A professional master’s thesis has proposed an integrated framework for electric vehicle charging load prediction and scheduling by focusing on the spatio-temporal distribution characteristics of electric vehicle charging demand. The research addresses two connected challenges: the difficulty of extracting spatio-temporal features from electric vehicle charging load data with high accuracy, and the operational pressure that large-scale electric vehicle integration places on distribution networks, including wider load peak-valley gaps, higher network losses, and rising operating costs. The work combines a graph convolutional network with a Transformer model for electric vehicle charging load prediction and develops a spatio-temporal adaptive scheduling strategy for electric vehicle charging and discharging. The study reports validation using electric vehicle charging station data from a city in China and an IEEE 33-node distribution network model.
The research is set against the continued expansion of electric vehicle ownership and the rapid growth of charging infrastructure. According to the study, China’s new energy vehicle sales reached 7.03 million in 2024, charging infrastructure increased by 3.28 million units, and cumulative charging infrastructure reached 11.884 million units. The national new energy vehicle fleet reached 31.4 million, including 22.09 million battery electric vehicles. The study also cites a forecast that new energy vehicle sales may exceed 13 million in 2025. These figures underline the scale of electric vehicle growth and the importance of accurate electric vehicle charging load prediction and coordinated electric vehicle scheduling for power system stability and resource allocation.
1. Industry Context and the Need for Electric Vehicle Load Prediction
The study places electric vehicle development within the broader context of energy transition and carbon reduction. It notes that fossil fuel use is a major source of carbon emissions and that transport accounts for about 10 percent of carbon emissions. China’s external dependence on oil is reported at 73 percent, and the research states that developing new energy vehicles is widely regarded as a way to reduce fossil energy depletion and environmental pollution. The study also refers to the global climate goal of limiting temperature rise to within 2 degrees Celsius before 2025 and to China’s “2030 carbon peak, 2060 carbon neutrality” planning targets. In this setting, electric vehicle adoption is presented not only as a transport transition but also as part of a wider effort to improve energy structure and support low-carbon development.
At the same time, the research emphasizes that electric vehicle charging demand can create substantial challenges for power grids. The study reports that in 2023 electric vehicle charging demand accounted for 15 percent to 25 percent of residential electricity load in China, and in first-tier cities the share reached about 30 percent. During evening peak periods, grid load rates exceeded 80 percent, 20 percent higher than in the fuel vehicle era. In highway and holiday scenarios, charging queues were longer, and single-day charging volume reached five times the level of ordinary days. Because electric vehicle user charging behavior has high spatio-temporal randomness and clustering effects, regional grid load peak-valley differences continue to expand. The study states that daily peak-valley rates in some urban distribution networks have exceeded 35 percent, and voltage violation risk has increased by 2 to 3 times, threatening power system safety margins.
The research also examines the economic dimension of electric vehicle integration. Random fluctuations in large-scale electric vehicle charging load can cause non-linear changes in grid operating costs. Under a traditional “supply follows demand” scheduling model, an additional 15 percent to 20 percent of spinning reserve capacity may be needed to control electric load peak-valley differences. The study further notes that existing vehicle-grid interaction technology is constrained by communication delay, control accuracy, and user acceptance, and that the economic value utilization rate of vehicle-grid interaction is less than 30 percent. Because business models are limited and benefit distribution is uneven, vehicle-grid interaction benefits are often captured by aggregators, while users may find that benefits do not cover battery degradation risk. This reduces user willingness to participate. These points frame the need for both better electric vehicle charging load prediction and more effective electric vehicle scheduling.
2. Spatio-Temporal Characteristics of Electric Vehicle Charging Loads
The research analyzes electric vehicle charging load characteristics from time, space, user behavior, and environmental perspectives. In the time dimension, electric vehicle charging load shows clear daily periodicity. On working days, charging demand often forms morning and evening peaks, while on non-working days the pattern may shift to a midday or night single peak. The study classifies date attributes into working days, weekends, and statutory holidays. Working-day charging facility load intensity is generally higher than on weekends because users travel regularly between residential and workplace areas. During holidays, long-distance travel increases battery energy consumption and creates concentrated charging demand. Night-time charging is also influenced by low electricity prices, which can produce a high night load peak. The study notes that peak timing may shift and fluctuate across regions because of differences in development levels and other factors.
