As the global shift toward transportation electrification accelerates, electric vehicles (EVs) have emerged as a cornerstone of sustainable mobility. In China, the rapid adoption of EVs has highlighted critical challenges in charging infrastructure, particularly in urban areas where public charging station distribution often fails to align with dynamic user demands. Traditional planning approaches, which rely on static data or uniform assumptions, struggle to address the spatial and temporal heterogeneity of charging needs. This study introduces a novel framework for differentiated charging facility planning, leveraging simulated trip chains to capture the complex interplay between user behavior, urban functionality, and energy constraints. By integrating multi-source data—including EV operational records, built-environment attributes, and charging station profiles—we extract multidimensional user characteristics and employ a two-step clustering model to identify distinct user groups. These groups inform a large-scale trip chain simulation that predicts charging demand patterns, while community detection algorithms partition urban areas into functionally distinct zones. The resulting planning model incorporates location-specific constraints and attractiveness weights, optimizing charging station placement and capacity to reduce grid stress, minimize carbon emissions, and enhance service efficiency. Empirical validation using data from Beijing demonstrates a 3.35% reduction in electricity consumption oversaturation and superior utilization rates for newly deployed facilities during peak hours.
The proliferation of electric vehicles in China has catalyzed unprecedented demand for public charging infrastructure. However, the uneven distribution of charging facilities across urban landscapes exacerbates range anxiety and undermines user confidence. Current public charger-to-vehicle ratios, such as 1:7 in China, reveal significant gaps in accessibility. Moreover, profitability for operators hinges on achieving daily utilization thresholds—e.g., 1.68 hours for a 50 kW DC charger—which are often unattainable under suboptimal planning. Existing studies frequently rely on Monte Carlo simulations or simplified assumptions, neglecting the nuanced behavioral patterns of EV users. In contrast, our approach harnesses high-resolution GPS data, point-of-interest (POI) information, and traffic networks to model realistic trip chains. By segmenting users based on temporal, spatial, and energy dimensions, we simulate charging decisions influenced by distance decay and Huff attractiveness models. This enables the identification of oversaturated zones and the formulation of differentiated strategies for charger deployment, accounting for regional variations in land use, pricing, and demand hotspots.
To quantify user heterogeneity, we extract features from EV operational data, including travel start/end times, state-of-charge (SOC) dynamics, and visited locations. Spatial entropy $H_{ ext{Space}}$ and temporal entropy $H_{ ext{Time}}$ are computed as follows:
$$H_{ ext{Space}} = -\sum_{i=1}^{N_A} P^S_{ ext{AOI},i} \log_2 P^S_{ ext{AOI},i}$$
where $P^S_{ ext{AOI},i}$ denotes the probability of a state event occurring in the $i$-th area of interest (AOI), and $N_A$ is the number of distinct visited spaces. Similarly, temporal entropy captures the dispersion of activities across time intervals. Jaccard similarity $J_{ ext{Daily}}$ assesses trajectory consistency:
$$J_{ ext{Daily}} = \frac{\sum_{\phi=1}^{D} \sum_{\varphi=1}^{D} \frac{|\Omega_{ ext{AOI},\phi} \cap \Omega_{ ext{AOI},\varphi}|}{|\Omega_{ ext{AOI},\phi} \cup \Omega_{AOI,\varphi}|}}{C^2_D}$$
where $\Omega_{ ext{AOI},\phi}$ represents the set of visited areas on day $\phi$, and $D$ is the total days analyzed. A two-step clustering model combines Gaussian mixture clustering for continuous features (e.g., travel distance, SOC change) and spectral clustering for categorical attributes (e.g., spatial entropy, functional properties). The Gaussian mixture probability density is expressed as:
$$p_{ ext{gau}}(x_{ ext{gau}}) = \sum_{k=1}^{K_{ ext{gau}}} \pi_k f(x_{ ext{gau}} | \mu_{ ext{gau},k}, \Sigma_{ ext{gau},k})$$
where $\pi_k$ are mixing coefficients, and $f$ is the Gaussian density function. Spectral clustering constructs an affinity matrix $A_{i,j}$ using Gaussian similarity:
$$A_{i,j} = e^{-\frac{\|x_i – x_j\|^2}{2\sigma^2}}$$
followed by eigenvalue decomposition and K-means clustering on the resulting eigenvectors.
