed by VanderWeele84 and the serial ignorability identification criterion proposed by Acharya et al.85. First, although the three mediating variables selected in this study—ISG, GTI, and ENC—are endogenous variables formed after policy implementation, they effectively satisfy the core assumption of serial ignorability from both an institutional and empirical perspective: First, NEDC cities were selected through a top-down, nationally coordinated process based on their pre-pilot resource endowments and developmental foundations, which are independent of the dynamic evolution of the mediating variables following the pilot phase. Furthermore, parallel trends held during the pre-policy period, satisfying the first-stage negligible-effect assumption; Second, this study employs two-way fixed effects to account for urban heterogeneity and macroeconomic shocks over time, while utilizing the DML method to pre-control for high-dimensional confounding variables, combined with a set of multidimensional city-level baseline covariates. By holding policy status, ex ante characteristics, and mediating variables constant, residual confounding interference is largely eliminated, satisfying the second-stage neglectability and effectively mitigating control bias arising from post-policy variables. Industrial structure upgrading The regression result of the mechanism of industrial structure upgrading is displayed in column (1) and column (4) of Table 7. The results show that the impact of NEDC on industrial structure upgrading is 0.0425 and significant at 5% level. After changing the algorithm, the results show that the impact of NEDC on industrial structure upgrading is 0.0127 and significant at 10% level. This suggests that the NEDC effectively contributes to the industrial structure upgrading, thereby enhancing ENR. Hypothesis H2 is tested. The main goals of the NEDC are to increase the production and consumption of renewable energy. The relevant industry chains such as smart grids, new energy vehicles and bioenergy are promoted. Under the influence of the green development model and the concept of environmental protection, other upstream and downstream industries are driven, prompting the industrial structure upgrading86. Industrial structure upgrading optimizes energy supply structure87 and improves energy efficiency. Pollution emissions are reduced. The carrying pressure on the ecosystem is eased and its adaptive capacity is enhanced, improving ENR. Green technology innovation The regression result of the mechanism of green technology innovation is displayed in column (2) and column (5) of Table 7. The results show that the impact of NEDC on green technology innovation is 0.1308 and significant at 1% level. After changing the algorithm, the results show that the impact of NEDC on green technology innovation is 0.1391 and significant at 1% level. The NEDC effectively improves green technological innovation, thereby enhancing ENR. Hypothesis H3 is tested. The fundamental goal of the NEDC is to reduce pollution emissions. This not only improves traditional end-of-pipe governance, but also incentivizes enterprises to expand green R&D investments and optimize resource allocation88,89. The green technology innovation is increased82. Knowledge spillovers from green technologies can reduce pollution emissions and carbon emissions90. This eases the pressure of eco-environmental self-repair and enhances ENR. Energy consumption The regression result of the mechanism of energy consumption are displayed in column (3) and column (6) of Table 8. The results show that the impact of NEDC on energy consumption is -0.2724 and significant at 1% level. After changing the algorithm, the results show that the impact of NEDC on energy consumption is -0.2137 and significant at 5% level. The NEDC reduces energy consumption, thus enhancing the ENR. Hypothesis H4 is tested. The NEDC effectively improves energy efficiency86 and reduces energy intensity91. Pollution emissions and carbon emissions are reduced17,78 and the eco-environmental quality is improved. This is conducive to alleviate ecological pressure and improve ecological resilience and elasticity23,92. Therefore, the NEDC reduces fossil energy consumption and leads to the transformation of the energy structure, thereby improving ENR. Table 8Results of the mechanism test.VariablesGradient boosting algorithmSupport vector machine algorithmISGGTIENCISGGTIENC(1)(2)(3)(4)(5)(6)NEDC0.0425**0.1308***-0.2724***0.0127*0.1391***-0.2137**(0.0100)(0.0316)(0.0519)(0.0070)(0.0304)(0.0508)Control variablesYesYesYesYesYesYesUrban fixed effectYesYesYesYesYesYesTime fixed effectYesYesYesYesYesYesObs339633963396339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Table 