Electric vehicles have reshaped the way societies approach mobility, energy consumption, and environmental sustainability. In my years of studying this domain, I have seen a dramatic increase in the number of electric vehicles on roads. According to public security statistics, the electric vehicle population in China grew from 2.11 million in 2018 to 18.13 million by June 2024. Such an explosive growth necessarily brings severe challenges to the charging infrastructure, the distribution grid, and the daily travel behaviour of users. The core problem is no longer whether electric vehicles can be produced, but whether they can be charged conveniently, economically, and safely at scale. In this first-person technical review, I focus on the key methodologies and technological pathways for planning and scheduling EV charging facilities. My objective is to create a comprehensive picture that links charging demand prediction, facility siting and capacity planning, and charging dispatch strategies, because these three layers are strongly coupled and cannot be treated independently. I will also include a discussion of demonstration projects and future directions, which are essential for translating algorithms into practical benefits.
1. Introduction
The rapid electrification of road transportation has brought unprecedented attention to the charging infrastructure. From the perspective of a power system engineer, an electric vehicle is not merely a mode of transportation; it is also a distributed energy resource that can shift load in time and even discharge back to the grid when suitable technologies are adopted. However, the same resource becomes a challenge when hundreds of thousands of vehicles charge at the same period in a distribution area, causing transformer overloads, voltage drops, and expanded peak demand. To cope with these issues, planners have to ask three basic questions:
First, where and when will the EV charging demand occur? Second, how many charging stations and chargers should be built, and at which geographical locations? Third, once the stations are built, how should the charging process be dispatched in real time or near real time to benefit all stakeholders?
These questions correspond to three major research streams in the literature, namely charging demand forecasting, charging facility planning, and charging scheduling. Existing review papers tend to cover only one of these streams. For instance, some reviews focus on load forecasting, others on sitting algorithms, and others on orderly charging control. In contrast, I present a unified view that starts from the fundamental predictor of charging behaviour and culminates in the operational dispatch of charging stations. I also deliberately distinguish between different objectives: peak shaving, frequency regulation, voltage control, renewable energy accommodation, and charging cost reduction. The coupling among prediction, planning, and operation is highlighted throughout this review, because a good plan cannot be formed without a credible demand prediction, and a good operation strategy cannot be implemented without a rational facility layout.
2. Charging Demand Prediction
Charging demand prediction is the foundation of all downstream decisions. In my analysis, charging demand prediction can be decomposed into two orthogonal dimensions: the time dimension and the space dimension. The time dimension can be subdivided into short-term prediction and medium-to-long-term prediction, while the spatial dimension mainly involves the distribution of charging requests over a city or a highway network. In this section, I first discuss time-series forecasting methods, and then I move to spatial distribution models based on origin-destination matrices and travel chains.
2.1 Time-Domain Prediction
Time-domain prediction is usually classified by the forecast horizon. Short-term prediction covers horizons from a few hours to one week, and its main use is to guide operational decisions such as charging scheduling, demand response, and real-time load management. Medium-to-long-term prediction covers months to years, and its principal use is to support capacity planning of charging infrastructure and distribution network reinforcement.
Short-term charging demand has strong randomness because it depends on users’ charging habits, travel schedules, weather, and even social events. Traditional physical models cannot easily describe all these stochastic factors. As a result, data-driven and machine-learning methods become attractive. Linear regression was one of the earliest methods. A typical regression equation can be written as
$$ \hat{y}(t) = \beta_0 + \sum_{k=1}^K \beta_k x_k(t) $$
where \(\hat{y}(t)\) is the predicted charging demand at time interval \(t\), \(x_k(t)\) are explanatory variables such as temperature, humidity, day type, the demand at the same time on the previous day, and the demand at the same time on the previous week, and \(\beta_k\) are regression coefficients. The main advantages are simplicity and explainability. However, linear forms are frequently insufficient to capture strongly nonlinear patterns when the number of charging events is large and heterogeneous.
Support Vector Machines (SVMs) provide a more flexible nonlinear approach. SVM performs linear regression in a high-dimensional feature space after a kernel mapping, and it possesses good generalisation capability, especially when the sample size is limited. For electric bus charging stations, SVM-based daily load prediction has shown better accuracy than linear regression. Still, when historical data are abundant, deep neural networks usually outperform SVM due to their higher capacity for representing complex interactions.