In the spatial dimension, electric vehicle charging load is non-uniformly distributed. The research divides urban transport network areas by function into residential areas, commercial areas, and work areas. Charging load distribution is closely related to the geographic location of urban functional zones. Residential areas mainly have night-time slow charging demand, with load density concentrated from 22:00 to 06:00. Commercial areas have fast charging demand, and midday load density can be much higher than in residential areas. The study also reports differences in charging pile utilization: utilization at urban center charging piles can exceed 90 percent, while charging facilities far from city centers may have low utilization. These spatial differences affect prediction, charging station planning, and distribution network expansion costs.
User behavior and environmental factors further shape electric vehicle charging load spatio-temporal distribution. The study distinguishes private cars from taxis and buses. Taxis and buses have regular charging characteristics because of fixed operating schedules, while private cars show greater randomness. Private electric vehicle users are further divided into three groups: intercity commuter users, transit intermittent users, and urban resident users. Intercity commuter users show high-frequency and high-capacity charging demand on working days and often choose intercity fast charging nodes. Transit intermittent users account for less than 15 percent of total users and show low-frequency, low-capacity charging behavior, with charging point selection highly related to transport network topology. Urban resident users show clear spatio-temporal coupling, and their charging demand peaks often appear around commuting periods. Temperature and weather also affect electric vehicle charging demand. Extreme high or low temperatures can increase battery energy consumption through cabin temperature control, shortening the charging demand cycle and increasing charging behavior density. Weather conditions such as rain, dust, ice, and snow can also change travel mode choice and electric vehicle use intensity.
3. Prediction Model Combining Graph Convolutional Network and Transformer
To address the limitations of conventional electric vehicle charging load prediction methods, the research proposes a prediction model that combines a graph convolutional network with a Transformer. The model is designed to extract spatial features among electric vehicle charging stations and to capture long-sequence temporal dependencies in electric vehicle charging load data. The overall architecture includes an input layer, a graph convolutional network module for spatial feature extraction, a Transformer module for temporal modeling, and an output layer. The research states that charging stations are mapped as graph nodes in a transport network. Topological connections are established based on road connectivity and geographic proximity, and multi-dimensional station features are fused to build a graph-structured dataset. The graph convolutional network extracts spatial features, while explicit positional encoding and a gated fusion strategy support spatio-temporal collaborative modeling. The Transformer’s global attention mechanism is used to address long-sequence dependency problems and improve load prediction efficiency.
The graph construction in the study defines a graph with charging stations as nodes and edges representing spatial association. When the Haversine distance between two charging stations is less than 2.5 km, an edge is established, and the edge weight is set according to the inverse of the distance. Otherwise, the weight is zero. The node feature matrix includes historical electric vehicle charging load values and environmental features such as temperature, humidity, wind speed, and rainfall. The study uses electric vehicle charging station data sampled every 15 minutes. The dataset includes 15 charging stations, and weather data are obtained through the Open-Meteo historical weather API by configuring geographic coordinates and time ranges. Data preprocessing includes format unification, time calibration, removal of clearly erroneous entries, and normalization. The research uses Min-Max normalization and Z-score standardization to improve data stability and help the deep learning model capture differences among features.
| Feature input | Considered factor | Feature description |
|---|---|---|
| Charging station name | Node matrix number in feature matrix | Spatial feature |
| Charging station address | Longitude and latitude coordinates for adjacency matrix construction | Spatial feature |
| Historical charging station load | Time-series historical load of each charging station | Temporal feature |
| Temperature | Temperature in the area of the charging station | External factor feature |
| Humidity | Humidity in the area of the charging station | External factor feature |
| Rainfall | Rainfall in the area of the charging station | External factor feature |
| Wind speed | Wind speed in the area of the charging station | External factor feature |
| Date type | Working day or holiday identification | User travel feature |
The graph convolutional network module first normalizes the adjacency matrix using a symmetric normalized Laplacian matrix to improve training stability and spatial feature aggregation efficiency. It then performs convolution operations over the graph. The Transformer module uses an encoder-decoder structure. In the input layer, positional encoding injects temporal information. A gated spatio-temporal fusion mechanism combines the encoded temporal features with the spatial feature matrix extracted by the graph convolutional network. The encoder uses multi-head attention, a feed-forward neural network, residual connections, and layer normalization to extract higher-order spatio-temporal features. The decoder includes a masked multi-head attention layer to prevent future information leakage and to generate prediction sequences step by step. To adapt the Transformer to continuous electric vehicle charging load prediction, the output layer replaces the traditional SoftMax function with a Sigmoid function, mapping hidden states to the interval from 0 to 1 and suppressing the influence of abnormal values. Inverse normalization then produces the final prediction result.