8Results of the mechanism test.VariablesGradient boosting algorithmSupport vector machine algorithmISGGTIENCISGGTIENC(1)(2)(3)(4)(5)(6)NEDC0.0425**0.1308***-0.2724***0.0127*0.1391***-0.2137**(0.0100)(0.0316)(0.0519)(0.0070)(0.0304)(0.0508)Control variablesYesYesYesYesYesYesUrban fixed effectYesYesYesYesYesYesTime fixed effectYesYesYesYesYesYesObs339633963396339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Results of the mechanism test. *** indicate significance at the 1% levels. Robust standard errors are in parentheses. Hierarchical regression mediation analysis and bootstrap testing To further explore the underlying causal framework, this study employed hierarchical regression mediation analysis to test the mediating effect and conducted a bootstrap mediation test based on these results. Regarding ISG, the results of the regression analysis are detailed in Table 9. The results in columns (1) through (3) indicate that NEDC effectively promotes ISG, thereby enhancing ENR; ISG serves as an effective mediating pathway. Furthermore, the bias-corrected 95% confidence interval for the bootstrap mediation test is [0.0001, 0.0005], indicating that the mediating effect of ISG is significant. Table 9Results of the industrial structure upgrading mechanism test.VariablesENRISGENR(1)(2)(3)NEDC0.0029***0.0347***0.0026***(0.0008)(0.0083)(0.0008)ISG0.0084***(0.0014)Bootstrap test[0.0001, 0.0005]Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Table 9Results of the industrial structure upgrading mechanism test.VariablesENRISGENR(1)(2)(3)NEDC0.0029***0.0347***0.0026***(0.0008)(0.0083)(0.0008)ISG0.0084***(0.0014)Bootstrap test[0.0001, 0.0005]Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Results of the industrial structure upgrading mechanism test. *** indicate significance at the 1% levels. Robust standard errors are in parentheses. Regarding GTI, the results of the regression analysis are detailed in Table 10. The results in columns (1) through (3) indicate that NEDC effectively promotes GTI, thereby enhancing ENR; GTI serves as an effective mediating pathway. Furthermore, the bias-corrected 95% confidence interval for the bootstrap mediation test is [0.0003, 0.0007], indicating that the mediating effect of GTI is significant. Table 10Results of the green technology innovation mechanism test.VariablesENRGTIENR(1)(2)(3)NEDC0.0029***0.3340***0.0024***(0.0008)(0.0462)(0.0008)GTI0.0016***(0.0002)Bootstrap test[0.0003, 0.0007]Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Table 10Results of the green technology innovation mechanism test.VariablesENRGTIENR(1)(2)(3)NEDC0.0029***0.3340***0.0024***(0.0008)(0.0462)(0.0008)GTI0.0016***(0.0002)Bootstrap test[0.0003, 0.0007]Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Results of the green technology innovation mechanism test. *** indicate significance at the 1% levels. Robust standard errors are in parentheses. Regarding ENC, the results of the regression analysis are detailed in Table 11. The results in columns (1) through (3) indicate that NEDC effectively promotes ENC, thereby enhancing ENR; ENC serves as an effective mediating pathway. Furthermore, the bias-corrected 95% confidence interval for the bootstrap mediation test is [0.0001, 0.0005], indicating that the mediating effect of ENC is significant. Table 11Results of the energy consumption mechanism test.VariablesENRENCENR(1)(2)(3)NEDC0.0029***-0.2926***0.0026***(0.0008)(0.0777)(0.0008)ENC-0.0009***(0.0002)Bootstrap test[0.0001, 0.0005]Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Table 11Results of the energy consumption mechanism test.VariablesENRENCENR(1)(2)(3)NEDC0.0029***-0.2926***0.0026***(0.0008)(0.0777)(0.0008)ENC-0.0009***(0.0002)Bootstrap test[0.0001, 0.0005]Control variablesYesYesYesUrban fixed effectYesYesYesTime fixed effectYesYesYesObs339633963396*** indicate significance at the 1% levels. Robust standard errors are in parentheses. Results of the energy consumption mechanism test. *** indicate significance at the 1% levels. Robust standard errors are in parentheses. Heterogeneity discussion of NEDC affecting ENR Discussion of geographic location The sample is divided into two groups (the Midwest and the East), according to the geographic location published by the National Bureau of Statistics of China. The results are displayed in Table 12 column (1) and column (2). The NEDC has no a significant impact on the ENR in the central-western regions. The effect of the NEDC on the ENR is 0.0037 and significant at the 10% level in the eastern region. This demonstrates that the NEDC only enhances the ENR in the eastern region compared to the central-western regions. The main explanation is, first, that there are significant disparities in the availability