Among neural networks, the Back-Propagation (BP) neural network is the most widely used baseline. Its drawbacks are local minima and overfitting. To mitigate these problems, researchers have proposed particle swarm optimization coupled with Bayesian regularization. More importantly, since charging demand is a time series, recurrent neural networks that explicitly model sequential dependencies have become the standard. Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) are two representative structures. Their gating mechanisms allow the model to retain relevant information over long intervals and to avoid vanishing gradients. In addition, the attention mechanism, initially developed for natural language processing, has been introduced into charging demand forecasting. An attention-enhanced LSTM or GRU can adaptively assign weights to different historical time steps, thereby improving sensitivity to sudden changes associated with holidays or extreme weather.
For medium-to-long-term prediction, the most common approach is to predict the future number of electric vehicles first, then multiply it by the average charging load per vehicle. One direct method is to extrapolate historical vehicle registration data with time-series models such as grey models or SVMs. However, electric vehicles have only been widely promoted for a short time, so historical data samples are scarce and the reliability of pure data-driven extrapolation is questionable.
By contrast, a more interpretable class of models is the Bass diffusion model, which was originally proposed to describe the adoption of new products. I find the Bass model particularly intuitive for electric vehicles because EV adoption depends on both external influences, such as advertising and government policy, and internal influences, such as word-of-mouth from existing owners. The model takes the following differential form:
$$ \frac{dN(t)}{dt} = \left( a + b \frac{N(t)}{m} \right)\left( m – N(t) \right) $$
where \(m\) is the maximum market potential, \(a\) is the innovation coefficient, \(b\) is the imitation coefficient, and \(N(t)\) is the cumulative number of adopters up to time \(t\). In practice, \(a\) and \(b\) can be adjusted dynamically to reflect policy changes and consumer satisfaction. Once the EV population trajectory is obtained, the medium-to-long-term charging demand can be estimated by simulating the charging behaviour of representative vehicles and multiplying by the population at each year.
| Method | Horizon | Typical use | Advantages | Limitations |
|---|---|---|---|---|
| Linear regression | Short | Quick baseline forecasting | Simple, transparent | Limited nonlinear capacity |
| SVM | Short | Electric bus charging stations | Good with small data | Slower with very large datasets |
| BP neural network | Short | General charging station load | Flexible nonlinear fitting | Risk of overfitting/local optima |
| LSTM/GRU | Short | Time-series charging load | Good at sequential patterns | Burden of hyperparameter tuning |
| Attention-based networks | Short | Ultra-short-term station load | Dynamic weighting of time steps | Needs large data and computation |
| Bass model | Medium/long | EV penetration and load evolution | Interpretable diffusion mechanism | Requires careful parameter calibration |
2.2 Spatial-Domain Prediction
The time dimension alone is insufficient for planning charging stations, because station siting requires us to know precisely where the charging demands are concentrated. Spatial prediction methods generally simulate the movement of individual EVs and record their charging requests. Two classes of models have emerged.
2.2.1 The OD-Matrix-Based Approach
The Origin-Destination (OD) matrix is a classic representation of travel demand. For a network with \(n\) traffic zones, the OD matrix is an \(n \times n\) matrix whose entry at row \(i\), column \(j\) denotes the number of EVs travelling from zone \(i\) to zone \(j\) in a given time interval. OD matrices can be obtained by household travel surveys or by inferring from observed road traffic counts. The latter method is more practical because it uses data from road sensors and avoids large-scale questionnaires. However, the inferred matrices are not unique: different inference models, such as the user-equilibrium approach and the stochastic user-equilibrium approach, may produce different results.
Given the OD matrix for each period, a Monte Carlo simulation can be performed. In each time step, an EV is assigned a destination according to the matrix, a feasible route is chosen based on traffic conditions, and the energy consumption is calculated. If the state of charge (SOC) drops below a threshold or if the remaining energy is insufficient for the next trip, the vehicle is routed to a nearby charging station and a charging request is created. The procedure is repeated until the simulation spans the desired study period. The main advantage of the OD-based method is that it can explicitly model traffic dynamics and route choices, making it suitable for systems with random trip demands such as electric taxis. The main disadvantage is the high cost of obtaining accurate time-dependent OD matrices.