The experimental environment used Windows 11 64-bit, Python 3.9.2, TensorFlow 2.6, an NVIDIA RTX 3060 6G graphics card, and a 2.30 GHz Intel i7 processor with 14 cores and 16.0 GB memory. The graph convolutional network consisted of three layers. The Transformer module used six encoder-decoder layers with four parallel attention heads. The batch size was 48, dropout was set to 0.1, the Adam optimizer was used with an initial learning rate of 0.001, and training continued for 300 epochs. The loss function was mean squared error. The dataset was divided into 80 percent for training and 20 percent for validation.
4. Prediction Accuracy and Spatial Feature Analysis
The research compares the proposed graph convolutional network and Transformer model with BP, LSTM, and graph convolutional network models. The evaluation uses root mean square error, mean absolute percentage error, mean absolute error, and the coefficient of determination. The reported results show that the combined model achieves the best performance across the four indicators. The root mean square error is 40.657, the mean absolute percentage error is 2.413, the mean absolute error is 28.473, and the coefficient of determination is 0.9260. Compared with the BP model, the root mean square error is reduced by 56.9 percent. Compared with the LSTM model, the mean absolute percentage error is reduced by 71 percent. Compared with the graph convolutional network model, the mean absolute error is reduced by 48.5 percent. The coefficient of determination is 5.3 percent higher than that of the BP model. The study states that these results confirm the model’s strong representation ability for non-linear relationships, robustness to time-series fluctuations, and capacity for spatial feature extraction.
| Prediction model | RMSE | MAPE | MAE (kW) | R2 |
|---|---|---|---|---|
| BP | 94.292 | 16.598 | 80.378 | 0.881 |
| LSTM | 55.328 | 8.325 | 45.363 | 0.855 |
| GCN | 64.371 | 13.753 | 55.235 | 0.823 |
| GCN-Transformer | 40.657 | 2.413 | 28.473 | 0.9260 |
The research also analyzes the spatial distribution of electric vehicle charging station use. A heat distribution map of urban charging stations shows that charging stations are more densely located in the city center, and intensity radiates outward along transport corridors. Based on this distribution, the study groups the 15 charging stations into functional areas. Stations 1, 2, 4, 13, and 14 are identified as commercial area stations. Stations 3, 5, 6, and 11 are identified as residential area stations. Stations 7, 8, 9, and 10 are identified as industrial area stations. Station 12 is identified as a highway charging station. The research states that this grouping is consistent with the urban transport network structure.
For daily electric vehicle charging load prediction across the 15 stations, the study finds clear differences in time distribution among charging stations. Stations 1, 2, and 4 show similar temporal trends. Their morning peak from 9:00 to 12:00 and evening peak from 18:00 to 21:00 account for 37 percent and 46 percent of charging demand, respectively. The study links this fluctuation to user commuting behavior and notes that these stations are located in transport hubs and commercial areas, where continuous charging demand is high during working hours. Stations 3, 5, and 6 show similar trends with higher night-time charging load, reflecting concentrated night charging in residential areas. Station 5, located at the intersection of commercial and residential areas, shows both commercial-area daytime demand and residential-area night demand. Other stations in industrial areas show relatively flat charging load and lower peak periods compared with commercial and residential stations, while remaining in working state for longer periods.
The study further tests the influence of spatial differences on prediction results by changing the number of nodes in the adjacency matrix. When nodes are removed, the spatial feature extraction among charging stations is reduced. The results show that reducing the number of nodes increases prediction error for each station, and the more nodes removed, the larger the prediction error. This indicates that spatial distribution relationships among stations are important for electric vehicle charging load prediction and that removing nodes weakens the association and influence among charging stations. The research also compares residuals among BP, graph convolutional network, LSTM, and the combined graph convolutional network and Transformer model. The BP residual plot shows a systematic high-value bias, the graph convolutional network performs well in low-value regions, and the LSTM model shows flattened clustering in the middle range with predicted values lower than actual values. The combined model shows the most concentrated residuals and remains stable in extreme value regions, consistent with the modeling advantage of global attention for complex dependencies.