of energy infrastructure between the eastern and western regions. The eastern region boasts a mature and stable energy infrastructure network (such as power grids, oil and gas pipelines, and district heating networks), as well as superior financial and technological resources. It can provide strong support for the construction and stable operation of new energy facilities. However, energy infrastructure in central and western regions is relatively underdeveloped, making it difficult to construct and utilize new energy facilities. The challenge in promoting the development of new energy cities in these regions lies in the difficulty of enhancing ENR in the short term. Second, the eastern region has a more developed economy. There are the well-developed transportation and information network, and a favorable geographical location in the eastern region. This not only prompts the knowledge mobility and enhances the technology diffusion, but also facilitates real-time exchanges among various subjects (government, enterprises, and the public) and accelerates the derivation of new ideas and thoughts93. Third, the accumulation of these differences will also lead to disparities in the effectiveness of energy policy implementation between cities in the east and west. For cities in the eastern region, the agglomeration of high-quality factor resources can shorten the research and development time of renewable energy, and improve the use of new energy areas and efficiency. When the NEDC is implemented, the government and enterprises adjust resource allocation and increases technological innovation. Renewable energy production and use is expanded, reducing pollution shocks and increasing ENR. The central-western regions are located in the interior of China. Economic development depends on labor-intensive industries. The lack of quality resources also constrains the development of this region. There are major dilemmas in implementing new energy demonstration policies. Enterprises innovative behavior is also constrained. This suggests that the NEDC hardly improves the energy structure and environmental quality. ENR is not improved. The sample is divided into two groups (third-fourth-fifth tier cities and first- second tier cities), according to the list in China’s City Charm Ranking. The results are displayed in Table 12 column (3) and column (4). The NEDC has no significant impact on ENR in the third-fourth-fifth tier cities. The effect of NEDC on ENR is 0.0047 and significant at the 10% level in the first-second tier cities. The main reason is the disparity in socioeconomic characteristics between first- and second-tier cities and third-, fourth-, and fifth-tier cities. First, there are differences in industrial structure and technological capabilities, with mid- to high-end industries concentrated in first- and second-tier cities. This attracts large inflows of high-quality factor resources and accelerates the knowledge accumulation and technology. The differences in infrastructure development and resource concentration; first- and second-tier cities have more comprehensive and well-developed public infrastructure94. With the gradual advancement of the NEDC, the government expands investment of new energy development and accelerates the energy transition. Enterprises actively raise investment in renewable energy technology research and accelerate the transformation of technological achievements. This reduces pollution emissions and enhances the recycling efficiency of natural resources, thereby increasing ENR. The economic base is poor. Quality resources are relatively lack in third-fourth-fifth tier cities. In particularly, the siphoning effect generated by first-second tier cities results in a lack of quality resources. This causes the NEDC to have no significant impact on ENR. Table 12Results of the heterogeneity regression: geographic location.VariablesCentral-western and eastern RegionThird-fourth-fifth and first-second tier citiesCentre-westEastThird-fourth-fifth tierFirst-second tier(1)(2)(3)(4)NEDC−0.00020.0037*−0.00120.0047*(0.0014)(0.0021)(0.0013)(0.0025)Control variablesYesYesYesYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs219612002808588** indicate significance at the 5% levels. Robust standard errors are in parentheses. Table 12Results of the heterogeneity regression: geographic location.VariablesCentral-western and eastern RegionThird-fourth-fifth and first-second tier citiesCentre-westEastThird-fourth-fifth tierFirst-second tier(1)(2)(3)(4)NEDC−0.00020.0037*−0.00120.0047*(0.0014)(0.0021)(0.0013)(0.0025)Control variablesYesYesYesYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs219612002808588** indicate significance at the 5% levels. Robust standard