Agent-based modeling of electric vehicle diffusion under the phase-out of charging infrastructure subsidies in China

Agent-based modeling of electric vehicle diffusion under the phase-out of charging infrastructure subsidies in China Lijing Zhu a , Runze Li a , Jingzhou Wang b , Haibo Chen c , Ondrej Havran c , Wen-Long Shang d,c,* a School of Economics and Management, China University of Petroleum, Beijing, China b Department of Agricultural Economics, Sociology, and Education, Pennsylvania State University, University Park, PA, USA c Institute for Transport Studies, University of Leeds, 34-40 University Road, Leeds, LS2 9JT, UK d Centre for Transport Studies, Imperial College London, SW7 2AZ, London, UK ARTICLE INFO Keywords: Electric vehicle Charging infrastructure Subsidy phase-outs Agent-based modeling ABSTRACT Government subsidies for electric vehicle charging infrastructure (EVCI) in China have accelerated the deployment of charging stations and promoted the diffusion of electric vehicles (EVs). However, these subsidies have also imposed a substantial fiscal burden on public finances. While much of the existing literature compares different types of EVCI subsidies, few studies explore the implications of phasing out EVCI-related subsidies for government spending and EV diffusion. This paper develops an agent-based model (ABM) incorporating EVCI operator, heterogeneous EV consumers, and the government to analyze how EVCI subsidies influence EV diffusion and proposes tailored phase-out policy combinations. A key innovation of this study is the integration of private charging pile-related factors into the consumer decision-making process through a discrete choice experiment. Additionally, regional disparities in EV diffusion between urban and suburban areas under EVCI subsidies are explored, and we find that by 2030, the EV penetration rate could reach 79.78 %, with suburban EV ownership surpassing that of urban areas. While EVCI subsidies significantly influence early and mid-stage EV adoption, their effectiveness diminishes in the later stages. Implementing phase-out subsidies under current standards can reduce cumulative government spending by approximately 91 % compared to a no-phase-out scenario, with only a marginal decline of 0.05 % in EV ownership. A comparative analysis of 50 subsidy phase-out policy combinations reveals that those featuring high initial operating subsidies with low initial construction subsidies under a rapid phase-out mode are the most cost-effective. The policy recommendations proposed alleviate fiscal burdens and promote more balanced EV development between urban and suburban areas. 1. Introduction Under the dual objectives of carbon peaking and carbon neutrality, accelerating the electrification of the transportation sector has become a key pathway for China to fulfill its emission reduction commitments (Bao et al., 2023; Zhong et al., 2024). Within this electrification agenda, the electric vehicle (EV) industry has experienced rapid growth, largely driven by supportive government subsidies. By the end of 2024, the number of EVs in China had reached 31.4 million (Kong, 2025). However, the development of electric vehicle charging infrastructure (EVCI) to support EV adoption remains hindered by significant imbalances. The first challenge is a quantity imbalance. By the end of 2024, China had installed approximately 3.579 million public charging piles (PCPs) and 9.239 million private charging piles (PrCPs). 1 With an EV fleet of 31.4 million, the resulting vehicle-to-charging pile ratio is roughly 2.45:1, which falls below the targeted 1:1 ratio (National Development and Reform Commission [NDRC], 2015). The second challenge is spatial distribution imbalance (Y. Shang et al., 2025). EVCI deployment in China shows a clear regional disparity, with higher densities generally observed in the south than the north and broader coverage in the east * Corresponding author. Centre for Transport Studies, Imperial College London, SW7 2AZ, London, UK. E-mail address: [email protected] (W.-L. Shang). 1 The number of public and private charging piles in 2024 is available at: https://www.evcipa.org.cn/newsinfo/8137834.html. Contents lists available at ScienceDirect Transport Policy journal homepage: www.elsevier.com/locate/tranpol https://doi.org/10.1016/j.tranpol.2025.103876 Received 21 July 2025; Received in revised form 25 October 2025; Accepted 28 October 2025 Transport Policy 175 (2026) 103876 Available online 30 October 2025 0967-070X/© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/bync-nd/4.0/ ). relative to the west. 2 In major cities, the average EVCI coverage rate in central urban areas reaches 80.8 %, compared with less than 5.0 % in rural areas. 3 EVCI plays a crucial role in alleviating range anxiety among EV users and is widely recognized as a critical influential factor of EV adoption (X. Li and Liu, 2023; Z. Wang, 2020). The effect of EVCI deployment on EV adoption can be explained through the lens of indirect network effect (INE), which refers to the phenomenon in which a user’s utility from adopting a particular product increases with the adoption of a complementary product (Katz and Shapiro, 1985). In the EV market, EVCI serves as a complementary good to EVs, and its quantity has an indirect yet substantial impact on the user experience of EV owners (Sun et al., 2018). Limited access to EVCI remains a major obstacle to EV adoption (Y. Liu et al., 2015). Since 2015, the Chinese government has implemented a series of subsidy policies for EVCI operators to accelerate the development of charging infrastructure (W. Shang et al., 2024). In the preliminary stages, these efforts primarily targeted charging station construction subsidies (CSs), aimed at alleviating the initial investment burden on operators and thereby stimulating the expansion of charging networks. Beginning in 2016, the government implemented operational subsidies (OSs) to incentivize greater efficiency among charging station operators, and under this policy, operators receive subsidies based on the actual volume of electricity dispensed. However, the extensive EVCI subsidies have imposed a growing fiscal burden on the governments. According to reports from the Ministry of Finance of the P.R.C., 4 cumulative subsidies for EVCI amounted to 7.556 billion RMB between 2016 and 2020. In response, several regions (provinces) in China have begun adjusting their subsidy policies. For example, Hunan and Hainan Provinces have adopted phased reduction policies for EVCI subsidies, gradually lowering the subsidy levels on an annual basis. 5 To date, limited attention has been given to the reduction of EVCI subsidies and their potential consequences. Moreover, most existing research has predominantly focused on PCPs, with few studies investigating the role of PrCPs in EV diffusion. In reality, PCPs account for only 27.92 % of China’s total EVCI stock, 6 and the ability of EV users to install PrCPs is another critical factor influencing purchase decisions (Qian et al., 2019). Therefore, incorporating the role of PrCPs into studies of EV diffusion is essential to more accurately reflect real-world EV adoption patterns. In addition to incentive policies and EVCI, the diffusion of EVs is shaped by other factors such as vehicle attributes, fuel prices, and individual income levels (Su and Diao, 2025). Accordingly, the EV diffusion process can be conceptualized as a complex system in which agents, such as consumers, operators, and policymakers, interact with each other over time within a dynamic environment. Agent-based modeling (ABM) approach provides a robust framework for analyzing such systems. By constructing micro-level agents with heterogeneous attributes and decision-making rules, ABM allows for autonomous decision-making and interactions within a defined environment, thereby enabling the exploration of diffusion pathways over time (Mehdizadeh et al., 2022). Building on this, this study develops an ABM framework incorporating heterogeneous EV consumers, EVCI operator, and the government to examine the impacts of EVCI incentive policies on the diffusion of EVs. The main contributions of this study are as follows: First, it examines the impact of EVCI subsidy phase-out on both EV diffusion and government fiscal expenditure, while also conducting a comparative analysis of alternative phase-out strategies. To the best of our knowledge, it represents the first quantitative evaluation of different EVCI subsidy withdrawal scenarios, which thereby offers valuable policy insights for the design of the phased subsidy reduction. Second, this study examines the spatial distribution of EVCI by comparing the diffusion patterns of EVs and charging piles in urban versus suburban areas and identifies effective policy strategies to address persistent imbalances in EVCI deployment across different geographic regions. Third, this study examines the role of PrCPs in consumer EV purchase decisions by integrating PrCPs installation into the decision-making process within an ABM framework and further simulates the diffusion of PrCPs, which addresses a notable gap in the existing literature on private charging infrastructure. The remainder of this paper is organized as follows: Section 2reviews the relevant literature on EVCI subsidies and EV diffusion modeling. Section 3outlines the ABM framework. Section 4details the parameter settings and model validation. Section 5presents simulation results and conducts sensitivity analysis. Finally, Section 6concludes with the main findings. 