5. Scheduling Strategy for Large-Scale Electric Vehicle Integration
Building on the electric vehicle charging load prediction analysis, the research proposes a scheduling strategy that considers spatio-temporal distribution characteristics. The strategy aims to address increased load peak-valley differences, higher network losses, and rising operating costs caused by large-scale electric vehicle access to distribution networks. The scheduling model considers user behavior spatio-temporal characteristics, the transport network, and distribution network factors. It constructs an electric vehicle scheduling spatio-temporal distribution model and formulates a multi-objective optimization problem that includes distribution network operating cost, load peak-valley smoothing, network loss reduction, and transport energy consumption. The study states that conventional scheduling strategies struggle to respond effectively to spatio-temporal dynamic changes, so the proposed strategy seeks a multi-objective balance among grid operation cost minimization, peak-valley smoothing, and network loss reduction.
In the user behavior model, the research uses a normal distribution to describe single-day travel distance for electric vehicle users. The study reports that 68.3 percent of users have travel distances within the mean plus or minus one standard deviation, and 95.4 percent are within the mean plus or minus two standard deviations. Electric vehicles are idle about 95 percent of the day, so idle time can be used for vehicle-to-grid discharge to help reduce distribution network losses. The model assumes that a proportion of electric vehicles connected to the grid agree to participate in vehicle-to-grid operation. When an electric vehicle is connected, the user submits an expected off-grid time and an expected battery state of charge at off-grid time. The electric vehicle then becomes dispatchable. The scheduling model ensures that the battery state of charge reaches the user-set expected value at the off-grid time while adjusting charging and discharging power to meet distribution network needs. State of charge is constrained between minimum and maximum values, and charging and discharging power are constrained by upper and lower limits.
The transport energy model considers road network conditions and calculates the shortest path from the current position to a charging station using Dijkstra’s algorithm. The transport network is represented as a graph with road intersections as nodes and road sections as edges. Dynamic weights are calculated from road section length and time-dependent road speed. The study classifies urban roads into four levels: expressways, arterial roads, secondary roads, and branch roads. Different road levels have different unit-mileage energy consumption under different traffic conditions. The model also includes a constraint that the remaining electric vehicle battery energy must be sufficient to support the travel distance. Transport energy cost is related to electricity price, and the distribution network pays a subsidy for electric vehicle travel to charging stations during scheduling. The study provides a time-of-use electricity price table with peak, flat, and valley periods.
| Period type | Time range | Electricity price (yuan/kWh) |
|---|---|---|
| Peak period | 08:00-12:00, 18:00-22:00 | 1.2 |
| Flat period | 12:00-18:00 | 0.8 |
| Valley period | 22:00-08:00 | 0.4 |
The distribution network model describes the power balance at each node during electric vehicle scheduling. The equation includes source injection, load consumption, line transmission, and electric vehicle charging or discharging power. When an electric vehicle is charging, its node power is positive; when discharging, its node power is negative. The model uses power flow equations to represent active and reactive power flow and calculates active network loss from current squared and resistance. The study also defines the load peak-valley difference before and after electric vehicle scheduling. Because electric vehicle charging load is coupled with the transport network and distribution network, the load is variable in space and time. The research introduces a continuous power flow model that accounts for electric vehicle behavior characteristics, in which the load parameter at a node is related to the number of connected electric vehicles.
6. Multi-Objective Optimization and Improved Particle Swarm Algorithm
The scheduling model formulates a multi-objective optimization problem. The total objective combines load peak-valley penalty cost, network loss cost, and transport energy cost, with priority coefficients. In the time dimension, a penalty cost is associated with the load peak-valley difference, and scheduling electric vehicles reduces this difference to obtain benefits. In the spatial dimension, network loss cost is calculated from the reduction in distribution network losses after electric vehicle access. From the spatio-temporal coupling perspective, transport energy consumption cost is calculated from electric vehicle travel distance and the number of electric vehicles participating in scheduling. The constraints include node voltage limits, branch flow limits, a voltage stability margin, time-of-use electricity price coordination, charging and discharging period coordination, user off-grid energy demand, and transport network energy consumption limits. The study states that a 5 percent voltage stability margin is reserved to respond to load fluctuations.