2.2.2 The Trip-Chain-Based Approach
Trip-chain models depict the daily activity sequence of an individual. A typical trip chain starts at home (H), goes to work (W), and then performs other activities (O) such as shopping and dining before returning home. Each activity is associated with a land-use type: residential, commercial, or workplace. Common trip-chain structures include H-W-H, H-O-H, H-W-O-H, and H-O-W-O-H. The travel behaviour is characterised by several stochastic distributions:
1) the distribution of the first departure time in a day; 2) the distribution of trip length for each segment; and 3) the distribution of parking duration after each trip. These distributions are often fitted to data from national travel surveys using commonly used probability densities such as the normal, lognormal, Gaussian-mixture, or Weibull distributions.
When simulating with a trip-chain model, one first samples a trip-chain type and a first departure time. Then the trip length for each segment is sampled. The SOC is updated by subtracting the energy consumed over the travelled distance. A charging event is triggered if the SOC is lower than the driver’s habitual threshold or if the remaining energy cannot cover the next segment. After all segments of the trip chain are completed, a new vehicle is simulated. The simulation results directly indicate the charging demand in three major land-use categories and can also assign the demand to individual traffic nodes if the exact coordinates of the trip ends are sampled.
| Feature | OD-matrix approach | Trip-chain approach |
|---|---|---|
| Most suitable vehicles | Electric taxis, random travel | Private EVs, buses with fixed patterns |
| Primary data required | OD matrices at multiple times | Probability distributions of chains |
| Traffic modelling | High fidelity, dynamic route choices | Rather simple, often distance-based |
| Spatial output granularity | Each traffic zone/node | Usually residential/work/business zones |
| Computational complexity | Moderate to high | Lower when route models are unnecessary |
3. Charging Facility Siting and Capacity Planning
Planning is the bridge between demand prediction and operational efficiency. In this section I explain the multiple objectives and constraints involved, and then I present the classic mathematical models used for siting. I also discuss how to carry out capacity planning inside a station.
3.1 Objectives and Constraints in Charging Facility Planning
A charging station plan must reconcile three opposing forces. First, the charging facility operator wants to minimise the capital expenditure for land, transformers, chargers, and construction, as well as the annual operation and maintenance costs. Second, the EV user wants to minimise the time spent driving to the station, waiting in queues, and conducting the charging process. Third, the distribution network operator wants to minimise grid losses, preserve voltage profiles, and avoid overloading transformers. The optimisation criterion is thus a weighted sum of several components:
$$ C_{\mathrm{total}} = C_{\mathrm{construction}} + C_{\mathrm{O\&M}} + C_{\mathrm{user}} + C_{\mathrm{grid}} $$
where \(C_{\mathrm{user}}\) includes driving distance, waiting time, and charging time converted into monetary value, and \(C_{\mathrm{grid}}\) is the cost of network losses and potential upgrade investments. Often these costs are contradictory; for instance, installing more chargers reduces user waiting cost but increases construction cost and may force expensive feeder upgrades.
Constraints in the planning problem can be classified as:
1) charging capacity constraints, which require the total charger power to at least satisfy the peak expected demand;
2) user satisfaction constraints, which may set a maximum allowed distance between demand and station or a maximum waiting time;
3) constraints due to existing stations, so that new investments avoid geographic clustering;
4) grid operational constraints, including voltage limits, line current limits, and transformer capacity limits.
Because these constraints are multidimensional, the planning problem is usually nonconvex and discrete, making it necessary to use heuristic algorithms, integer programming solvers, or decomposition methods.
3.2 Facility Location Models
I now present the most common facility location models in the EV context. The notation is as follows. Let \(I\) be the set of demand points and \(J\) the set of candidate station sites. Let \(h_i\) denote the charging demand at point \(i\), \(d_{ij}\) the shortest travel distance from \(i\) to \(j\), and \(p\) the prescribed number of stations to build. The binary variable \(x_j\) equals \(1\) if a station is built at candidate \(j\), and \(0\) otherwise. The binary variable \(y_{ij}\) equals \(1\) if demand point \(i\) is served by station \(j\).