2. Literature review 2.1. EV and EVCI policies The rapid expansion of EVs in the Chinese market has been largely driven by strong policy support. To encourage EV adoption, the government has primarily relied on a series of policies such as license plate restrictions, EV purchase tax exemptions, and EV purchase subsidies. The effectiveness of these policy instruments has been extensively documented in the existing literature (Shen et al., 2021; Sun et al., 2018; T. Zhang et al., 2024). While license plate control policies can strongly influence EV adoption, their impact is geographically restricted as they are implemented in only a limited number of cities. By contrast, purchase tax exemptions and purchase subsidies are applied more broadly and play a more important role on EV diffusion (Zhu et al., 2022). However, with the rapid expansion of the EV market and growing fiscal pressures, direct purchase subsidies were gradually reduced and ultimately phased out by the end of 2022. Given the phase-out of EV purchase subsidies, scholars have increasingly shifted their attention to EVCI subsidy policies. Li et al. (2017) identified a feedback loop between EV adoption and charging infrastructure deployment. Their findings indicate that, ceteris paribus, CSs for EVCI have twice the effect on EV diffusion as EV purchase subsidies. Zhu et al. (2019) argued that, in the context of EV purchase subsidy withdrawal, the government should increase CSs for EVCI to sustain EV market growth. Zhang et al. (2024) examined the effects of different subsidy combinations, including EV purchase subsidies and EVCI CSs, on EV sales. Their findings suggest that, as the EV market matures, purchase subsidies should be gradually redirected toward supporting EVCI construction, with CSs progressively phased out in the later stages of market development. Following the introduction of OSs for charging stations, some scholars have examined the combined effects of CSs and OSs on EVCI development and diffusion. Luo et al. (2023) assessed the effectiveness of different subsidy types in EV adoption and found that the cost of acquiring one additional EV user through EV purchase subsidies is much higher than through either CSs or OSs. Ling et al. (2024) analyzed how different combinations of CSs and OSs affect the decision-making of EVCI operators and concluded that CSs are more suitable for operators with low operating cost coefficients (which measure the operational capability, with lower values indicating lower costs 2 Additional details on the spatial pattern of EVCI allocation across China can also be found at: https://www.evcipa.org.cn/newsinfo/8137834.html. 3 Coverage rates of EVCI in urban and rural areas are provided at: https:// www.cnenergynews.cn/focus/2025/03/13/detail_20250313204160.html and https://mp.weixin.qq.com/s/DAaxqE7dO0YMVBILaYj7Hw. 4 Reports by the Ministry of Finance of the P.R.C. can be accessed at: http://jj s.mof.gov.cn/zxzyzf/jnjpbzzj/202204/t20220420_3804302.htm and http://jjs. mof.gov.cn/zxzyzf/jnjpbzzj/202405/t20240507_3934118.htm. 5 Details on the phasing out EVCI subsidies in Hunan and Hainan provinces are available through: https://plan.hainan.gov.cn/sfgw/0400/201907/af 69db13fd15463b870fe8806dfd2577.shtml and https://gxt.hunan.gov.cn/gxt /xxgk_71033/zcfg/gfxwj/202212/t20221230_29171166.html. 6 Information on the number of public charging piles and the EV stock in China is sourced from the China Electric Vehicle Charging Infrastructure Promotion Alliance(EVCIPA)’s report: https://www.evcipa.org.cn/newsinfo/813 7834.html. L. Zhu et al. Transport Policy 175 (2026) 103876 2 to achieve the same level of service), whereas OSs are more effective for those with relatively high cost coefficients. Chen et al. (2023) found that the marginal effects of investment subsidies, CSs, and OSs on EVCI deployment differ, with investment subsidies exhibiting the most pronounced diminishing marginal effect. In summary, with the withdrawal of EV purchase subsidies and the implementation of charging station subsidies in China, scholars have increasingly focused on the impact of these policy shifts. However, few studies have analyzed CSs and OSs within an integrated analytical framework, and even fewer have quantitatively assessed the consequences of EVCI subsidy phase-outs. As with EV purchase subsidies, the subsidies for charging stations are also expected to be gradually phased out as the EVCI industry matures. It is therefore essential to design optimized phase-out strategies that reduce fiscal expenditure while minimizing potential adverse effects on EV development. 2.2. Indirect network effects The influence of charging infrastructure on EV diffusion can be understood through the lens of INE. First introduced by Katz and Shapiro (1985), INE describes situations in which the utility a consumer derives from a product is correlated with the availability of a complementary product. The existence and significance of INE have been confirmed across various industries, such as video games (Clements and Ohashi, 2005), DVDs (Inceoglu and Park, 2011), and television sets (X. Zhang, 2007). In recent years, numerous studies have examined the INE between EVs and EVCI, which can generally be classified into macro-level and micro-level analyses. Macro-level studies primarily assess the impact of EVCI quantity on EV adoption using aggregate data. Drawing on data from China’s EV market, Li and Liu (2023) demonstrated the positive impacts of EVCI on EV diffusion in terms of both EVCI stocks and density. Based on data from the U.S. market, Li et al. (2017) further confirmed the significant influence of INE on EV diffusion through an empirical study. Similarly, Koch et al. (2022) analyzed the INE in the Norwegian EV market by constructing an econometric model and found that its intensity is lower than in the United States. Micro-level studies, by contrast, focus on how the quantity of EVCI in a user’s vicinity affects perceived utility, typically relying on questionnaires to capture user preferences. Sun et al. (2018a,b), for example, conducted a discrete choice experiment with individual-level data and found that INE positively influences EV choice, with their heterogeneity analysis further revealing substantial variation in sensitivity to INE across different user groups. Likewise, Liu et al. (2015) confirmed the presence of INE in the EV market using questionnaire data and noted that EVCI-related INE issues hinder EV purchases. Overall, the existence and importance of INE are well established in previous research. Building on this foundation, our study incorporates INE into an ABM framework to simulate the dynamic relationships between the diffusion of EVs and EVCI. 2.3. EV diffusion research methodologies A substantial body of research has explored EV diffusion through methodologies such as the Bass model, system dynamics (SD), and ABM. The Bass model, originally proposed by Bass (1969), is widely used to quantify innovation diffusion processes and has been applied by several scholars to analyze EV adoption. For instance, Shi et al. (2022) developed a Bass model incorporating vehicle scrappage and internal combustion engine vehicle (ICEV) competition mechanisms to predict EV diffusion in Shanghai under policy influences. Yang et al. (2025) integrated uncertainty theory into the Bass model to simulate EV sales growth under disruptive events, while Fan et al. (2025) proposed a Bass model with a green premium factor to forecast EV sales in the Chinese market. Nevertheless, although the Bass model is relatively simple and data-efficient, its inability to account for external influential factors has limited its applicability to the early stages of EV diffusion research. Unlike the Bass model, SD allows for the incorporation of a wider range of influencing factors from a macro-level perspective. Li et al. (2023) developed an SD-based EV diffusion model to analyze the impacts of factors such as vehicle performance and charging convenience. Similarly, Zhu et al. (2024) constructed an EV diffusion model incorporating vehicle performance, EVCI, and policy factors, which enabled a comparative analysis of different policy measures in promoting EV diffusion. Using an SD framework that involves government, firms, and consumers, Kong et al. (2020) assessed the effects of phasing out EV purchase subsidies on EV market share. Shen et al. (2021) analyzed the impact of subsidy phase-out policies in Shanghai on local EV adoption, and similarly, Kim et al. (2021) projected EV diffusion trends in South Korean cities through an SD approach and evaluated the effectiveness of subsidy policies on EV penetration. However, neither the Bass model nor SD is capable of explicitly capturing the impact of individual-level behavior on the overall system. In contrast, the ABM approach enables the construction of multiple heterogeneous agents within a studied environment. These agents interact with each other and with their environment, thereby generating complex phenomena and patterns at the macro level (Mehdizadeh et al., 2022). Consequently, a growing number of scholars have studied EV adoption with the ABM approach. For example, Eppstein et al. (2011) developed a micro-level ABM of EV diffusion that incorporates agent behaviors shaped by media information dissemination and peer effects. Zhang et al. (2011) combined the multinomial logit model with an ABM framework to represent heterogeneous consumers based on empirical data. Other scholars have further leveraged ABM’s capabilities to analyze how social relationships among users influence overall EV development (Shafiei et al., 2012; Xu and Bi, 2024). Table 1 below summarizes the literature employing the aforementioned methods to study EV adoption. Besides, several scholars have also leveraged the spatial modeling capabilities of ABM to explore the spatial diffusion of EVs and EVCI and to forecast their spread across geographical regions. Wang et al. (2023) modeled the spatial development of the EV market from 2015 to 2040 and found that EV users gradually form clusters over time. Silvia and Krause (2016) divided an urban area into commercial and residential zones, allocating PCPs in varying quantities across these zones to reflect a realistic charging network structure. Similarly, Hu´ etink et al. (2010) adopted a spatial segmentation approach and divided the study region into urban and rural areas to analyze how different hydrogen refueling infrastructure deployment strategies influence the diffusion of hydrogen vehicles. Moreover, by integrating ABM with geographic information systems (GISs), some researchers have developed spatial environments that reflect urban characteristics, thereby enabling more accurate analyses of EV and EVCI diffusion within specific city contexts (Luo et al., 2023; Zhuge et al., 2021). Although these studies leverage ABM’s spatial modeling capabilities to segment the simulation space, the demographic characteristics assigned to different regions often rely on subjective assumptions rather than empirical evidence. Table 2 summarizes the relevant literature on EV diffusion with an ABM approach. This paper seeks to simulate the diffusion of EVs and charging stations under real-world conditions by incorporating individualized factors such as consumer location, the number of nearby PCSs, and the feasibility of installing PrCPs. The ABM approach, characterized by its bottom-up framework for modeling individual behavior to derive system-level outcomes, is particularly suited for simulating EV diffusion under heterogeneous consumer preferences. Moreover, by utilizing spatial modeling capabilities, this study analyzes the spatial dynamics of EV diffusion and proposes differentiated strategies for promoting EV uptake in urban and suburban areas based on the simulation results. L. Zhu et al. Transport Policy 175 (2026) 103876 3 3. Methodology 3.1. Agent-based modeling description This paper constructs a multi-agent system within an ABM framework, incorporating EV consumers, EVCI operator, and the government, to evaluate the impacts of policy interventions on the EV market. As vehicle performance and charging infrastructure continue to evolve, potential consumers may shift their preferences from ICEVs to EVs, thereby reshaping overall charging demand. In response to this changing demand and government subsidy policies, EVCI operator adjusts its PCS deployment strategies, which in turn further influence the development of the charging environment. Furthermore, considering the influence of PrCPs on consumer purchase decisions and the differences in installation conditions between urban and suburban areas in China, this paper adopts an urban-suburban classification to capture regional heterogeneity. Sub-models for each agent are described in detail in the following sections. 