As electric vehicle adoption accelerates, power systems are facing a new planning challenge: how to forecast charging demand with greater precision, how to understand where and when electric vehicle users will need energy, and how to turn charging and discharging into a grid-supporting resource rather than a source of stress. A research project at Southwest University of Science and Technology has examined these questions through three linked studies covering electric vehicle charging load forecasting, spatio-temporal electric vehicle charging demand forecasting, and multi-objective electric vehicle charging-discharging optimization. The work reports improvements in forecasting accuracy and a strategy that can reduce user costs while smoothing grid load in simulation.
The research arrives at a moment when electric vehicle growth is no longer only a transport story. It is also an energy-management story. Every electric vehicle connected to a charger becomes part of the load curve, and every charging decision can either add pressure to peak hours or help absorb low-valley electricity. The study therefore treats electric vehicle charging not as a fixed demand but as a flexible process that can be forecast, modeled, and optimized. The central message is that better electric vehicle data modeling and better electric vehicle dispatch can work together to support grid stability, user benefits, and cleaner energy use.
1. Electric Vehicle Growth Makes Charging Management a Grid Priority
The research begins from a straightforward observation: electric vehicle charging load is highly random and complex. Users arrive at different times, charge for different durations, and respond to prices, weather, and daily schedules in different ways. When large numbers of electric vehicle users charge during peak periods, the result can be sharper load peaks, greater peak-valley differences, and higher stress on distribution networks. When electric vehicle charging is coordinated, however, the same fleet can become a distributed energy storage resource. This is the logic behind vehicle-to-grid, or V2G, in which electric vehicle batteries can charge from the grid and, when conditions allow, discharge back to the grid.
The study positions electric vehicle charging load forecasting as an early warning and planning tool. If grid operators can anticipate electric vehicle charging load, they can prepare generation, reserve capacity, and network operation more effectively. The study also positions electric vehicle charging demand forecasting as a spatial planning tool. Charging demand is not evenly distributed across a city, and an electric vehicle driver in one traffic zone may face a very different charger availability situation from a driver in another zone. Finally, the study positions charging-discharging optimization as an operational tool. Once electric vehicle demand is known, prices and constraints can be used to shift charging toward valleys and encourage discharging during peaks.
2. Three Research Gaps Addressed by the Electric Vehicle Study
The research identifies three practical gaps in existing electric vehicle charging research. These gaps shape the structure of the work and explain why the study combines decomposition, attention mechanisms, graph-based learning, and multi-objective optimization in one framework.
- Electric vehicle charging load forecasting methods often fail to mine deep and multi-scale features from load sequences, limiting accuracy when electric vehicle charging behavior is volatile.
- Electric vehicle charging demand forecasting methods often do not fully integrate spatial and temporal information, which weakens prediction when charging demand varies across both time and location.
- Electric vehicle charging-discharging strategies often focus on only one side of the problem, either user interests or grid interests, and sometimes do not fully account for electric vehicle charging demand.
To address these gaps, the research develops a multi-scale electric vehicle charging load forecasting method, a spatio-temporal electric vehicle charging demand forecasting method, and a multi-objective electric vehicle charging-discharging strategy. Each part uses the output or logic of the others: more accurate electric vehicle load forecasting supports grid planning, more accurate electric vehicle demand forecasting supports spatial charging management, and a multi-objective strategy supports the practical use of electric vehicle flexibility.
3. Multi-Scale Electric Vehicle Charging Load Forecasting
The first part of the study focuses on electric vehicle charging load forecasting. The research argues that directly feeding raw electric vehicle load data into a forecasting model is not enough, because the original sequence contains random fluctuations and mixed frequency components. The proposed method therefore combines variational mode decomposition, a whale optimization algorithm, attention mechanisms, and multi-scale temporal convolutional network residual blocks. The aim is to reduce the complexity of the electric vehicle load sequence and then extract features at several time scales.