3.2.1 The P-Median Model
The p-median model minimises the demand-weighted travel distance between charging demand points and their assigned stations. I can write it as
$$ \min \sum_{i \in I} \sum_{j \in J} h_i d_{ij} y_{ij} $$
subject to
$$ \sum_{j \in J} x_j = p, \qquad \sum_{j \in J} y_{ij} = 1 \;\; \forall i \in I, \qquad y_{ij} \le x_j \;\; \forall i,j $$
This model is widely used because it guarantees a globally efficient service. By applying the p-median model to charging stations, one usually minimises the total cost of the operator and the users. The model can be extended to account for grid constraints through a bi-level structure where the upper-level makes investment decisions and the lower-level simulates network operation.
3.2.2 The P-Center Model
The p-center model changes the objective from total sum to the maximum distance. It seeks to minimise the largest distance from any demand point to the nearest open station. The formulation is
$$ \min D $$
subject to
$$ \sum_{j \in J} h_i d_{ij} y_{ij} \le D \;\; \forall i \in I $$
together with the same constraints as the p-median problem regarding the selection of \(p\) stations and the assignment of demand points. The p-center model is useful when one must guarantee a worst-case emergency service, but for daily EV charging it often leads to overinvestment in remote regions at the expense of efficient areas.
3.2.3 The Set-Covering and Maximum-Covering Models
The set-covering model answers a different question: given a maximal acceptable travel distance \(R\), what is the least number of charging stations needed to cover every demand point? Its objective is
$$ \min \sum_{j \in J} x_j $$
subject to
$$ \sum_{j \in J} y_{ij} \ge 1 \;\; \forall i \in I, \qquad y_{ij} \le x_j \;\; \forall i,j \in J, i \in I, \qquad y_{ij} = 0 \;\; \text{if } d_{ij} > R$$
Although the model is simple, it ignores budget constraints. The maximum-covering model instead assumes a limited budget and maximises the total amount of demand that can be covered within the allowed distance. The objective is
$$ \max \sum_{i \in I} h_i z_i $$
where \(z_i\) is a binary variable indicating whether demand point \(i\) is covered by at least one opened station, and \(z_i\) is subject to the coverage logic with respect to \(x_j\).
3.2.4 Flow-Based Models
All previous models treat charging demand as static zone-based demand. However, highway charging stations serve vehicles driving along routes rather than vehicles residing in fixed zones. The flow-capturing model considers each road segment as a candidate site and each path as a source of charging flow. Let \(Q\) be the set of paths, \(f_q\) the traffic flow on path \(q\), and \(y_q\) a binary indicator stating that at least one station is located on path \(q\). With at most \(p\) stations, we maximise the number of paths whose traffic is captured:
$$ \max \sum_{q \in Q} f_q y_q $$
subject to
$$ \sum_{j=1}^{n} x_j \le p, \qquad y_q \le \sum_{j \in q} x_j \;\; \forall q \in Q $$
Extensions include the flow-refuelling model that accounts for limited driving range and the capacitated flow-refuelling model that accounts for station capacity. In my opinion, flow-based models are the preferred choice for intercity highway charging networks, while the node-based p-median and coverage models work well for urban charging networks.
| Model | Objective | Main use | Advantages | Disadvantages |
|---|---|---|---|---|
| p-median | Min total weighted distance | Urban station siting | Cost-efficient overall | Large distances may remain |
| p-center | Min maximum distance | Emergency/equity service | Guarantees worst case | May be expensive for EV users |
| Set covering | Min number of stations | Coverage guarantee | Clear coverage criterion | Ignores capacity and budget |
| Maximum covering | Max covered demand under budget | Budget-constrained planning | Effective with limited resources | Does not account for volume at station |
| Flow capturing | Max captured paths | Highway charging network | Directly models through traffic | Requires path flow data |
3.3 Capacity Planning inside a Charging Station
Capacity planning is concerned with the number of chargers and their power ratings inside a station. The internal capacity is determined by the forecasted arrival process of EVs. A standard model is to treat the station as a multi-server queue, where EVs arrive according to a Poisson process with rate \(\lambda\) and the service rate of each charger is \(\mu\). The objective is to minimise the sum of construction cost and user waiting cost:
$$ \min_{s} \left[ C_f(s) + c_w \lambda W_q(s) \right] $$
where \(s\) is the number of chargers, \(C_f(s)\) is the fixed and operational cost, \(c_w\) is the value of waiting time per user per hour, and \(W_q(s)\) is the average waiting time, which in an \(M/M/c\) queue is expressed as
$$ W_q(s) = \frac{ P_0 \displaystyle \frac{\rho^{s}}{s!} \frac{\rho}{s-\rho} }{ \lambda (1 – \rho) } $$
with some care regarding the allowable value of traffic intensity \(\rho=\lambda/(s\mu)\). It is important to choose \(s\) such that \(\rho < 1\). In many practical planning problems, one first computes the envelope of charging demand over the day and then selects the number of chargers that keeps the waiting time below a pre-specified threshold even in the busiest period. Some recent works have coupled the capacity planning with renewable generation and storage, resulting in a more complex economic optimisation where the power ratings of photovoltaic arrays and batteries are also optimised.