3.2. Consumer sub-model The ABM framework centers on the decision-making behavior of consumers, assuming that they follow a utility maximization approach that accounts for vehicle attributes, charging station characteristics, and individual socioeconomic profiles. Selection and parameterization of variables in the decision model are informed by statistical analysis of data obtained from a discrete choice experiment. 3.2.1. Discrete choice model This study assumes that consumers behave rationally and consider factors such as vehicle attributes, INE, and the accessibility of charging or gas stations when making purchase decisions. To assess the influence of these factors on purchase intentions, this paper employs a stated preference approach within a discrete choice experiment framework, which analyzes consumer preferences by observing choices across hypothetical purchase scenarios with varying attribute combinations (Higgins et al., 2017; Qian et al., 2019; Rudolph, 2016). Data was collected through a questionnaire comprising two main components: demographic information and a stated preferences survey. Following Patt et al. (2019), the demographic section includes questions on respondents’ income, acceptable vehicle purchase price, place of residence, and availability of private charging infrastructure. The stated preferences section focuses on vehicle attributes, INE, and PrCP installation status. Specifically, the vehicle attributes considered are purchase price, driving range, refueling or recharging time, cost per 100 km of driving, and maintenance and insurance expenses (Rudolph, 2016). INE is measured by the number of charging or gas stations within a predefined radius of the respondent’s residence (Sun et al., 2018). A discrete choice model is developed under the assumption that users choose between two vehicle types: ICEVs and EVs. Utility Uij that user i derives from selecting vehicle type j consists of three components, as specified in Equation (1). Xj represents the explanatory variables, including vehicle attributes (purchase price, driving range, refueling/- recharging time, cost per 100 km of driving, and insurance and maintenance costs), INE (measured by the number of charging stations within a specified radius), and PrCP installation status (represented by a binary variable and equal to 1 if a PrCP is installed and 0 otherwise). Υi denotes individual-level demographic characteristics, including gender, annual income, and age (Qian et al., 2019). ε ij is the error term. Uij =βʹXj+ α ʹΥi+ ε ij (1) To ensure the validity of the questionnaire design, we conducted a one-week pilot survey in December 2024, combining in-person visits to Table 1 Overview of literature on EV adoption. Methodology Reference Main content Summary Bass model Shi et al. (2022) A Bass model incorporating vehicle scrappage and technological competition to analyze EV diffusion in Shanghai, China. Although the Bass model offers advantages such as relative simplicity and low data requirements, its applicability is constrained by limited flexibility in capturing the influence of external environmental factors.X. Yang et al. (2025) Integration of uncertainty theory with the Bass model to analyze EV diffusion under the influence of uncertain events. Fan et al. (2025) A Bass model incorporating the green premium of EVs. System dynamics (SD) Y. Li et al. (2023) A multi-agent interaction SD model to analyze the impacts of different subsidy schemes on EV diffusion and carbon emissions. SD can capture interactions among heterogeneous agents and allows for a simulation of feedback in complex systems. However, it faces limitations in explicitly representing system-level phenomena that arise from individual responses to external factors. Shen et al. (2021) An SD model considering dynamic technological maturity evolution and incentive phase-out. Zhu et al. (2024) An SD model incorporating multi-agent interactions based on INE to compare the effectiveness of different policies in promoting EV diffusion. Kong et al. (2020) An SD model of EV diffusion incorporating interactions among the government, firms, and consumers. Kim et al. (2021) An SD model of EV diffusion incorporating policy incentives and environmental benefits. Agent-based modeling (ABM) Zhang et al. (2011) An ABM exploring the impact of exogenous factors on EV adoption in the context of multi-agent interactions between manufacturers, consumers, and the government. ABM can capture complex interactions among heterogeneous agents and between agents and their environment at a macro level. Eppstein et al. (2011) An ABM investigating the nonlinear impact of social influence and media exposure on EV market penetration. Shafiei et al. (2012) An ABM forecasting EV adoption through perceived utility and social influence within a choice-based diffusion framework. Xu and Bi (2024) An ABM considering the impact of word-of-mouth effects and social networks on consumers’ decisions to purchase EVs. Sun et al. (2018) An INE-based ABM evaluating the influence of spatial EVCI deployment policies on EV diffusion. Y. Wang et al. (2023) An INE-based ABM evaluating the impact of the government and operator EVCI deployment on EV diffusion. Luo et al. (2023) An ABM modeling the role of policy and external factors in overcoming the EV diffusion bottleneck.  Zhuge et al. (2021) A GIS-based ABM evaluating the impacts of EV price, range, and EVCI on EV diffusion across different regions in Beijing, China.  L. Zhu et al. Transport Policy 175 (2026) 103876 4 4S stores across multiple districts of Beijing with online data collection. Feedback from respondents regarding the clarity and appropriateness of the questionnaire was collected during the survey and subsequently used to refine its content. During the formal survey, questionnaires were collected from April 2 through April 30, 2025. The regional distribution of the responses is reported in Appendix A, Table A3, while descriptive statistics of the survey data are presented in Tables A1 and A2. In total, 549 responses were collected through the Wenjuanxing platform, 7 of which 533 were deemed valid after data trimming. The cleaned data was analyzed using Stata 16.0, with a mixed logit model to estimate consumer preferences for the relevant factors. Estimated coefficients were obtained through the maximum likelihood method, and the results are summarized in Table 3. As shown in Table 3, several vehicle attributes, including vehicle price, driving range, refueling/recharging time, cost per 100 km of driving, and maintenance and insurance cost, are found to be statistically significant. Particularly, the estimated coefficients for vehicle price, refueling/recharging time, cost per 100 km of driving, and maintenance & insurance cost are negative, suggesting that higher values of these attributes reduce consumer utility, which is consistent with theoretical expectations and aligns with prior research (Qian et al., 2019). With respect to the charging environment, both INE and PrCP installation status exhibit statistically significant coefficients, each contributing positively to consumer utility. Notably, the impact of PrCP installation is much greater than that of having one PCS within a 5 km radius. This finding supports the conclusions of Qian et al. (2019) and Helveston et al. (2015), who argued that the availability of private charging infrastructure has a stronger influence on EV adoption compared to the convenience of public charging infrastructure. Regarding demographic characteristics, only the income categories of ‘150,000–200,000 RMB’ and ‘≥300,000 RMB’ exhibit statistically significant effects. The magnitude and direction of these estimated coefficients indicate that higher income levels are associated with a greater likelihood of EV purchase, which is consistent with the findings of Sun et al. (2018a,b). By contrast, the coefficients for gender and age turn out to be not statistically significant, indicating that these characteristics have little influence on EV purchase decisions. 