3.1 Decomposition and Weather Feature Selection
The electric vehicle charging load data were aggregated to an hourly level. The study also examined weather data and calendar information, because electric vehicle charging behavior can be influenced by temperature, pressure, humidity, wind, weekday patterns, and holidays. A multiple linear regression model was used to assess the influence of weather variables on electric vehicle charging load. The reported coefficients are shown below.
| Weather variable | Coefficient |
|---|---|
| Air temperature | 0.81 |
| Horizontal atmospheric pressure | -0.86 |
| Sea-level atmospheric pressure | 0.85 |
| Three-hour pressure change | 1.33 |
| Surface relative humidity | 0.05 |
| Mean wind speed | 0.28 |
For the electric vehicle load sequence, the study used the whale optimization algorithm to tune the parameters of variational mode decomposition. The method searched over decomposition numbers and penalty factors, using a minimum envelope entropy criterion to evaluate candidate parameter sets. In the reported experiment, the electric vehicle charging load sequence was decomposed into five subseries. These subseries showed different behaviors: one was relatively smooth, another showed periodic fluctuation, and others were more irregular and may reflect random electric vehicle user behavior.
3.2 Attention-Based Temporal Convolutional Residual Blocks
The forecasting model uses temporal convolutional network residual blocks improved with gating, temporal attention, and channel attention. The gating mechanism helps regulate information flow inside the residual block. Temporal attention assigns different weights to different time steps, allowing the model to focus on important moments in the electric vehicle load sequence. Channel attention adjusts the importance of feature channels, helping the model emphasize useful representations. Three improved residual blocks with different filter sizes and dilation rates are used in parallel to capture multi-scale electric vehicle load features. Their outputs are combined and weighted by an attention layer before the final prediction layers.
The reported model settings included filter sizes of 3, 5, and 7; dilation rates of 1, 2, and 4; 50 training epochs; the Adam optimizer; mean absolute error as the loss function; and a learning rate of 0.001. The dataset came from the ElaadNL electric vehicle charging project and covered January 1 to June 30, 2019, with 163,255 charging records. Weather data for the same period contained 4,344 hourly records. The final 1,200 hours of electric vehicle charging load were used as the test set.
3.3 Reported Electric Vehicle Load Forecasting Results
The study compared the proposed electric vehicle charging load forecasting method with several models, including GRU, LSTM, CNN-Attention-LSTM, FCA-TCN, EMD-LSTM, VMD-CNN, VMD-LSTM, and ITCN-AT-BiGRU. The reported error results are summarized below.
| Model | MSE (kW2) | MAE (kW) | R2 |
|---|---|---|---|
| GRU | 46.28 | 5.02 | 0.618 |
| LSTM | 43.26 | 4.89 | 0.643 |
| CNN-Attention-LSTM | 41.15 | 4.67 | 0.653 |
| FCA-TCN | 47.33 | 4.97 | 0.609 |
| EMD-LSTM | 25.84 | 3.80 | 0.787 |
| VMD-CNN | 7.31 | 2.05 | 0.937 |
| VMD-LSTM | 4.53 | 1.63 | 0.960 |
| ITCN-AT-BiGRU | 3.43 | 1.42 | 0.970 |
| Proposed method | 1.47 | 0.91 | 0.987 |
The reported results show that the proposed method achieved the lowest mean square error and mean absolute error among the compared models, and the highest R2 value. The study also conducted an ablation experiment to isolate the contribution of decomposition, improved residual blocks, and multi-scale feature fusion. The ablation results are shown below.
| Model variant | MSE (kW2) | MAE (kW) | R2 |
|---|---|---|---|
| TCN | 47.55 | 4.97 | 0.607 |
| VMD plus TCN | 7.59 | 2.09 | 0.934 |
| VMD plus improved residual TCN | 2.39 | 1.21 | 0.980 |
| VMD plus multi-scale residual TCN | 1.67 | 1.01 | 0.985 |
| Proposed method | 1.47 | 0.91 | 0.987 |
According to the study, decomposition helped reduce the complexity of the electric vehicle charging load sequence, while the improved residual block helped the model focus on important time points and feature channels. Multi-scale feature fusion further improved the model’s ability to handle diverse electric vehicle load patterns. For grid operators, this line of work suggests that more accurate electric vehicle charging load forecasting can provide better input for scheduling and capacity planning.