To unify dimensions and allocate weights among objective functions, the research applies multi-objective vector normalization. A decision matrix is constructed from three indicators: the change in load peak-valley penalty cost, total network loss energy, and total transport energy consumption distance. The matrix is normalized using range normalization. Entropy divergence coefficients are calculated, and entropy weights are dynamically allocated with spatio-temporal correction factors. The final objective function combines the normalized indicators with dynamic weights. In this structure, the first weighted term represents priority for reducing the load peak-valley difference, the second weighted term represents secondary reduction of distribution network losses, and the third weighted term represents control of transport energy consumption.
The research improves the particle swarm optimization algorithm to solve the spatio-temporal collaborative optimal scheduling objective. Standard particle swarm optimization updates particle velocity and position using inertia weight, individual best position, and global best position. The improved algorithm introduces a time-adaptive inertia weight that adjusts according to scheduling period characteristics. During peak periods, the inertia weight adjustment is increased to improve global search capability and respond to load fluctuations. During valley periods, the adjustment is reduced to accelerate convergence to the optimal solution. The algorithm also introduces space-adaptive learning factors that are adjusted according to distribution network node load distribution. At high-load nodes, one learning factor is increased to emphasize individual historical best experience; at low-load nodes, another learning factor is increased to emphasize group best experience.
To reduce the risk of local optima, the research introduces Tent chaotic mapping to generate the initial population, ensuring that particles cover the solution space uniformly and improving exploration efficiency for spatio-temporal coupled solutions. Cauchy mutation is also introduced to help the algorithm escape local optima and prevent premature convergence. The mutation intensity coefficient decreases with iteration number, and the standard Cauchy distribution random number has long-tail characteristics that promote large-scale disturbance. The study also describes a fuzzy adjustment process in which clear data are converted into fuzzy sets and then defuzzified into clear data, with normalized fitness values used as input variables between 0 and 1. These mechanisms are combined in the improved spatio-temporal adaptive particle swarm algorithm.
7. Case Study and Scheduling Results
The case study uses the IEEE 33-node distribution network system, located within an urban transport network topology. The study identifies transport network nodes 1 to 5 and 7 to 10 as residential areas, nodes 14, 15 to 18, and 20 to 24 as commercial areas, and nodes 6, 11, 12, 13, 19, 25, and 26 as industrial areas. The simulation includes 50 charging piles supporting 200 electric vehicles. These charging piles are distributed across selected key nodes with different charging types and power configurations. The research designs three scenarios. Scenario 1 includes 200 electric vehicles, does not consider transport energy consumption, and uses particle swarm optimization. Scenario 2 includes 200 electric vehicles, considers transport energy consumption, and uses particle swarm optimization. Scenario 3 includes 200 electric vehicles, considers transport energy consumption, and uses the improved spatio-temporal adaptive particle swarm optimization algorithm. Scenario 1 serves as the base experiment for comparison with the other scenarios.
| Transport node number | Area type | Charging pile type | Number of charging piles | Power configuration |
|---|---|---|---|---|
| 3 | Residential area | Slow charging pile | 15 | 7 kW |
| 9 | Residential area | Slow charging pile | 15 | 7 kW |
| 6 | Industrial area | Fast charging pile | 10 | 120 kW |
| 11 | Transport hub | Ultra-fast charging pile | 5 | 250 kW |
| 14 | Transport hub | Ultra-fast charging pile | 5 | 250 kW |
| 15 | Commercial area | Fast charging pile | 10 | 120 kW |
In the voltage deviation analysis, the research uses distribution network node 26 as an example. Under scenario 1, the voltage deviation range reaches -5.0 percent to -2.0 percent and exceeds the lower limit. Under scenario 3, the voltage deviation range is -3.0 percent to -1.5 percent and remains within the stable allowed range. The study states that electric vehicle spatio-temporal charging and discharging behavior effectively balances distribution network power distribution and alleviates voltage violation problems in load-concentrated areas. The result indicates that the improved scheduling strategy can better support voltage security during electric vehicle charging and discharging.