3.2.2. Utility function In the consumer sub-model, a mixed logit regression method based on survey data is used to construct a utility evaluation framework that incorporates vehicle attributes and environmental factors. As shown in Fig. 1, the framework consists of four components: (1) changes in vehicle lifespan over time (within the blue box), (2) comparison of vehicle attributes (within the red box), (3) evaluation of the EV charging environment (within the yellow box), and (4) calculation of EV purchase probability (within the green box). The blue box represents consumers who already own a vehicle (either EV or ICEV). For these consumers, vehicle lifespan decreases with each period, and once it reaches zero, they are reclassified as non-owners. The red box denotes vehicle attribute comparison, in which EV driving range and price are first evaluated. Consumers proceed to consider the charging environment only if both attributes meet expectations; if either falls short, they shift to purchasing an ICEV. Given that ICEVs are not subject to range anxiety, it is assumed that consumers will purchase an ICEV if its price falls below a Table 3 Maximum likelihood estimation results of the mixed logit regression. Variables Coeff. Std. error Pvalues Vehicle price (10,000 RMB) −0.060*** 0.011 0.000 Driving range (100 km) 0.004*** 0.0004 0.000 Refueling/Recharging time (minutes) −0.009*** 0.001 0.000 Cost per 100 km (RMB) −0.026*** 0.005 0.000 Maintenance and insurance cost (10,000 RMB) −0.540*** 0.063 0.000 Number of gas or charging stations (within 5 km) 0.018*** 0.005 0.001 PrCP installation (=1 if PrCP installed) 0.504*** 0.065 0.000 Income (100,000–150,000 RMB) 0.193 0.158 0.221 Income (150,000–200,000 RMB) 0.305*0.167 0.066 Income (200,000–300,000 RMB) 0.182 0.170 0.285 Income (≥300,000 RMB) 0.489** 0.210 0.020 Gender (Male) 0.031 0.103 0.767 Age (25–34) 0.060 0.147 0.682 Age (35–44) 0.054 0.161 0.739 Age (45–54) 0.309 0.266 0.245 Age (≥55) 0.219 0.471 0.642 Constant 0.357 0.249 0.153 Log likelihood − 2780.2399   Notes: *p <0.1; **p <0.05; ***p <0.01. The number of observations is 533. The omitted reference categories are ‘Income (50,000–100,000 RMB)’ for income level, ‘Female’ for gender, and ‘Age (18–24)’ for age, due to multicollinearity. Table 2 ABM-based studies on EV diffusion. Reference Consumer heterogeneity PCP PrCP PCS subsidy PCS subsidy phaseout PCS spatial allocation Zhang et al. (2011) ✓     Eppstein et al. (2011) ✓    ✓ Shafiei et al. (2012) ✓     Xu and Bi (2024) ✓    ✓ Sun et al. (2018) ✓ ✓    ✓ Y. Wang et al. (2023) ✓ ✓ ✓   ✓ Luo et al. (2023) ✓ ✓ ✓  Zhuge et al. (2021) ✓ ✓ ✓   ✓ Zhuge et al. (2021) ✓ ✓    ✓ Wolbertus et al. (2021) ✓ ✓    ✓ Huang et al. (2021) ✓ ✓ ✓  Pagani et al. (2019) ✓ ✓ ✓   ✓ Silvia and Krause (2016) ✓ ✓    ✓ This study ✓ ✓ ✓ ✓ ✓ ✓ 7 This platform specializes in professional questionnaire distribution and offers access to an extensive sample pool of approximately 6.2 million individuals, encompassing a wide spectrum of income levels, age groups, occupations, and other demographic characteristics. L. Zhu et al. Transport Policy 175 (2026) 103876 5 preset acceptable threshold; otherwise, they exit the market without making a purchase. The yellow box corresponds to the assessment of the charging environment. When an EV’s driving range and sale price meet expectations, consumers first check whether they already possess a PrCP or have permission to install one. If either condition is satisfied, the model calculates the probability of purchasing an EV with access to a PrCP. 8 If not, it is then assessed whether a PCS exists within 5 km 9 ; if available, the probability of purchasing an EV without PrCP (but with access to PCS) is calculated. If no PCS is available, consumers instead consider an ICEV and evaluate its price. Finally, the green box refers to the calculation of EV purchase probability. After this probability is determined, a random number is drawn from a uniform distribution [0,1] to simulate the consumer’s purchasing decision mechanism (Zhu and Ma, 2025): If the calculated probability is greater than or equal to the random number, the consumer purchases an EV; otherwise, they consider an ICEV and evaluate whether its price meets expectations. To account for EV consumer heterogeneity, we build on Wang et al. (2023) by analyzing the correlations between annual income and key parameters such as annual driving mileage and the proportion of time spent charging at PCSs. Based on the direction and strength of these correlations, we differentiate parameter values across income groups. Furthermore, drawing on the survey data, we establish income distribution profiles for urban and suburban populations. Details of the income distribution and correlation analysis are provided in Appendix A (Table A2 and Table A4, respectively). Based on the previously introduced discrete choice model, the utility consumer i derives from choosing vehicle type j at time t, which is Ui,j,t, can be specified in Equation (2): Ui,j,t=β1*(Pj,t−Spur,t)+β2*Cdri,i,j+β3*Cother,j+β4*Rj,t+β5*Tj+β6* Ni,j,t+β7*Bi+ ε i,j,t (2) Where Pj,t is the price of vehicle type j at period t, and Spur,t denotes the EV purchase subsidy at period t. Cdri,i,j represents the cost per 100 km of driving for consumer i using vehicle type j, while Cother,j is the maintenance and insurance cost of vehicle type j. Rj,t denotes the driving range of vehicle type j at period t, and Tj is the time required to fully refuel or recharge vehicle type j. Ni,j,t refers to the number of charging or gas stations within a 5-km radius of consumer i with vehicle type j during period t. Bi is a binary indicator equal to 1 if consumer i has access to a PrCP, and 0 otherwise. β is a vector of estimated coefficients, and ε i,j,t is the error term. It is worth noting that the influence of demographic attributes on consumer purchase decisions is omitted from the model specification in this paper. While this simplification could be seen as problematic, as a p-value above 0.1 does not necessarily imply a lack of effect (Scorrano and Danielis, 2025). We proceed with this assumption for the sake of model parsimony. Accordingly, the probability user i chooses an EV during period t, denoted as Pri,e,t, is formulated as shown in Equation (3): Pri,e,t=e(Ui,e,t− ε i,e,t) e(Ui,e,t− ε i,e,t)+e(Ui,f,t− ε i,f,t)(3) 3.2.2.1. Cost of 100-km driving. This paper assumes that EV users fall into two categories: those who rely exclusively on PCS and those who have access to PrCPs (although the latter may also use public charging facilities). The charging capacities of public and private stations are denoted as Kpub and Kpri, respectively. For users who rely solely on PCSs, the per-100-km driving cost, Cdri,i,e,pub, is determined by three factors: the composite electricity price at PCS Pelc, the service fee per unit (kWh) of electricity Pser, and the parking fee incurred per unit (kWh) of electricity charged while charging Ppark. 10 For consumers with PrCPs, the overall charging cost, Cdri,i,e,pri, is calculated by combining the costs of PCS and PrCP usage. This total cost depends on the share of charging time that individual i spends using PCP, represented by yi. The specific cost is calculated using Equation (4) for users who rely only on PCSs, and Equation (5) for those with access to PrCPs. Cdri,i,e,pub =he*(Pelc +Pser +Ppark)(4) Cdri,i,e,pri =he*Kpub*(Pelc +Pser +Ppark)*yi+Kpri*Ppri*(1−yi) Kpub*yi+Kpri*(1−yi)(5) Where yi denotes the percentage of total charging time that consumers with PrCPs spend using PCSs, 11 and he represents the electricity consumption per 100 km of driving for an EV. Ppri is the charging price of PrCPs and is assumed to be time-invariant. 12 The composite electricity price at PCS, Pelc, is calculated by multiplying the share of charging volume in each time-of-use period by the corresponding hourly electricity rate 13 and then summing across all time periods, as shown in Equation (6). Pelc =∑ 23 s=0 σ s*Pelc,s(6) Where σ s represents the share of charging volume in hour s, and Pelc,s is the corresponding electricity pricing rate. For ICEV consumers, the cost per 100 km of driving, Cdri,i,f, is calculated as the product of the fuel price Pg and the fuel consumption per 100 km of driving hf, as shown in Equation (7). In this case, Bi is set to 0, indicating that access to PrCPs will not affect the utility from using ICEVs. Cdri,i,f=hf*Pg(7) 3.2.2.2. Maintenance and insurance costs. The total maintenance and 8 In the Chinese market, most automobile manufacturers provide EV consumers free PrCPs and installation services, making it relatively easy for the consumers to access PrCPs. But in fact, the installation of a PrCP requires both a dedicated parking space and approval from the property management company. But according to a report by the China Consumer Association, the main reasons EV users do not install PrCP are the absence of a dedicated parking space or property management company’s refusal, which is consistent with the findings of EVCIPA (2020). Also, although some consumers do not install PrCPs even the conditions above are met, such case are not the main reason. Therefore, this paper makes an assumption and simplification: once installation conditions are satisfied, consumers are assumed to install PrCPs. Source: htt ps://www.cca.org.cn/Detail?catalogId=475800366178373&contentType=art icle&contentId=521575306829893. 9 According to the investigation by the China Consumers Association, 96 % of EV users live within 5 km of PCSs. Based on this, this study assumes that EV consumers would consider the presence of a PCS within 5 km when making purchase decisions. Source: https://www.cca.org.cn/Detail?catalogId=4758 00366178373&contentType=article&contentId=521575306829893. 10 For computationally simplicity, we assume a constant parking fee of approximately 4 RMB per hour. Given that the capacity of a public charging pile is set at 30 kW, the corresponding parking fee incurred per unit (kWh) of electricity charged is calculated as 4/30 =0.13 RMB per kWh. 11 In the main text, yi refers to the proportion of total charging time consumer i spent using PCS (i.e., Time of charging at PCS/(Time of charging at PCS + Time of charging at PrCP)). However, in the questionnaire, yi was defined in opposite way for ease of respondent understanding, namely, as the share of charging time spent using PrCP (i.e., Time of charging at PrCP/(time of charging at PCS +Time of charging at PrCP)). Accordingly, the ‘share of charging time at PrCP’ reported in Table A1 does not appear in the main text. 12 The government has mandated regulations on electricity pricing for PrCPs (NDRC, 2014). Under this policy, a unified residential electricity rate is applied, which means that Ppri remains constant over time. Source: https://zfxxgk.ndrc. gov.cn/web/iteminfo.jsp?id=19564. 13 The time-of-use charging volume shares are from: https://www.evcipa.org. cn/newsinfo/8137317.html. L. Zhu et al. Transport Policy 175 (2026) 103876 6 insurance cost for vehicle type j is calculated by multiplying the estimated maintenance and insurance costs per period by the expected number of periods that the vehicle will be in use, as shown in Equation (8). Specifically, Cins,j denotes the insurance cost, Cmai,j represents the maintenance cost, and η j refers to the estimated service life (in periods) of vehicle type j. Cother,j=(Cins,j+Cmai,j)*nj(8) 3.2.2.3. EV pricing. With respect to EV pricing, this paper follows the approach of Hu´ etink et al. (2010) for hydrogen vehicle pricing, assuming that EV prices gradually decrease as market stock increases, as shown in Equation (9). Pe,t=Pe,0*(Qe,0 Qe,t−1) ω (9) Where Pe,0 is the initial price of EVs, and Qe,0 is the initial stock of EVs. Qe,t−1 represents the stock of EVs in period t-1, and ω denotes the learning rate, with higher values of ω indicating a faster decline in EV prices. 