errors are in parentheses. Results of the heterogeneity regression: geographic location. ** indicate significance at the 5% levels. Robust standard errors are in parentheses. Resource-richness discussion More water resources mean more water energy reserves. According to the median river length, the sample is divided into lower and higher water resource groups. The results are displayed in column (3) and column (4) of Table 13. The effect of NEDC on ENR was not significant in the group with lower water resources, but was 0.0039 in the group with higher water resources, and was statistically significant at the 1% significance level. This indicates that the NEDC’s promotional effect on ENR is not significant in regions with scarce water resources, whereas it effectively enhances ENR in regions with abundant water resources. The primary reason lies in the fact that abundant water resources offer immense potential for hydropower generation, which provides a foundation for new energy demonstration cities to vigorously promote hydropower development and water transportation95,96. On the one hand, regions with abundant water resources can leverage high-quality hydropower resources to optimize regional energy supply structures, thereby facilitating the substitution of traditional, energy-intensive sources with clean energy. On the other hand, these regions can utilize their ample water resources to support the green transformation of the energy sector, effectively alleviating pressure on both the ecological environment and the energy system. These factors collectively improve the quality of ecological and environmental systems and enhance their restorative capacity, ultimately fostering a virtuous cycle among various systems and thereby increasing ecological resilience. However, in cities with relatively scarce water resources, insufficient hydropower reserves and a lack of adequate clean hydropower resources as a practical vehicle for factor substitution make it difficult for these cities to optimize and adjust their energy structures by leveraging their water resource advantages. Furthermore, water scarcity itself exacerbates regional ecological constraints and resource supply pressures, which in turn further constrain the space for the transformation and upgrading of the energy system. This makes it difficult to effectively unleash the core mechanisms of factor substitution and pressure relief during policy implementation. Policy benefits are easily offset by inherent local resource constraints, and the role of pilot policies in boosting energy resilience is significantly hampered. Ultimately, this manifests as the NEDC policy’s effect on improving ENR being statistically insignificant in water-scarce regions. According to the list in the National Sustainable Development Plan for Resource-Based Cities (2013–2020), the sample is categorized into resource-based and non-resource-based cities. The results are displayed in columns (5) and (6) of Table 13. The impact of NEDC on ENR is not significant in resource-based cities, but is 0.0055 in non-resource-based cities, which is statistically significant at the 1% significance level. This indicates that the NEDC has a negligible effect on promoting ENR in resource-based cities, whereas it significantly enhances ENR in non-resource-based cities. The primary reason lies in the fact that resource-based cities have long relied on traditional resources such as minerals and fossil fuels, leading to path dependence that has left these cities mired in the “resource curse” dilemma97. At the same time, resource-based cities have a heavily skewed industrial structure, making transformation particularly challenging. Existing energy-intensive industries crowd out the development space for new energy, making it difficult to rapidly complete the substitution and transition of energy inputs. Furthermore, these cities face long-term ecological degradation and development pressures resulting from resource extraction, bearing a heavy burden of ecological restoration and energy system optimization, which makes it difficult for policy implementation to effectively alleviate these pressures. Traditional development inertia has significantly weakened the enabling effect of the NEDC pilot policy, making it difficult for the policy to demonstrate a positive impact on energy resilience, and resulting in insignificant outcomes. In contrast, non-resource-based cities have relatively diversified industrial structures and have accumulated a large number of high-quality production factors, significantly promoting knowledge accumulation and iterative upgrading98. With lower dependence on traditional fossil fuels, these cities face less resistance to industrial