4. Spatio-Temporal Electric Vehicle Charging Demand Forecasting
The second part of the study addresses electric vehicle charging demand forecasting. Here, the problem is not only how much electric vehicle charging load will appear, but where charging demand will be concentrated. The research models charging stations or traffic zones as nodes in a graph, with spatial relationships represented by an adjacency matrix. Each node has features such as charger occupancy, charging price, and geographic information. The forecasting task uses past electric vehicle charging demand to predict future demand over different horizons.
4.1 Graph Embedding and Adaptive Attention
The proposed electric vehicle charging demand forecasting framework contains a graph embedding module, a multivariate decoding module, and a pretraining module. The graph embedding module first uses a two-dimensional convolutional neural network to extract time-series features from electric vehicle charging data. It then uses a graph attention network with adaptive attention weighting to capture relationships among charging nodes. A two-layer graph attention structure is used to extract and fuse spatio-temporal features. Weighted aggregation of multiple attention heads is used instead of fixed aggregation, and skip connections help preserve information from earlier layers.
The study used an urban electric vehicle charging dataset covering 247 traffic zones, with a five-minute sampling interval and 8,640 timestamps. The data included charger occupancy, node relationships, longitude and latitude, and charging price. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. The model used mean square error as the loss function and was tested at 15-minute, 30-minute, 45-minute, and 60-minute horizons. Reported training settings included a batch size of 512, 200 pretraining epochs, 1,000 fine-tuning epochs, the Adam optimizer, a weight decay of 0.00001, and a learning rate of 0.001.
4.2 Bidirectional Decoding and Pretraining
After graph embedding, the multivariate decoding module uses a bidirectional gated recurrent unit with attention to process the embedded spatio-temporal features. The bidirectional structure captures both forward and backward dependencies in the electric vehicle charging demand sequence. The attention mechanism then assigns different weights to hidden states, allowing the model to emphasize important time points. A pretraining module uses generated tuning samples and a meta-learning strategy to improve model initialization and generalization. The pretraining process is designed to help the model adapt to price fluctuations and neighbor-area effects in electric vehicle charging demand.
4.3 Reported Electric Vehicle Demand Forecasting Results
The study compared the proposed method with LSTM, GCN, GAT, STGCN, DCRNN, and AST-GAT. The reported 15-minute results are shown below.
| Model | RMSE (10-2) | MAPE (%) | RAE (10-2) | MAE (10-2) |
|---|---|---|---|---|
| LSTM | 3.59 | 8.91 | 10.66 | 1.90 |
| GCN | 4.20 | 40.28 | 35.70 | 6.35 |
| GAT | 3.39 | 10.24 | 11.30 | 2.01 |
| STGCN | 3.55 | 31.41 | 28.06 | 4.99 |
| DCRNN | 3.72 | 11.32 | 12.80 | 2.28 |
| AST-GAT | 3.41 | 9.93 | 11.39 | 2.03 |
| Proposed method | 2.92 | 8.12 | 9.74 | 1.36 |
The reported 60-minute results are shown below. They indicate that the proposed electric vehicle charging demand method maintained an advantage as the forecasting horizon lengthened.
| Model | RMSE (10-2) | MAPE (%) | RAE (10-2) | MAE (10-2) |
|---|---|---|---|---|
| LSTM | 9.01 | 24.15 | 29.07 | 6.36 |
| GCN | 8.06 | 52.03 | 45.69 | 9.05 |
| GAT | 8.89 | 29.38 | 29.57 | 5.35 |
| STGCN | 8.74 | 49.62 | 40.45 | 7.34 |
| DCRNN | 7.82 | 24.68 | 28.94 | 5.13 |
| AST-GAT | 7.54 | 27.69 | 28.74 | 4.62 |
| Proposed method | 6.27 | 22.20 | 25.42 | 3.54 |
The average results across the 15-minute, 30-minute, 45-minute, and 60-minute horizons are summarized below.