3.2.2.4. EV driving range. To model the evolution of EV driving range over time, this paper adopts an S-shaped technology lifecycle curve to represent technological maturity, which in turn affects driving range. Since EV driving range is closely tied to advances in battery and related technologies (Z. Liu et al., 2023), and given that the number of patents can serve as a proxy for technological progress, we follow the approach of Shen et al. (2021) by linking technological maturity to driving range. In their work, technology diffusion theory is applied to fit an S-curve to the growth of EV-related patents, with the maximum number of patents estimated. Technological maturity in each period is then calculated as the ratio of the cumulative number of patents to this estimated maximum. While the same method is adopted in this study, adjustments are made to account for differences in the simulation period: in Shen et al. (2021), each simulation step represents one month, whereas in this study, each step corresponds to three months. EV driving range and technological maturity are calculated using Equations (10) and (11), respectively. Re,t=Re,0 (1−Techt+Tech0)2(10) Techt=1 1+e− τ *[(3*t−2)+667−θ](11) Where Techt denotes the technological maturity of EVs at time t, and Tech0 is the initial level of technological maturity. τ represents the growth rate of the S-curve, while θ indicates its inflection point—the Fig. 1. The decision-making process of consumers. L. Zhu et al. Transport Policy 175 (2026) 103876 7 time at which the growth rate begins to slow. To avoid unrealistic or uncontrolled growth in EV driving range over time, an upper bound is also set for the maximum driving range. 3.2.2.5. Driver population. China’s automotive market has experienced steady growth in both vehicle ownership and the number of licensed drivers(Kong, 2025). Although the annual growth rate of new drivers has gradually declined, the market is expected to remain in an expansion phase until approximately 2028. From a long-term perspective, the Chinese automotive market is projected to sustain growth through 2050 (Hao et al., 2011). In this study, the simulation period covers the real-world timeframe from 2019 to 2031. To simplify the model, and based on the observed annual trend, we assume that the growth rate of licensed driver quantity each period is time-invariant and represented by γd. The population size of the simulated space in period t, denoted as Mt, is calculated using Equation (12), while the number of newly added potential consumers in each period, ΔMt, is determined using Equation (13). Mt=Mt−1*(1+γd)(12) ΔMt=Mt−Mt−1(13) 3.2.2.6. PrCP installation eligibility. The installation of PrCP requires several conditions, most notably sufficient load capacity and access to dedicated parking spaces. To promote EV adoption, the Chinese government issued a policy in 2015 mandating that newly built neighborhoods include provisions for charging infrastructure (General Office of the State Council of the P.R.C, 2015). 14 However, many communities built before 2015 were not equipped with sufficient electrical infrastructure to support PrCPs, and the majority of parking spaces in these neighborhoods do not meet the technical requirements for PrCP installation, particularly in older communities built before 2000. These aging communities account for nearly 40 % of residential areas in 20 major cities across China. 15 Based on the ongoing renovations of older communities, this study assumes that the proportion of the population meeting PrCP installation conditions will gradually increase, reflecting continuous improvements in residential infrastructure. Nevertheless, due to the persistent shortage of dedicated parking spaces in China, 16 our model constrains this proportion from even reach 100 %. To simplify the modeling process, it is further assumed that, in each period, a portion of the population without PrCP installation conditions, denoted as Mt,np, acquires the necessary conditions at a fixed rate, γn. This specification is formulated in Equation (14): Mt+1,np =Mt,np*(1−γn)(14) Where γn represents the conversion rate of PrCP installation eligibility in each period. 3.3. EVCI operator sub-model In this study, the EVCI operator evaluates both operational performance and market charging demand to determine whether to deploy new PCSs. The operator’s revenue primarily comes from charging fees and government subsidies, while costs consist of construction, land rent, maintenance, and electricity procurement (Q. Zhang et al., 2018). ROI serves as a comprehensive metric for evaluating the balance between operating profits and costs (Schroeder and Traber, 2012). Although charging demand is directly linked to the number of EV users, it is also shaped by a complex interplay of charging prices, electricity procurement costs, operating expenses, and construction costs (Q. Zhang et al., 2018). Therefore, analyzing charging demand requires a dual focus on both the scale of EV users and the characteristics of the existing public charging infrastructure. The utilization rate serves as an effective indicator for the relationship between the number of consumers and the number of PCSs across different regions. By analyzing utilization rates, the operator can better identify regional charging demand patterns and adjust charging station deployment strategies accordingly. As the primary objective of this paper is not to examine competitive dynamics among operators, the model is simplified by assuming a single operator in the market. The decision-making process of this operator is illustrated in Fig. 2. If the operating profit in a given period is positive, the operator undertakes cost improvement measures for PCSs. Based on the updated costs and current profit, ROI is recalculated. When ROI exceeds a predefined threshold ROImin, the operator deploys additional stations. Deployment decisions are guided by a comparison of regional utilization rates with the overall average, with new PCSs allocated in areas showing above-average demand. 3.3.1. Operating profit The operating profit of PCSs mainly depends on service fees, parking fees, operating costs, OSs, and advertising revenue. Revenue from parking and service is directly tied to consumer charging demand— higher demand leads to higher operating profit. At the same time, as the number of PCSs increases, operating costs rise correspondingly, while advertising revenue also grows due to the expansion of infrastructure. 17 The operator’s operating profit in period t, π t, is given by Equation (15). π t=∑ n i=1 Ec i*(Pser +Ppark)−∑ n g=1 Cg op +Sop,t+Aads*∑ m h=1 Nh,e,t(15) Where Ec i represents the electricity demand of user i during each period and is calculated in Equation (16). Aads denotes the advertising revenue generated by each PCS per period, and Cg op is the operating cost of PCS g per period. Nh,e,t indicates the number of the stations in region h during period t, while Sop,t represents the government OSs provided in period t. Ec i=Di 100*he(16) Cg op =γc*Cg con (17) As shown in Equation (16), the electricity demand of user i per period, denoted as Ec i, is determined by the user’s driving distance in that period Di and the EV’s energy consumption per 100 km he. The operating cost of PCS g in each period, Cg op, is calculated based on its construction cost Cg con and the ratio of operating to construction costs γc, as expressed in Equation (17). Especially, Ec i is time-invariant, and the construction cost of PCS g, Cg con, equals the unit construction cost in period t, Ccon,t, when the station was built. 3.3.2. Research and development cost The growth in the number of EVs increases charging demand, which requires the operator to expand the number of PCSs to accommodate more users. However, the construction of new stations directly affects the operator’s ROI. To address this, the operator allocates a portion of its funds to research and development (R&D), with a particular focus on reducing construction costs of PCSs. Following the description of operator R&D investment in Sun et al. (2016), this study constructs a 14 This report is accessed at: https://www.gov.cn/zhengce/content/2015-10 /09/content_10214.htm. 15 The information is from the Beike Research Institute report: https://rese arch.ke.com/121/ArticleDetail?id=274. 16 Relevant information is accessed from: https://capital.people.com. cn/n1/2020/0708/c405954-31775313.html. 17 The installation of PCSs does not directly affect the operator’s profit. This is because, based on our field investigation, most PCSs are provided by automobile manufacturers at the time of vehicle purchase, with both equipment and installation services offered to users free of charge. L. Zhu et al. Transport Policy 175 (2026) 103876 8 simplified cost-improvement R&D model. In this model, the operator allocates a portion of its operating profit in period t to R&D investment, denoted as RDt, as defined in Equation (18). RDt=λt* π t(18) Where λt represent the proportion of profit allocated to R&D investment in period t. The R&D investment in period t, denoted as RDt, determines the construction cost Ccon,t for that period. However, construction costs cannot decrease indefinitely and are subject to a lower bound, denoted as Cmin con . The reduction in construction cost for each period, ΔCcon,t, is calculated using Equation (19). ΔCcon,t= μ *RDt*v*(Ccon,t−1−Cmin con )(19) Where μ represents the proportion of R&D investment allocated specifically to cost reduction, and v reflects the effectiveness of R&D in lowering construction costs. After calculating the cost reduction with Equation (19), the operator updates the construction cost according to Equation (20). Ccon,t=Ccon,t−1−ΔCcon,t(20) 3.3.3. Public charging stations deployment The deployment of PCSs by EVCI operators is primarily driven by two factors: RIO and charging demand. The former is evaluated against a minimum acceptable threshold, ROImin, which must be exceeded for new investments to proceed. The latter is assessed through regional utilization rates. 