transformation. Non-resource-based cities can smoothly leverage the NEDC initiative to advance energy structure adjustments, replacing traditional high-carbon energy sources with clean energy to achieve factor substitution. At the same time, by utilizing green development strategies to alleviate dual pressures on energy supply and ecological carrying capacity, policy measures can be fully implemented and take effect, continuously driving the enhancement of energy resilience. Table 13Results of the heterogeneity regression: Resource-richness.VariablesWater resourcesResource-based cityLowHighYesNo(3)(4)(5)(6)NEDC−0.00360.0039***−0.00230.0055***(0.0023)(0.0013)(0.0020)(0.0014)Control variablesYesYesYesYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs1704169213442052***, ** indicate significance at the 1%, 5% levels. Robust standard errors are in parentheses. Table 13Results of the heterogeneity regression: Resource-richness.VariablesWater resourcesResource-based cityLowHighYesNo(3)(4)(5)(6)NEDC−0.00360.0039***−0.00230.0055***(0.0023)(0.0013)(0.0020)(0.0014)Control variablesYesYesYesYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs1704169213442052***, ** indicate significance at the 1%, 5% levels. Robust standard errors are in parentheses. Results of the heterogeneity regression: Resource-richness. ***, ** indicate significance at the 1%, 5% levels. Robust standard errors are in parentheses. Discussion of ENR sub-indices and NDVI To avoid potential issues of endogeneity and circularity within the ENR indicator, clarify the multidimensional impact pathways of policies, and validate the empirical findings. Following the ENR index construction methodology, the ENR was decomposed into three major sub-dimensions—state, pressure, and response—for separate regression analysis to identify the policy’s primary focus. Additionally, the normalized difference vegetation index (NDVI), which is unaffected by pollution emission data, was selected as a substitute dependent variable for re-estimation. This approach eliminates mechanical associations resulting from policy-driven emission reductions, thereby accurately identifying the policy’s actual ecological effects. The results are presented in columns (1) and (4) of Table 14. The NEDC’s impact on state resilience is 0.0004, on pressure resilience is 0.0030, and on response resilience is 0.0043. Among these, the estimated effect on state resilience did not pass the significance test, while the effect on stress resilience was significant at the 10% level, and the effect on response resilience was significant at the 1% level. This indicates that the NEDC’s role in enhancing urban ecological resilience exhibits significant dimensional heterogeneity. The policy did not significantly improve the state resilience of urban ecosystems, suggesting that substantive improvements in ecological conditions may be subject to a lag effect; in contrast, the policy was highly effective in enhancing urban coping capacity and regulatory mechanisms, primarily by alleviating external environmental pressures and stimulating urban governance and institutional responses. Furthermore, regression results using the NDVI as a substitute for the dependent variable show that the coefficient for NEDC is 0.0146 and is significant at the 1% level. This result further confirms the robustness of the baseline conclusions. This effectively rules out potential endogeneity and mechanical redundancy issues in the construction of the indicators, accurately identifying the genuine ecological benefits brought about by the policy. Table 14Results of the heterogeneity regression: ENR Sub-Indices and NDVI.VariablesState resiliencePressure resilienceResponse resilienceNDVI(1)(2)(3)(4)NEDC0.00040.0030*0.0043***0.0144***(0.0013)(0.0016)(0.0010)(0.0053)Control variablesYesYesYesYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs3396339633963396***, ** indicate significance at the 1%, 5% levels. Robust standard errors are in parentheses. Table 14Results of the heterogeneity regression: ENR Sub-Indices and NDVI.VariablesState resiliencePressure resilienceResponse resilienceNDVI(1)(2)(3)(4)NEDC0.00040.0030*0.0043***0.0144***(0.0013)(0.0016)(0.0010)(0.0053)Control variablesYesYesYesYesUrban fixed effectYesYesYesYesTime fixed effectYesYesYesYesObs3396339633963396***, ** indicate significance at the 1%, 5% levels. Robust standard errors are in parentheses. Results of the heterogeneity regression: ENR Sub-Indices and NDVI. ***, ** indicate significance at the 1%, 5% levels. Robust standard errors are in parentheses. Research findings and policy implications Research conclusions In the context of sustainable development, the enhancement of ENR not only meets the growing needs of the