| Model | Average RMSE (10-2) | Average MAPE (%) | Average RAE (10-2) | Average MAE (10-2) |
|---|---|---|---|---|
| LSTM | 6.70 | 17.25 | 21.16 | 4.06 |
| GCN | 6.76 | 46.95 | 42.08 | 7.72 |
| GAT | 6.36 | 20.00 | 21.08 | 3.77 |
| STGCN | 6.20 | 39.17 | 34.30 | 6.14 |
| DCRNN | 5.93 | 18.00 | 21.13 | 3.75 |
| AST-GAT | 5.73 | 18.34 | 20.28 | 3.49 |
| Proposed method | 4.74 | 15.30 | 18.11 | 2.52 |
The ablation study reported that removing the graph embedding module, the multivariate decoding module, or the pretraining module increased forecasting errors at all tested horizons. The largest increases appeared when the graph embedding or multivariate decoding module was removed. This suggests that spatial relationship modeling and temporal decoding both play important roles in electric vehicle charging demand forecasting. The study argues that electric vehicle charging demand is shaped by both time-of-day patterns and spatial dependencies among charging areas, and that a model must integrate both to perform well.
5. Multi-Objective Electric Vehicle Charging-Discharging Optimization
The third part of the study moves from forecasting to operation. The research proposes a multi-objective electric vehicle charging-discharging strategy that considers both user-side interests and grid-side interests. On the user side, the strategy aims to minimize a comprehensive cost that includes charging cost, battery consumption cost, electric vehicle discharge revenue, and discharge time revenue. On the grid side, the strategy aims to minimize peak-valley difference and load fluctuation. The strategy is formulated as a mixed-integer programming problem and solved with a combination of linear programming, the simplex method, and branch-and-bound search.
5.1 User-Side Cost and Grid-Side Objectives
The user-side objective is designed to encourage electric vehicle owners to participate in charging and discharging. Charging cost is based on charging power and charging price. Battery consumption cost is modeled with a battery degradation rate and battery replacement cost. Discharge revenue is based on discharging power and discharge price. Discharge time revenue compensates electric vehicle users for the time they make their vehicles available for discharging. The grid-side objectives include minimizing the peak-valley difference of the load curve and minimizing load fluctuation through variance.
The study used time-of-use pricing to guide electric vehicle charging and discharging. The reported charging and discharging prices are shown below.
| Period | Charging price | Discharging price |
|---|---|---|
| Peak period | 1.2 | 1.6 |
| Flat period | 0.8 | 0.4 |
| Valley period | 0.34 | 0.3 |
The optimization model includes constraints on charger capacity, charging and discharging power, battery energy, state of charge, user travel needs, and charging time. Charger capacity is linked to the electric vehicle charging demand at each time step, so the strategy does not assume unlimited charging capacity. Battery state of charge is constrained between minimum and maximum levels. The expected energy requirement of each electric vehicle is also included so that the strategy does not sacrifice user mobility for grid benefits. The reported parameter settings included a minimum state of charge of 0.1, a maximum state of charge of 0.9, a charging and discharging efficiency of 0.95, and a battery capacity of 64. The weights for user cost, peak-valley difference, and load fluctuation were 0.5, 0.25, and 0.25, respectively.
5.2 Constraints and Solution Method
The solution process relaxes the binary connection variable into a continuous variable in the interval from zero to one and solves the resulting linear programming problem with the simplex method. It then applies branch-and-bound to recover an integer-feasible solution. The study describes the lower bound, upper bound, branching, pruning, and termination logic used in the search. The overall aim is to find a charging-discharging schedule that satisfies electric vehicle charging demand while reducing user cost and improving the grid load profile.
5.3 Reported Electric Vehicle Optimization Results
The study compared the proposed strategy with three alternatives. Strategy 1 used uncoordinated charging with no discharging. Strategy 2 used uncoordinated charging and discharging. Strategy 3 used time-of-use pricing for charging and discharging. The reported optimization results are shown below.
| Strategy | Cost (yuan) | Peak-valley difference (kW) | Load fluctuation (kW2) |
|---|---|---|---|
| Strategy 1 | 207.5 | 95 | 20529.33 |
| Strategy 2 | 25.7 | 117 | 24609.96 |
| Strategy 3 | 42.63 | 57 | 6835.64 |
| Proposed strategy | 1.57 | 57 | 4414.76 |
The reported results show that the proposed electric vehicle charging-discharging strategy achieved the lowest cost, the lowest load fluctuation, and a peak-valley difference equal to the best of the comparison strategies. Strategy 1 had the highest cost, reflecting the inefficiency of uncoordinated charging and the absence of discharge revenue. Strategy 2 reduced cost relative to Strategy 1 but produced the highest peak-valley difference and the highest load fluctuation, showing that random charging and discharging can increase grid stress. Strategy 3 improved the peak-valley difference relative to Strategy 1 and Strategy 2, but its cost and load fluctuation remained higher than the proposed strategy.