18 Following Luo et al. (2023), ROI per period, ROIt, is calculated as the ratio of operating profit to initial construction cost, as expressed in Equation (21). ROIt= π t Ccon,t− (z*K1*scon)(21) Where z denotes the number of PCPs, K1 is the charging capacity of a single station, scon represents the CS for each station, and π t is the average operating profit per station, which is calculated based on Equation (22). For computational simplicity, it is assumed that all charging stations have identical charging capacity and an equal number of charging piles, with variations across stations limited to construction and operation costs. π t= π t ∑ m h=1 Nh,e,t (22) When the ROI exceeds the threshold ROImin, the operator proceeds with the construction of a new PCS. The overall average utilization rate of all PCSs in the simulated space, denoted as rt, is calculated using Equation (23). rt=∑m h=1∑n i=1Ec i,h ∑ m h=1(Nh,e,t*Emax) (23) Where Emax denotes the maximum charging volume that a single station can offer per period. Once the overall average utilization rate is obtained, the regional utilization rate for region h in period t, rht , is determined through Equation (24). rh,t=∑n i=1Ec i,h Nh,e,t*Emax (24) The average utilization rate of PCSs effectively reflects EV charging demand within a region. A higher utilization rate indicates stronger demand and signals an undersupply of charging infrastructure, thereby indicating the need for additional PCS deployment. The operator identifies regions with utilization rates higher than the average by comparing the regional average utilization rate rh,t with the overall average rt, and deploys new PCSs in these regions. 19 However, the expansion of PCSs is inherently constrained, as additional construction and operation lead to higher costs. Given this fact, it is assumed that the operator deploys additional PCSs only in regions with above-average demand, with the number of newly deployed stations calibrated to ensure that regional utilization does not exceed the overall average rate, while no new PCSs are deployed in regions without excess demand. To ensure that regional utilization rates do not surpass the overall average, operators apply Equation (25) to calculate the number of additional stations, ΔNh,e,t, needed in each region. The results are rounded up to the nearest integer to avoid fractional values in ΔNh,e,t. Since the primary objective of this study is to examine the impact of public policies on EV diffusion rather than the market effects of specific PCS locations, the model assumes that newly added PCSs are randomly distributed within targeted regions. Furthermore, charging demand is assumed to be evenly distributed across all PCSs in each region. ΔNh,e,t=⌈∑n i=1Ec i,h rt*Emax −Nh,e,t⌉(25) 3.4. Government sub-model The government provides subsidies to both EV consumers and EVCI operator, which aligns with real-world policy practices. Consumers receive EV purchase subsidies, while operator obtains two forms of financial support: OSs and CSs for charging infrastructure. 3.4.1. EV purchase subsidy The simulation begins in 2019, coinciding with the period when EV purchase subsidies significantly boosted EV adoption before being Fig. 2. The decision-making process of the EVCI operator. 18 Relevant information is available through: https://research.gszq.com/rese arch/report?rid=8ae505846943c3450169537578e22c1a. 19 This study assumes that regional utilization rate decreases as additional charging stations are built in the region. L. Zhu et al. Transport Policy 175 (2026) 103876 9 short run. This phenomenon also sheds light on why many charging infrastructure providers in China experienced financial difficulties during the early stages of EV market expansion. 33 Faced with persistent operational losses, several charging infrastructure providers suspend services, resulting in the proliferation of so-called “zombie chargers”, which refer to the inactive or non-functional PCSs that hinder EV diffusion. Introducing OSs during this stage can help alleviate the issue by incentivizing continued operation. Therefore, a policy shift from CSs to OSs emerges as a necessary measure to promote the sustainable development of the charging infrastructure industry. 5.3. Impact analysis of subsidy phase-out policies 5.3.1. Setting for subsidy phase-out modes Given the substantial fiscal burden associated with charging infrastructure subsidies, their gradual phase-out has become an inevitable policy trend. Indeed, many Chinese cities have already announced specific withdrawal plans. For example, Shanghai has outlined a twostage reduction of OSs beginning in 2025, with a full termination scheduled by 2028. Hainan Province has retained CSs exclusively for Fig. 7. Temporal evolution of EV penetration rate by region under baseline scenario. Fig. 8. Temporal evolution of EV purchase price and driving range. 33 More details on this can be assessed at: https://report.iresearch.cn/report /202006/4456.shtml. L. Zhu et al. Transport Policy 175 (2026) 103876 16 rural areas, while lowering the subsidy rate from 10 %–15 % to 5 %–10 %. Similarly, Chongqing has announced that, beginning in 2026, its current OS of 0.1 RMB/kWh will be reduced by an additional 20 %, placing additional pressure on EVCI operators to enhance operational efficiency. In the sensitivity analysis of EVCI subsidies, we find that their impact on the diffusion of EVs and EVCI is most pronounced in the early and middle stages of the simulation (before period 16). Building on this observation, and in line with real-world policy trends toward subsidy phase-outs, we design two alternative phase-out scenarios that begin in the middle stage (period 17), as summarized in Table 6. Mode 1 (gradual phase-out) assumes that subsidies are reduced by 25 % from their initial level starting in period 17, followed by additional 25 % reductions at regular intervals until a full withdrawal is achieved by period 37. Mode 2 (rapid phase-out) assumes that subsidies are immediately halved in period 17 and fully eliminated by period 33. To further investigate the effects of different initial subsidy levels on EV adoption, EVCI deployment, and government expenditure, this paper adopts a policy-mix approach. Specifically, we combine multiple initial levels of CSs and OSs with the two previously defined phase-out strategies. By pairing the five CS levels and five OS levels listed in Table 5 with the two phase-out modes, a total of 50 distinct policy combinations is yielded. For instance, Policy Combination 1 (0.05, 150) represents a scenario with an initial OS level of 0.05 RMB/kWh and a CS level of 150 RMB/kW, followed by a gradual phase-out. Similarly, Policy Combination 2 (0.25, 250) represents another scenario with an initial OS level of 0.25 Fig. 9. Temporal evolution of vehicle market stock by type and region. Fig. 10. Temporal evolution of EVCI scale and density by type and region. L. Zhu et al. Transport Policy 175 (2026) 103876 17 RMB/kWh and a CS level of 250 RMB/kW, under a gradual phase-out strategy as well. The following section provides a comparative analysis of outcomes across all 50 policy combinations. 5.3.2. Comparison of policy combinations Fig. 16 presents the cumulative subsidy expenditure and EV stock levels at the end of the simulation period for each of the 50 policy combinations. Overall, the results indicate that, regardless of the phaseout mode, higher initial subsidy levels always lead to greater EV ownership. However, when comparing the two phase-out modes under identical initial subsidy configurations, the gradual phase-out mode results in significantly higher cumulative government spending, on average 22.30 % more than the rapid phase-out mode. By contrast, the corresponding increase in EV stocks is negligible, with an average difference of only 0.028 %. To evaluate the efficiency of different subsidy phase-out strategies, which is defined as achieving higher EV adoption while minimizing Fig. 11. Temporal evolution of EVCI operator profitability and OS-to-Profit ratio by region. Fig. 12. Temporal evolution of cumulative subsidies by type and region. Table 5 Specification of OS and CS levels. Subsidy type Level Unit Very low Low Baseline High Very high OS 0.05 0.10 0.15 0.20 0.25 RMB/ kWh CS 50 100 150 200 250 RMB/kW L. Zhu et al. Transport Policy 175 (2026) 103876 18 government expenditure, we conducted an additional analysis of the 50 policy combinations. The baseline subsidy level is defined as an OS of 0.15 RMB/kWh and a CS of 150 RMB/kW. Initial subsidy levels below this baseline are categorized as low subsidies, which include OS values of 0.05 RMB/kWh and 0.10 RMB/kWh, and CS values of 50 RMB/kW and 100 RMB/kW. Conversely, OS values of 0.20 RMB/kWh and 0.25 Fig. 13. Temporal evolution of PCP quantity and proportional difference by OS level. Fig. 14. Temporal evolution of EV ownership and proportional differences by OS level. Fig. 15. Temporal evolution of PCP quantity and operator profitability percentage differences by CS level. L. Zhu et al. Transport Policy 175 (2026) 103876 19 RMB/kWh, along with CS values of 200 RMB/kW and 250 RMB/kW, are classified as high subsidies. Under both phase-out modes, we focus on two distinct groups of policy combinations: (1) low OS with high CS, and (2) high OS with low CS. For ease of comparison, these combinations are plotted as connected points in Fig. 16. As shown in the figure, policy combinations with high initial OS and low initial CS cluster in the upperleft region of the graph, whereas those with low initial OS and high initial CS appear in the lower-right region. The comparative analysis reveals that, on average, combinations with high initial OC and low initial CS reduce cumulative government expenditure by 32.15 %, while simultaneously achieving 0.54 % higher EV ownership compared to the combinations with low initial OS and high initial CS. Taken together, these results suggest that policy combinations involving a rapid phaseout strategy with high initial OS and low initial CS are the most costeffective. 