people for a better life, but is also an important part of realizing harmonious coexistence between human beings and nature. Based on a panel data of 283 prefecture-level cities from 2011 to 2022, the dual machine learning mothed is used to predict the impact of NEDC on ENR. On average, this policy has improved the ENR indicator, with even more pronounced effects in regions with abundant natural resources. The conclusion still holds after some series of robustness tests. Mechanism analysis shows that the NEDC enhances ENR by prompting industrial structure upgrading, improving green technology innovation and reducing energy consumption. The heterogeneity results indicate that the NEDC significantly enhances the ENR in the eastern region and first-second tier cities compared to the central-western regions and third-fourth- fifth tier cities. At the same time, the NEDC does not have a significant impact on promoting ENR in water-scarce cities and resource-based cities, whereas its impact on promoting ENR in water-abundant cities and non-resource-based cities is more pronounced. This paper theoretically clarifies the important mechanisms where the NEDC shapes the stability of the structure and function of the eco-environmental system. The deep integration of state governance modernization and ecological civilization construction is also provided with empirical evidence and policy references. Policy implications First, the NEDC is an important institutional guarantee for the promotion of ENR. Local governments dynamically summarize the experience and initiatives of pilot cities and extend them to more areas according to local conditions. Based on the strategic framework of regional integration, the central government guides the establishment of joint mechanism for eco-environmental co-regulation, information sharing and co-management among regions. This improves the synergy and overall resilience of eco-environmental governance. Relevant departments avoid the practice of blindly following the trend. Based on the findings of this study, the experience gained from the NEDC could be replicated in relevant countries. The sustainable development model, in which human beings live in harmony with nature, is gradually being realized on a global scale. Secondly, with the NEDC as an opportunity, industrial structure upgrading is promoted and green technology innovation is accelerated. This prompts the transformation of old and new energy sources and promotes the development of new energy industries on a large scale and in clusters. Traditional fossil energy sources are replaced. The research and development of cutting-edge technologies of the new energy industry is intensified. The governments vigorously attract social investment to the new energy industry and realizes the transformation of green technology achievements. Relevant departments should also guide the public to actively participate in the construction of new energy demonstration cities and cultivate an awareness of energy conservation and environmental protection. In addition, a combination of policy instruments is emphasized to create policy complementarities. ENR is enhanced. Green transformation and sustainable development are progressively realized. Finally, China has a vast geographical scope and differential resource endowments. Governments and enterprises in different regions combine realistic conditions, economic structures, and resource endowments to develop more appropriate policy regimes. The share of new energy production and consumption is more effectively increased to enhance ENR. Meanwhile, in central and western regions and third- and fourth-tier cities, policies the promotes the NEDC have not effectively enhances ENR. This requires local governments and businesses to actively learn from the advanced knowledge and experience from more developed regions. Governments in eastern regions and first- and second-tier cities are encouraging relevant enterprises to share their experiences more widely, thereby improving the ecological and environmental quality in less developed regions. In cities with scarce water resources that rely heavily on natural resources, it is essential to accelerate the shift away from a resource-dependent development model and plan water-efficient clean energy projects tailored to local conditions. At the same time, efforts should be made to strengthen targeted policy support, address shortcomings in resource endowments, and steadily leverage the enabling role of policies to enhance ENR. This can enhance the policy effectiveness of NEDC. Governments and enterprises in peripheral cities take the initiative to actively learn from advancement experience, knowledge and technology. ENR is improved.