4. Charging Scheduling Strategies
After stations are built, the next question is how to schedule the charging loads in time and space. Uncontrolled charging may create an even stronger evening peak because many EV owners plug in immediately upon returning home. Scheduling strategies transform EV charging from a passive resistive load into a flexible resource. The literature has investigated different scheduling objectives: peak shaving, frequency regulation, voltage support, renewable integration, and charging cost reduction. Before deriving strategies, however, it is necessary to estimate the available regulation capacity of an EV fleet.
4.1 Schedulable Potential of an EV Fleet
Most strategies control an aggregation of EVs rather than an individual car. The schedulable potential of the fleet is the set of feasible charging/discharging power trajectories that satisfy each vehicle’s travel requirements. For a single EV, the feasible charging power \(p_{e,t}\) at time \(t\) is bounded by the charger limit:
$$ – \underline{P}_e \le p_{e,t} \le \overline{P}_e $$
where \( \overline{P}_e\) is the maximum charging power, and \(\underline{P}_e\) is the maximum discharging power if vehicle-to-grid (V2G) is allowed. The battery state of charge evolves according to
$$ SOC_{e,t+1} = SOC_{e,t} + \eta \frac{p_{e,t} \Delta t}{E_{e}} $$
where \(\eta\) is the conversion efficiency, \(\Delta t\) is the time step, and \(E_e\) is the battery capacity. Terminal constraints impose that the SOC at the departure time must be above the driver’s minimum requirement. For a fleet of \(N\) vehicles, the aggregated power \(P_t = \sum_{e=1}^N p_{e,t}\) can be estimated via Monte Carlo simulation, deep-learning prediction, or by summing the individual feasible intervals using techniques from set theory. This schedulable potential is then used as the input to the scheduling optimisation.
4.2 Peak-Valley Shifting Strategies
Peak-valley shifting is perhaps the oldest and most widely applied scheduling objective. The goal is to flatten the net load curve by encouraging EVs to charge during valley hours and curtail charging during peak hours. Without control, many EVs start charging at 18:00, coinciding with evening residential load. A simple remedy is the dual-sequence valley-filling strategy, which splits EVs into two groups: one group starts charging at the beginning of the valley-price period, and the other group delays charging such that the charging ends exactly when the valley-period tariff ends. By carefully determining the fraction assigned to each group, the daily maximum load can be reduced.
More generally, the scheduling problem can be formulated as follows. Let \(P_t^{\text{base}}\) be the non-EV base load and let \(p_{e,t}\) be the controlled charging power of EV \(e\). We minimise the variance of the total load:
$$ \min \; \frac{1}{T} \sum_{t=1}^{T} \left( P_t^{\text{base}} + \sum_{e=1}^{N} p_{e,t} – \bar{P} \right)^2 $$
subject to battery dynamics, user departure constraints, and station power limits. By solving this quadratic optimisation problem, EVs are automatically shifted to low-load periods. In a residential area, such optimisation can significantly reduce the peak-to-valley ratio and smooth the transformer loading.
4.3 Frequency and Voltage Regulation
Frequency regulation needs fast power response. Although a synchronous generator can respond within seconds, power-electronic-interfaced EV chargers can respond even faster. To participate in frequency regulation, the charging controller must adjust the active power of the fleet based on a frequency deviation or a regulation signal. Some studies adopt a sliding-mode proportional-integral controller to allocate power commands among EVs while minimising the total regulation mileage and maintaining battery life.