5.3.3. Comparison of phase-out modes Fig. 17 presents the EV stocks and EVCI quantities corresponding to each of the 50 subsidy phase-out policy combinations. For ease of comparison, two reference lines are added to indicate the levels of EV ownership and EVCI deployment under the baseline non-phase-out scenario (where baseline subsidy levels are maintained throughout the simulation). These reference lines divide the figure into four subregions. Policy combinations falling in the upper-right region result in increases in both EV adoption and EVCI deployment compared to the baseline. This indicates that, provided initial subsidy levels are relatively high, implementing a phase-out policy does not hinder the diffusion of EVs or charging infrastructure. However, it is also noteworthy that the gap in cumulative government spending between phase-out and nonphase-out policies is substantial. Compared with the baseline nonphase-out scenario, subsidy phase-out combinations reduce total government subsidy expenditure by 83.94 %–97.19 %. In other words, the upper-right region of Fig. 17 highlights a set of policy options that not only sustain or enhance EV adoption and EVCI deployment, but also dramatically alleviate fiscal burdens. Specifically, we compare EV ownership and cumulative subsidy expenditure under the policy combination (0.15, 150) across both gradual and rapid phase-out scenarios, using the baseline non-phase-out scenario as a benchmark. Compared to the baseline, cumulative subsidy expenditures decrease by 90.07 % under the gradual phase-out and by 91.88 % under the rapid phase-out. Meanwhile, changes in EV ownership remain within 0.05 %, a difference that is practically negligible. These findings are consistent with our results in Section 5.1, which indicate that the impact of subsidies on EV diffusion is concentrated in the early and middle stages, with only limited effects observed in the later stages. If current subsidy levels were to be maintained throughout the entire simulation period, the rapid increase in EVCI deployment during the mid-to-late stages would result in substantially higher expenditures on both OSs and CSs. By contrast, the phase-out policies proposed in this study substantially reduce fiscal burdens without compromising the long-term diffusion of EVs. 5.4. Impact analysis of external factors 5.4.1. Gasoline and electricity prices Fig. 18 illustrates the sensitivity of EV ownership to changes in gasoline and electricity prices. The results indicate that both factors significantly influence EV adoption decisions. As shown in Fig. 18(a), EV ownership is positively correlated with gasoline prices, while Fig. 18(b) illustrates a negative correlation between electricity prices and EV adoption. These findings are consistent with previous research, such as Luo et al. (2023) and Shafiei et al. (2012). 5.4.2. Electricity consumption per hundred kilometers Fig. 19(a) shows the sensitivity analysis of EV ownership with respect to variations in electricity consumption per 100 km. Similar to electricity prices, EV ownership is negatively correlated with electricity consumption per 100 km, as both factors influence consumer purchase decisions through their effect on the driving cost of EVs. Meanwhile, as shown in Fig. 19(b), lower electricity consumption per 100 km is associated with a reduction in the number of PCSs. This occurs because improved energy efficiency reduces overall charging demand, thereby discouraging the further deployment of PCSs. In light of ongoing improvements in energy efficiency, the charging infrastructure industry should strategically shift its focus from simply expanding the number of facilities toward enhancing service quality. Such a transition would signal the emergence of a value-driven development phase, characterized by fewer but more efficiently operated charging facilities, rather than continued scale-oriented expansion. 6. Conclusions and implications 6.1. Conclusions This paper examines the effects of EVCI subsidies and their phase-out in China within an ABM framework and yields three main conclusions: (1) The Chinese EV market is projected to sustain rapid growth, with the penetration rate expected to reach 79.78 % by 2030. A notable trend is the accelerated diffusion of EVs in suburban areas, where EV ownership is anticipated to eventually surpass that of urban areas. In terms of charging infrastructure, PrCPs are expected to remain dominant, while government subsidies will continue to play a crucial role in supporting the expansion of public charging facilities. Although suburban areas initially lag behind urban areas in PCP deployment, they are expected to overtake urban levels as EV adoption grows, albeit at a lower spatial density. From a profitability perspective, suburban charging piles may incur financial losses in the early stages but are likely to yield higher long-term profits for operators than their urban counterparts. (2) EVCI subsidies are found to be most effective in promoting EV diffusion during the early and middle stages, with their influence diminishing in the mid-to-late stages. Specifically, our results indicate that raising the OC level from the current 0.15 RMB/ kWh to 0.25 RMB/kWh can enhance charging pile deployment by up to 20 % in the early and middle stages but yields only about a 2 % increase in the later stages. While CSs can facilitate early infrastructure expansion, they risk causing overinvestment and operational inefficiencies, which may undermine the long-term sustainability of the industry. If current subsidy levels were maintained without phase-out, annual government spending in 2030 would be approximately 4.14 times higher than in 2024. These findings underscore the importance of shifting from CSs to Table 6 EVCI subsidy phase-out modes. Period Gradual subsidy phase-out (Mode 1) Rapid subsidy phase-out (Mode 2) OS CS Max OS OS CS Max OS [1,16] Initial level 400 kW h/kW/ period Initial level 400 kW h/kW/ period [17,28] Initial level*75 % 300 kW h/kW/ period Initial level*50 % 200 kW h/kW/ period [29,32] Initial level*50 % 200 kW h/kW/ period Initial level*25 % 100 kW h/kW/ period [33,36] Initial level*25 % 100 kW h/kW/ period 0 0 0 [37,52] 0 0 0 0 0 0 L. Zhu et al. Transport Policy 175 (2026) 103876 20 OSs and adopting a phased reduction in subsidy policies in the later stages of market development. (3) To assess the effectiveness of subsidy phase-out strategies, this study constructs 50 policy combinations by varying initial CS and OS levels across two phase-out modes. The simulation results indicate that maintaining current subsidy standards while adopting a phase-out mode can reduce cumulative government spending by 91 % compared to a no-phase-out scenario, while resulting in only a marginal 0.05 % decline in EV ownership. A comparison of the 50 policy combinations further reveals that, under identical initial CS and OS levels, the rapid phase-out mode reduces total expenditure by 22.30 % compared to the gradual mode. Moreover, a policy combination with high initial OS and low initial CS lowers cumulative expenditure by 32.15 % compared to the opposite setup (a combination with low initial OS and high initial CS). In summary, these findings suggest that the most cost-effective policy mix consists of a rapid phase-out mode combined with high initial OS and low initial CS, as it significantly reduces fiscal expenditure without impeding EV diffusion. 6.2. Implications Our findings provide insights for policymakers and generate practical guidance for EVCI operators. Based on the simulation results, the following specific recommendations are proposed: (1) When designing subsidy policies, governments should account for the growing fiscal burden arising from rapid EV diffusion. In the early stages of market development, cities with adequate fiscal budgets and an urgent need to accelerate EV adoption may implement high levels of both CSs and OSs. Such a strategy Fig. 16. Cumulative subsidy expenditure and EV ownership across policy combinations. Fig. 17. Comparison of phase-out policy combinations and the baseline non-phase-out scenario. L. Zhu et al. Transport Policy 175 (2026) 103876 21 supports the rapid deployment of public charging infrastructure and helps operators overcome profitability challenges in the early stage. However, it is also essential to have strict supervision of CSs for the sake of efficiency in using public resources. Introducing capacity utilization assessments for operators can help prevent overinvestment in regions with limited charging demand. For cities facing tighter fiscal constraints, a strategy with relatively higher OS and lower CS, with a focus on enhancing operational performance, would be more appropriate. It is also recommended to link subsidy amounts to key performance indicators such as utilization rates and fault rates, which can facilitate the identification and elimination of underperforming charging stations. As the market develops, governments should adopt a dynamic subsidy phase-out mechanism that balances the reduction of subsidy expenditure and the promotion of EV diffusion. (2) To promote balanced EVCI development across urban and suburban areas, policymakers should account for regional heterogeneity and adopt targeted measures. In terms of subsidy policy design, greater financial support should be allocated to suburban EVCI to accelerate infrastructure deployment and enhance operator profitability during the early stages of market development. On the technical and planning aspects, the government can leverage the relative ease of PrCP installation in suburban areas by expanding charging pile capacity and streamlining installation requirements. Such efforts would help address issues such as limited grid capacity and cumbersome approval processes that currently constrain PrCP expansion. Through these measures, policymakers can foster a suburban EVCI development mode centered on PrCPs, with PCSs playing a complementary role, thereby fully leveraging the potential of PrCPs while overcoming key barriers to EV and EVCI adoption in suburban areas. (3) EVCI operators should take advantage of the current subsidy window by prioritizing charging pile deployment in highpotential suburban areas, leveraging high OS levels to offset financial losses in the early stages. At the same time, implementing dynamic pricing strategies can help enhance utilization rates of piles and mitigate the impact of future subsidy phaseouts. Beyond pricing, operators should explore innovative business models to reduce operational and maintenance costs in suburban areas. For example, given the widespread presence of PrCPs in suburban areas, operators could collaborate with homeowners to pilot shared charging schemes. Furthermore, considering the unique characteristics of suburban land ownership, a