| Dimension | Level 1 | Level 2 | Level 3 |
|---|---|---|---|
| Time | Time nodes | Start/end times of states | Temporal entropy |
| Process duration | Duration of states | — | |
| — | — | — | |
| Space | Spatial nodes | Start/end locations | Spatial entropy, functional attributes |
| Spatial transfer | Euclidean distance, actual mileage | Jaccard similarity | |
| — | — | — | |
| Energy | Energy nodes | SOC at start/end | — |
| Energy change | Electricity consumption/gain | — |
Trip chain simulation generates daily itineraries for synthetic EV users, incorporating stochastic decisions based on SOC, proximity to chargers, and time-dependent factors. The simulation assumes mandatory charging when SOC falls below 20% and prioritizes slow charging when feasible. Key equations include the actual travel distance $d_{ ext{travel}}(r)$ derived from Euclidean distance $d_{ ext{euc}}(r)$:
$$d_{ ext{travel}}(r) = 5.243 \cdot d_{ ext{euc}}(r)^{0.873}$$
and travel time $t_{ ext{travel}}(r)$:
$$t_{ ext{travel}}(r) = 1200 \cdot \left( \frac{d_{ ext{travel}}(r)^{1.875}}{1.4002 imes 10^{-8}} \right)$$
SOC update after a trip is modeled as:
$$SD_{ ext{end}}(r) = SD_{ ext{start}}(r) – \frac{d_{ ext{travel}}(r) \cdot ext{ECR}_\kappa}{100 \cdot ext{CAPA}_\kappa}$$
where $ ext{ECR}_\kappa$ is the energy consumption rate and $ ext{CAPA}_\kappa$ is the battery capacity of vehicle type $\kappa$. Charging time $t_{ ext{charge}}(r)$ depends on the energy replenished $\Delta ext{CAPA}$ and charging power $P_{ ext{ch}}^r$:
$$t_{ ext{charge}}(r) = \frac{\Delta ext{CAPA}}{P_{ ext{ch}}^r} \cdot 3600$$
User charging station selection employs a two-step floating catchment area (2SFCA) method combined with a dynamic Huff model. The accessibility score $A_i$ for grid $i$ is:
$$A_i = \sum_{j=1}^m \frac{ heta_j f(\alpha t_{i,j})}{\sum_{i=1}^g u_i f(\alpha t_{i,j})}$$
where $f(\alpha t_{i,j}) = \gamma e^{-\zeta (\alpha t_{i,j} – t_0)^2}$ is the distance decay function, $ heta_j$ is the average number of chargers per station in grid $j$, and $u_i$ is the user count in grid $i$. The probability $P_{T, u_{ ext{ev}}, k_{ ext{cs}}}$ of user $u$ choosing station $k$ at time $T$ is:
$$P_{T, u_{ ext{ev}}, k_{ ext{cs}}} = \frac{A_{T, k_{ ext{cs}}} / t_{T, u_{ ext{ev}}, k_{ ext{cs}}}}{\sum_{k=1}^{M_{ ext{travel}}} A_{T, k_{ ext{cs}}} / t_{T, u_{ ext{ev}}, k_{ ext{cs}}}}$$
Station attractiveness $A_{T, k_{ ext{cs}}}$ incorporates factors like charger count, fast-charger availability, pricing, POI density, and information completeness, weighted by coefficients derived from empirical data.