The study interprets these results as evidence that electric vehicle charging-discharging optimization should consider both user economics and grid operation. A strategy that only minimizes cost may create new load peaks. A strategy that only smooths the grid may discourage electric vehicle users from participating. The multi-objective approach attempts to balance these concerns by combining charging cost, battery degradation, discharge revenue, discharge time compensation, peak-valley difference, and load fluctuation in one scheduling framework.
6. What the Findings Mean for Electric Vehicle Charging Infrastructure
For electric vehicle charging infrastructure planning, the research suggests that forecasting and optimization should not be separated. Electric vehicle charging load forecasting can inform how much capacity is needed and when. Electric vehicle charging demand forecasting can inform where capacity is needed. Charging-discharging optimization can inform how existing capacity can be used more flexibly. Together, these tools can support a more coordinated approach to electric vehicle grid integration.
- Electric vehicle charging load forecasting can help grid operators anticipate peak pressure and plan reserve capacity.
- Electric vehicle charging demand forecasting can help planners identify spatial hotspots and allocate charging resources more efficiently.
- Electric vehicle charging-discharging optimization can help shift electric vehicle charging to lower-load periods and use electric vehicle batteries as flexible storage.
- Time-of-use pricing and discharge compensation can encourage electric vehicle users to participate in grid support.
- Battery degradation and travel needs must remain part of electric vehicle scheduling so that grid benefits do not come at the expense of user mobility.
The research also highlights the role of data quality. The electric vehicle load forecasting study used charging records and weather data. The electric vehicle demand forecasting study used charger occupancy, charging price, geographic relationships, and node adjacency. The optimization study used time-of-use prices and battery constraints. In each case, the quality and completeness of electric vehicle data directly affect the usefulness of the result. The study therefore supports broader data collection across regions, vehicle types, and charging scenarios.
7. Limitations and Next Steps for Electric Vehicle Research
The research acknowledges several limitations. The electric vehicle charging load forecasting study used a specific charging dataset and weather dataset, and electric vehicle charging behavior can vary across cities, seasons, and vehicle types. Future work could collect electric vehicle charging data from different regions and from passenger cars, commercial vehicles, and logistics vehicles to test whether the method remains effective across a wider range of electric vehicle use cases. The electric vehicle charging demand forecasting study used an urban dataset with traffic zones and charging prices, but future models could incorporate traffic flow, socioeconomic factors, and other cross-domain data to capture more dimensions of electric vehicle charging demand.
The charging-discharging optimization study used a simulation framework with time-of-use prices and battery constraints. Future work could incorporate real-time grid operation data, battery health monitoring, and user travel trajectories to build a more detailed electric vehicle scheduling model. Environmental benefits could also be added to the optimization framework by considering how electric vehicle charging and discharging affect renewable energy use and carbon emissions. Field pilots with grid operators, charging operators, and electric vehicle users would help validate the strategy in real conditions.
8. Conclusion
The research presents an integrated view of electric vehicle grid integration. Accurate electric vehicle charging load forecasting provides the foundation for grid planning. Spatio-temporal electric vehicle charging demand forecasting provides the spatial intelligence needed for charging infrastructure management. Multi-objective electric vehicle charging-discharging optimization provides the operational flexibility needed to turn electric vehicle charging from a passive load into an active grid resource. The reported experiments show improved forecasting performance and a simulation-based strategy that reduces user cost and load fluctuation while maintaining a favorable peak-valley difference.
For a power system preparing for wider electric vehicle adoption, the broader lesson is clear: electric vehicle charging should be measured, modeled, and managed as a dynamic process. Better electric vehicle load forecasts can reduce uncertainty. Better electric vehicle demand forecasts can improve placement and operation of charging facilities. Better electric vehicle charging-discharging strategies can align user behavior with grid needs. As electric vehicle penetration continues to rise, these capabilities will become increasingly important for stable, efficient, and cleaner energy systems.