Voltage regulation requires the coordination of active and reactive powers. In a distribution network, a charging station can inject reactive power through its bidirectional converter even while it is delivering active power to the battery. An integrated model may include capacitor banks, on-load tap changers, and charging stations in a joint optimisation. One approach is a day-ahead scheduling model that treats the reactive power capability of chargers as a resource. The day-ahead stage determines the state-of-charge targets for each EV, while the real-time stage adjusts the reactive power to hold voltages within acceptable limits. This coordinated architecture has been tested in simulations and has shown that EV chargers can reduce the number of tap operations and lower the need for static var compensators.
4.4 Renewable Energy Accommodation
Renewable sources like solar and wind have a variable output. EV batteries are an attractive buffer because they can be charged when the renewable generation is abundant and, if V2G is allowed, they can release the stored energy when renewables drop. The scheduling problem then becomes a price-based or direct-control matching problem. Let \(G_t\) denote the renewable generation profile. The objective could be to maximise the renewable self-consumption or to minimise the curtailed energy:
$$ \min \sum_{t=1}^{T} \left( G_t – \sum_{e=1}^{N} p_{e,t} – P_t^{\text{battery}} \right)^2 $$
where \(P_t^{\text{battery}}\) represents other storage devices. In practice, renewables are uncertain, so deterministic optimisation is insufficient. A robust option is a two-stage scheduling method: in the day-ahead stage, one predicts renewable output; in the real-time stage, one corrects deviations by charging or discharging EVs. Using such a two-stage strategy, researchers have reported improved utilisation of wind and solar resources and reduced load shedding.
4.5 Charging Cost Reduction
Cost reduction can be examined from the viewpoint of the user and the charging station operator. Some strategies rely on direct control, where users delegate their charging rights in exchange for a discount or incentive. The coordinator performs centralised optimisation that balances the electricity cost and the battery degradation cost. Other strategies rely on indirect control through price signals, such as time-of-use tariffs or real-time dynamic prices.
Under a time-of-use tariff, an EV owner naturally intends to postpone charging to the low-price hours. However, if all owners postpone simultaneously, a new peak at the beginning of the valley period may appear. To overcome this rebound peak, the operator can introduce a game-theoretic interaction among EVs. The station charges a price that varies slightly according to the instantaneous demand, and each EV responds to the price by choosing its start time and power. A Nash equilibrium usually yields a more evenly distributed load profile than the naive price following. The cooperation between users and the operator can also be analysed as a cooperative game, where the total welfare gain from orderly charging is fairly allocated between the participants.
Table listing the scheduling objectives and their appropriate control variables:
| Objective | Controlled quantity | Typical actor | Time scale | Common techniques |
|---|---|---|---|---|
| Peak valley shifting | Active charging power start time | Aggregator / grid operator | Day-ahead to real-time | Quadratic programming, heuristics |
| Frequency regulation | Active power fast response | TSO/aggregator | Seconds to minutes | Droop control, PI sliding mode |
| Voltage regulation | Reactive power of chargers | DSO/station operator | Minutes to hours | OPF, hierarchical control |
| Renewable accommodation | Charging and discharging power | Microgrid / VPP | 15 min to day | Two-stage stochastic/robust |
| Cost minimization | Charging price and schedule | Station or owner | Day-ahead | TOU pricing, game theory |
5. Demonstration Programs
While simulation models illustrate the theoretical benefits, field demonstrations reveal the practical barriers and possibilities. I have followed several demonstration projects around the world with great interest. They share a common theme: users respond to pricing and incentives, but the response rate varies by charging context.
5.1 A Demand Response Pilot in Shanghai, China, 2019
In this demand response project, private residential chargers, workplace destination chargers, and battery-swap stations were all included. Different compensation prices were designed for downward load-shifting (valley filling) and upward load-shifting (peak shaving). The results showed that private chargers had a low participation rate of only about 5.3 percent for valley filling, but their potential capacity was very large due to the massive number of chargers. Workplace chargers achieved roughly 75 percent response for peak shaving because fleet managers could control the chargers remotely. Battery-swap stations achieved the highest response rate of 81.2 percent when notified 30 minutes in advance, because the station could schedule battery charging in a more flexible manner. This project taught me that the type of charging infrastructure drastically affects the ability for an EV fleet to engage in demand response.