crowd-funded installation model, where homeowners provide space and EVCI operators contribute equipment and technical expertise, could be an effective solution. Fig. 18. EV ownership under varying gasoline and electricity prices. Fig. 19. EV ownership and PCP quantity under varying electricity consumption levels. L. Zhu et al. Transport Policy 175 (2026) 103876 22 6.3. Limitations and future work This study is subject to several limitations. First, it remains challenging to accurately model variations in EVCI capacity due to the complexity of influencing factors and the heterogeneous composition of EVCI with different capacity levels. As a result, EVCI capacity in this study is set to be static (exogenous) rather than dynamically evolving. Future research could apply alternative forecasting approaches to better capture the dynamic evolution of charging pile capacity. Second, the electricity consumption per 100 km of EV driving is also treated as an exogenous variable in this study because of the lack of reliable forecasting methods, and thus a sensitivity analysis was conducted in this study to partially address this limitation. In addition, this model treats EV charging time and ICEV attributes (such as energy consumption per 100 km of driving, driving range, and vehicle price) as static. This simplification may lead to either overestimation or underestimation of the long-term attractiveness of EVs. Future research could address this limitation by developing dynamic models for both EVs and ICEVs to improve accuracy and more closely reflect real-world technological progress and usage conditions. Third, automobile manufacturers are not incorporated into the model due to its complexity, as the primary focus of this study is on EVCI. For the sake of simplification, this paper also follows previous studies by considering only a single operator without accounting for competition among multiple operators. Future research could address these limitations by incorporating automobile manufacturers and introducing competitive mechanisms among multiple operators to more accurately reflect the evolution of the EV market in the real world. CRediT authorship contribution statement Lijing Zhu: Conceptualization, Methodology, Writing – original draft. Runze Li: Data curation, Methodology. Jingzhou Wang: Software, Validation. Haibo Chen: Data curation. Ondrej Havran: Writing – review & editing. Wen-Long Shang: Supervision, Writing – original draft. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements This work was supported by Ministry of Education in China (MOE) Project of Humanities and Social Sciences (No. 24YJA790105) and Beijing Natural Science Foundation, China (No. 9232003). Besides, this research was also partially supported by the ZEV-UP and ePowerMove projects co-funded by the European Union under Grant agreement ID: 101138721 and 101192753. Appendix A Table A1 Sample characteristics of survey respondents Variable Category Sample size Percentage Gender Female 305 57.2 % Male 228 42.8 % Age 18–24 78 14.6 % 25–34 244 45.8 % 35–44 162 30.4 % 45–54 37 6.9 % ≥55 12 2.3 % Income <100,000 RMB 73 13.7 % 100,000–150,000 RMB 154 28.9 % 150,000–200,000 RMB 131 24.6 % 200,000–300,000 RMB 109 20.4 % ≥300,000 RMB 66 12.4 % Place of residence Central Urban area 261 49.0 % Inner suburban area 195 36.6 % Outer suburban area 30 5.6 % Rural area 47 8.8 % EV ownership and PrCP installation status Yes 173 32.5 % EV only 111 20.8 % No 249 46.7 % PrCP charging time (PrCP +PCS)charging time Not applicable 360 67.5 % ≤20 % 14 2.6 % 21–40 % 54 10.1 % 41–60 % 51 9.6 % 61–80 % 35 6.6 % >80 19 3.6 % Vehicle purchase intention in the next year Yes 385 72.2 % No 148 27.8 % Annual driving mileage (unit: 10,000 km) <1 64 12.0 % 1–1.5 116 21.8 % 1.5–2 250 46.9 % 2–3 89 16.7 % ≥3 14 2.6 % L. Zhu et al. Transport Policy 175 (2026) 103876 23 Table A2 Income characteristics of respondents from different regions Variable Category Sample size Percentage Income (Urban area) <100,000 RMB 18 6.9 % 100,000–150,000 RMB 57 21.9 % 150,000–200,000 RMB 70 26.9 % 200,000–300,000 RMB 68 26.2 % ≥300,000 RMB 47 18.1 % Income (Suburban and rural areas) <100,000 RMB 52 19.0 % 100,000–150,000 RMB 98 35.9 % 150,000–200,000 RMB 63 23.1 % 200,000–300,000 RMB 41 15.0 % ≥300,000 RMB 19 7.0 % Table A3 Geographic Distribution of the Responses Province Sample size Province Sample size Beijing 29 Hubei 32 Tianjin 10 Hunan 15 Hebei 32 Guangdong 57 Shanxi 18 Guangxi 6 Inner Mongolia 7 Hainan 2 Liaoning 23 Chongqing 11 Jilin 7 Sichuan 19 Heilongjiang 8 Guizhou 16 Shanghai 23 Yunnan 9 Jiangsu 40 Shaanxi 13 Zhejiang 30 Gansu 8 Anhui 15 Qinghai 1 Fujian 25 Ningxia 2 Jiangxi 10 Xinjiang 2 Shandong 40 Henan 23 Table A4 Correlation matrix of demographic variables Income Annual driving mileage Minimum acceptable EV driving range Maximum acceptable vehicle purchase price Share of charging time at PrCP Income 1    Annual driving mileage 0.259*** 1    Minimum acceptable EV driving range 0.275*** 0.106*** 1   Maximum acceptable vehicle purchase price 0.614*** 0.242*** 0.329*** 1  Share of charging time at PrCP 0.254*** 0.176*** 0.112*** 0.306*** 1 Notes: *p <0.1; **p <0.05; ***p <0.01. Appendix B Table B1 Description of parameters and variables Category Symbol Description Symbol Description Parameters PfICEV purchase price Pj,0Initial purchase price of vehicle type j Rj,0Initial driving range of vehicle type j TjRefueling/recharging time for vehicle type j heElectricity consumption per 100 km of driving for an EV hfFuel consumption per 100 km for an ICEV Pelc Composite electricity price at a public charging station Ppark Parking fee while charging Pser Service fee per unit (kWh) of electricity charged Ppri Charging price of private charging piles Pgas Gasoline price Kpub Capacity of public charging station Kpri Capacity of private charging station σ sShare of charging volume in hour s njEstimated service life of vehicle type j Cins.jInsurance cost for vehicle type j per period Cmai,jMaintenance cost for vehicle type j per period θInflection point of the technological maturity curve τ Growth rate of the technological maturity curve γdGrowth rate of licensed driver quality per period γnConversion rate of private charging pile installation eligibility per period Aads Advertising income for a public charging station per period γcOperating-to-construction cost ratio per period Cmin con Minimum construction cost for a charging station (continued on next page) L. Zhu et al. Transport Policy 175 (2026) 103876 24 Table B1 (continued) Category Symbol Description Symbol Description zNumber of charging piles in a public charging station sop Operational subsidy rate per unit of electricity charged scon Construction subsidy for each public charging station sop,max Maximum subsidized charging volume per unit of charging pile capacity (kW) Emax Maximum charging volume for a station per period   Variables Ui,j,tUtility of user i from vehicle type j in period t Pe,tPurchase price of an EV in period t Qj,tMarket stock of vehicle type j in period t Re,tDriving range of an EV in period t spur,tEV purchase subsidy in period t Cdri,i,fCost per 100 km of driving for consumer i when using ICEV Cdri,i,e,pub Cost per 100 km of driving for consumer i without private charging piles when using EV Cdri,i,e,pri Cost per 100 km of driving for consumer i with private charging piles when using EV Ni,j,tNumber of charging or gas stations within a 5-km radius of consumer i with vehicle type j in period t BiPrivate charging piles access indicator (1 if yes, and 0 otherwise) yiTime of charging at PCS/(Time of charging at PCS +Time of charging at PrCP) Pri,j,tPurchase probability of vehicle type j in period t for consumer i TechtTechnological maturity of EVs in period t ΔMtNewly added licensed drivers in period t MtNumber of licensed drivers in period t Mt,np Number of consumers without private charging pile installation eligibility in period t π tOperating profit of EVCI operator in period t Ec iElectricity demand of user i per period Cg op Operating cost for public charging station g per period Nh,e,tNumber of public charging stations in region h in period t DiDriving distance of consumer i per period Cg con Construction cost of public charging station g RDtR&D investment in period t λtR&D investment share of profit in period t μ Share of R&D for cost reduction ν Effectiveness of R&D in lowering construction costs Ccon,tNew public charging station construction cost in period t ΔCcon,tPublic charging station construction cost reduction in period t ROItRate of return for EVCI operator in period t π tAverage operating profit for a public charging station in period t Sop,h,tAmount of operational subsidy in region h in period t Sop,tTotal amount of operational subsidy in period t Scon,tTotal amount of construction subsidy in period t ΔNh,e,tNumber of newly built charging stations in region h in period t rtOverall average utilization rate of public charging stations across the simulated region in period t rh,tRegional average utilization rate of public charging stations in region h in period t Appendix C Table C1 Consumer attribute settings Parameter Description Value Source IAnnual income (unit: 10,000 RMB) Urban area Uniform [5,10), prob =7 % The survey study in Appendix A Uniform [10,15), prob =22 % Uniform [15,20), prob =27 % Uniform [20,30), prob =26 % Uniform [30,100], prob =18 % Suburban area Uniform [5,10), prob =19 % Uniform [10,15), prob =36 % Uniform [15,20), prob =23 % Uniform [20,30), prob =15 % Uniform [30,100], prob =7 % DiDriving distance of user i per period (unit: 1000 km) Uniform [1.25,3.75] if I =[5,10) The survey study in Appendix A Uniform [2.50,5.00] if I =[10,15) Uniform [3.75,6.25] if I =[15,20) Uniform [4.375,6.875] if I =[20,30) Uniform [5.00,7.50] if I =[30,100] yiProportion of user’s total charging time at PCP Uniform [0.6,0.8] if I =[5,10) The survey study in Appendix A Uniform [0.5,0.7] if I =[10,15) Uniform [0.4,0.6] if I =[15,20) Uniform [0.3,0.5] if I =[20,30) Uniform [0.2,0.4] if I =[30,100] Pke Maximum acceptable EV purchase price (unit: 10,000 RMB) 1.8 *I a. The survey study in Appendix A b. (Sun et al., 2018) Pkf Maximum acceptable ICEV purchase price (unit: 10,000RMB) 1.4 *I a. The survey study in Appendix A b. (Sun et al., 2018) RkMinimum acceptable EV driving range (unit: km) 200 if I =[5,10) The survey study in Appendix A 350 if I =[10,15) (continued on next page) L. Zhu et al. Transport Policy 175 (2026) 103876 25