| EV Model | Battery Capacity (kWh) | Energy Consumption (kWh/100 km) | Max Fast-Charging Power (kW) |
|---|---|---|---|
| Tesla Model Y | 60 | 13.8 | 250 |
| Tesla Model 3 | 60 | 13.7 | 250 |
| BYD Qin Plus EV | 57.6 | 11.9 | 70 |
| BYD Yuan Plus | 60.48 | 12.5 | 80 |
| BYD Han EV | 85.4 | 13.5 | 120 |
| VW ID.3 | 52.8 | 13.1 | 100 |
Community detection partitions the urban grid into 13 distinct zones using a link network $G = (V, L, W)$, where nodes $V$ represent grids, edges $L$ denote direct trips with charging events, and weights $W$ reflect trip frequency and energy consumption. The Fast Unfolding algorithm maximizes modularity $Q$:
$$Q = \frac{1}{2\varpi} \sum_{i,j} \left( w_{ij} – \frac{\xi_i \xi_j}{2\varpi} \right) \delta(c_i, c_j)$$
where $\varpi$ is the total edge weight, $\xi_i$ is the sum of weights incident to node $i$, and $\delta(c_i, c_j)$ is 1 if nodes $i$ and $j$ belong to the same community. This yields zones with strong internal connectivity and minimal inter-zone reliance, facilitating targeted infrastructure planning.
The planning model adopts a two-phase approach: candidate site selection and charger capacity determination. The objective function for site selection minimizes weighted distance to oversaturated grids while incorporating location-specific attractiveness:
$$\min \sum_{i,j \in V} o_i \left[ -\left( \sum_{p=1}^P \beta_{i,p} G_{ ext{POI},i,p} + \chi_1 G_{ ext{park},i} + \chi_2 G_{ ext{jam},i} \right) + \chi_3 \sum_{j \in V} d_{i,j} G_{ ext{CDs},j} \right]$$
subject to $d_{i,j} \leq R$ and $\sum o_i \leq O_{ ext{total}}$, where $o_i$ is a binary variable for selection, $\beta_{i,p}$ are geographically weighted regression coefficients for POI type $p$, $G_{ ext{POI},i,p}$ is the POI count, $G_{ ext{park},i}$ is parking lot availability, $G_{ ext{jam},i}$ is traffic congestion index, $d_{i,j}$ is Euclidean distance, and $G_{ ext{CDs},j}$ is excess charging demand. Capacity optimization maximizes demand coverage while minimizing costs:
$$\max \sum_{i \in V} \left[ \lambda Z_i – M_i C_{ ext{op,s}} – N_i (C_{ ext{ch}} + C_{ ext{op,p}} + C_{ ext{ca}} P_{ ext{ch}} + C_{ ext{inv}}) \right]$$
with constraints $0 \leq N_i \leq \min(N_{\mu, ext{max}}, M_{ ext{park},i} N_{\mu, ext{avg}})$ and $0 \leq M_i \leq M_{ ext{park},i}$, where $N_i$ is the number of new chargers, $M_i$ is the number of new stations, $Z_i$ is the分担ed excess demand, and $C$ terms represent operational and investment costs.
| Scheme | Candidate Sites | New Chargers | Energy Consumption (MWh) | CO2 Emissions (t) | Potential Excess Demand |
|---|---|---|---|---|---|
| Baseline | 0 | 0 | 3943.4 | 2698 | — |
| Scheme 1 | 400 | 7812 | 3921.5 | 2683 | 4868 |
| Scheme 2 | 500 | 9234 | 3975.6 | 2720 | 4970 |
| Scheme 3 | 600 | 10543 | 3914.1 | 2678 | 4803 |
Validation using Beijing data shows that Scheme 3 reduces electricity consumption in oversaturated areas by 3.35% and achieves higher utilization rates for new chargers during peak hours compared to existing facilities. Carbon emissions are calculated as:
$$B_{q,\mu} = \sum_{\vartheta=1}^{\eta_q} E_\vartheta$$
where $B_{q,\mu}$ is the indirect emissions of greenhouse gas $q$ in community $\mu$, $\eta_q$ is the provincial grid emission factor, and $E_\vartheta$ is the energy consumption. The differentiated planning approach yields a more resilient charging network, with emissions reductions observed in both urban cores and suburban areas.
This study underscores the importance of data-driven, behavior-aware planning for electric vehicle infrastructure in China. By accounting for user heterogeneity and urban spatial structure, our model enables scalable and sustainable EV integration. Future work will incorporate real-time routing algorithms and renewable energy variability to further refine planning accuracy. As China’s EV market continues to expand, such tailored strategies will be crucial for balancing economic viability with environmental goals.