5.2 An Industrial-Park V2G Demonstration in Baoding, China, 2021
A vehicle-to-grid demonstration project was launched in an industrial park in Baoding. The project installed 50 bidirectional charging points and encouraged employees to discharge their EVs during peak load hours and recharge during off-peak hours. The operator developed a reward mechanism based on the monthly discharged energy. The highest monthly reward reached as much as 2000 RMB for a participant, which was a strong incentive. The project also highlighted the importance of battery warranty: the original equipment manufacturer promised to maintain the battery warranty for all vehicles participating in the project. This demonstration is an excellent example of aligning the interests of drivers, vehicle manufacturers, and grid operators.
5.3 A Large-Scale Vehicle-Grid Interaction Centre in Wuxi, China, 2023
One of the largest vehicle-grid interaction centres in China was put into operation in Wuxi in August 2023. The centre has an area exceeding 14,500 square metres and can support reverse power injection from 50 electric vehicles at the same time. The maximum reverse power is nearly 2,000 kW within 30 minutes. During the first period of operation, the centre completed more than 300 discharge events, with a cumulative discharge energy exceeding 1,800 kWh. Furthermore, the aggregated resource has been connected to a virtual power plant platform, enabling the building to contribute more than 2.2 million kWh of electricity during peak periods over the year. The second phase of the project plans to add superchargers, mobile charging, and battery swapping facilities, which will allow the centre to serve 144 vehicles for charging, 50 vehicles for discharging, and 400 vehicles for battery swapping simultaneously.
5.4 Parker Project in Denmark
The Parker project ran from 2016 to 2018 and was one of the earliest attempts to commercialize vehicle-to-grid frequency regulation in Denmark. Fifty 10 kW bidirectional chargers were used to demonstrate that EVs can deliver the frequency regulation services required by the Danish transmission system operator. The project confirmed that the CHAdeMO protocol was a technically capable V2G standard. It also exposed the economic challenge: the revenue from frequency regulation was highly dependent on market prices, and the project did not produce a stable profit. This underlines that technical feasibility does not automatically translate into a robust business model.
5.5 SunnYparc Project in Switzerland
The SunnYparc project is a multi-year demonstration of solar-powered charging combined with V2G. The project aims to install a photovoltaic system generating about 1 GWh per year, serving local residences and 250 parking spaces. Fifty of those charging stations are bidirectional. During sunny periods, surplus solar power is stored in EV batteries; during evening peak hours, the stored energy is discharged to cover residential demand. Since Switzerland permits private microgrids, the project uses market-based pricing to coordinate the individual decisions of vehicle owners. The lessons from this project will be very valuable for future prosumer communities with high renewable penetration.
| Project | Country | Primary aim | Response or outcome | Key lesson |
|---|---|---|---|---|
| Shanghai demand response | China | Peak shaving and valley filling | Private: 5.3%, workplace: 75%, swap: 81.2% | Response rate depends heavily on charging type |
| Baoding V2G | China | Vehicle-grid interaction | Rewards up to 2000 RMB per month per user | Battery warranty enables user trust |
| Wuxi vehicle-grid centre | China | Fast & discharging services | 50 vehicles discharge at nearly 2 MW | VPP integration unlocks large-scale flexibility |
| Parker | Denmark | Frequency regulation | Technically provided all required frequency services | Market economics still barrier |
| SunnYparc | Switzerland | Solar self-consumption | 1 GWh PV and 250 chargers | Market pricing coordinates distributed decisions |
6. Existing Research Limitations and Future Directions
Despite the progress I have reviewed, several important open problems remain. In my opinion, these challenges represent the most fruitful areas for future research.
6.1 More Realistic Modelling of EV User Behaviour
Charging behaviour is stochastic and strongly influenced by individual preferences. Existing OD-matrix and trip-chain models use fairly simple decision rules for when to charge, often based only on SOC thresholds. They rarely capture the heterogeneity of users, the impact of real-time traffic information, holiday effects, or the strategic behaviour of users reacting to prices. A path forward is to leverage large-scale data from charging networks and travel surveys, while ensuring privacy-preserving data sharing. Federated learning and distributed optimization could allow models to be trained on private user data without centralizing personal information.
6.2 Coevolution of Charging Demand and Infrastructure
There is a bidirectional feedback loop: charging infrastructure determines how convenient electric vehicles are, and the growing number of EVs creates demand for new infrastructure. Most studies treat the future EV population as an input, but the diffusion of EVs should depend on the coverage and reliability of the charging network. An enhanced Bass model where the imitation coefficient is made a function of the charging station density could produce more credible long-term demand forecasts. This is particularly important for regulators who need to decide how to phase charging investment over years.
6.3 Coordinated Planning of Charging Facilities and Distribution Networks
EV growth may require reinforced distribution transformers and feeders. Planning charging stations without considering active network management can lead to excessive connection costs or the need to reinforce the grid. Conversely, a coordinated multi-stage planning model would jointly decide the location of charging stations and the upgrades of network assets. Stochastic programming and robust optimisation are suitable tools because future charging demand is highly uncertain.
6.4 Design of Fair and Effective Pricing Mechanisms
Most simulation studies assume that an aggregator can directly control EVs. In practice, EV users must be persuaded by an appropriate price mechanism. The challenge is to design a tariff that reflects the marginal cost of electricity, compensates users for their flexibility, and discourages gaming behaviour. Cooperative game theory can provide a fair distribution of the total benefits among users and operators. In addition, price-sensitivity models should be built from real demonstration data, including the data from the demand response and V2G projects I described in the previous section.
6.5 New Architectural Models for EV Participation in Ancillary Services
The use of large fleets of electric vehicles for ancillary services is still fragmented. Some projects control private chargers, others control fleet chargers or battery-swap stations. A unified framework may be needed to aggregate resources from different sectors. For private chargers, blockchain-based peer-to-peer trading could enable decentralized transactions. For public fast-charging stations, a two-level game between the station operator and individual EVs could coordinate the charging schedule. For dedicated fleet and swapping stations, the resources can be treated as a virtual power plant that has a relatively predictable schedule and can thus participate in day-ahead markets. Developing such multi-tier architectures is an important future step.
| Research gap | Explanatory example | Potential solution |
|---|---|---|
| User behaviour modelling | Charging decisions depend on psychological thresholds | Privacy-preserving data gathering, federated learning |
| Demand-infrastructure coevolution | More chargers encourage more EV purchases | Endogenous diffusion models with availability feedback |
| Distribution network coordination | Charging station placement may cause feeder congestion | Multi-stage joint planning, stochastic optimisation |
| Pricing mechanism | Direct control ignores user willingness | Cooperative game theory, real price-response models |
| Participation architecture | Various charging contexts need separate aggregation | Structured VPP, blockchain P2P, hierarchical games |
7. Conclusions
In this first-person review, I have systematically examined the key technologies for planning and operating electric vehicle charging facilities. I began with charging demand prediction, where electric vehicle charging demand is split into temporal and spatial dimensions. Temporal forecasts are obtained by regression and AI models for the short term and by Bass-type diffusion models for the long term. Spatial forecasts rely on OD matrices for taxis and random traffic and on trip-chain models for private vehicles with regular schedules. I then moved to facility planning, reviewing the p-median, p-center, set-covering, maximum-covering, and flow-based models. Every model has a place: p-median is a workhorse for urban systems, p-center is suitable for guaranteeing equity, and flow-based models are best for long highways. Capacity planning inside the station can use queueing theory to balance user waiting time with operator costs.
I also discussed charging scheduling strategies from the perspective of different system-level objectives. EV loads can be shifted to achieve valley filling, quickly modulated for frequency control, coordinated with reactive power for voltage support, used to absorb renewable generation, and managed to minimise charging cost for users and operators. The demonstration programs in China, Denmark, and Switzerland show that technological readiness is high, but the user participation rate and economic viability still require careful incentive and market design. Future challenges call for more realistic user behaviour models, coevolution of demand and infrastructure, coordinated planning of grid and chargers, advanced pricing mechanisms, and multi-level aggregation frameworks. I believe that only by integrating demand forecasting, infrastructure planning, and operational scheduling can we meet the enormous charging needs of modern and future electric vehicles.