The research also analyzes network loss with transport energy consumption considered. Before scheduling, the network loss peak at node 6 appears at 17:00 and is directly related to high load current on its main line. After scheduling, Dijkstra’s algorithm is used to plan the shortest-time path, reducing the selection of long-distance, high-impedance charging paths. The network loss peak falls to 8.7 kW, a reduction of 30.4 percent. The study states that this result verifies the effectiveness of the transport energy consumption constraint and shows the coordinated effect between the transport energy model and power grid power flow.
| Scenario | Peak-valley difference before scheduling (kW) | Peak-valley difference after scheduling (kW) | Network loss (kW) | Grid cost (yuan) | Transport cost (yuan) | Total cost (yuan) |
|---|---|---|---|---|---|---|
| Scenario 1 | 2853.39 | 1508.42 | 35.20 | 68255.25 | 0 | 68255.25 |
| Scenario 2 | 2853.39 | 1303.84 | 28.84 | 62315.63 | 4212.43 | 66528.06 |
| Scenario 3 | 2853.39 | 1169.73 | 21.76 | 55828.97 | 4803.50 | 60632.47 |
The scenario comparison shows the role of spatio-temporal adaptation. In scenario 1, only grid-side objectives are optimized. The peak-valley difference is reduced by 22 percent, but transport energy consumption is ignored, and improvements in network loss and voltage violations are limited. In scenario 2, path planning reduces ineffective charging paths, and the peak-valley difference is further reduced by 41.8 percent. However, algorithm convergence is constrained by static parameters. In scenario 3, dynamic inertia weight adjustment and Cauchy mutation are introduced. The peak-valley difference reduction reaches 59 percent, and the distribution network total cost is the lowest among the three scenarios. The research states that this achieves multi-objective collaborative optimization.
The study also examines different electric vehicle penetration ratios. As the electric vehicle ratio increases from 15 percent to 25 percent, user participation rises, and the peak-valley difference reduction rate improves, making the load more balanced. When the electric vehicle ratio increases to 35 percent, the peak-valley difference reduction rate instead declines, because a high proportion of electric vehicles creates new load concentration. The reported results are as follows.
| Electric vehicle ratio | User participation rate | Peak-valley difference reduction rate | Total cost (yuan) |
|---|---|---|---|
| 15% | 55% | 48% | 65820 |
| 25% | 65% | 59% | 60632 |
| 35% | 70% | 52% | 64210 |
The scheduling results also show improved spatio-temporal balance of total system load. The total load curve changes from an obvious dual-peak characteristic before scheduling to a smoother distribution after scheduling. The study presents the spatio-temporal distribution of charging and discharging load, where positive values represent electric vehicle charging and negative values represent electric vehicle discharging. This distribution reflects how electric vehicle scheduling can shift charging demand across time and space while considering user travel needs and distribution network constraints. The research states that the proposed strategy can effectively reduce load peak-valley differences, lower distribution network losses, and reduce total cost while satisfying node voltage deviation rate constraints.
8. Implications for Electric Vehicle and Grid Coordination
The research has implications for electric vehicle charging infrastructure planning, distribution network operation, and transport-power system coordination. Accurate electric vehicle charging load prediction can help reduce fluctuations caused by random user charging behavior and support interaction between electric vehicle charging piles and the power grid. Although electric vehicle charging load has randomness, large data sets and similar time points can reveal statistical patterns that support future electric vehicle charging load prediction. Improved electric vehicle charging load prediction can also support charging facility planning, user experience improvement, and scheduling algorithm design. The study emphasizes that incorporating spatio-temporal distribution characteristics into electric vehicle charging load prediction is not only a way to improve accuracy but also a foundation for optimization scheduling and charging infrastructure planning.
For electric vehicle scheduling, the research states that coordinated scheduling is important for the joint development of the power system and transport system. Effective scheduling strategies can relieve peak-valley pressure and support dynamic balance between supply and demand. They can also reduce electric vehicle user costs and comprehensive grid costs while improving coordination efficiency between the grid and the transport network. The study links electric vehicle scheduling with vehicle-to-grid technology, describing how electric vehicles can serve as distributed energy storage resources. Electric vehicles have advantages such as high spatio-temporal flexibility, large aggregate capacity, and agile response mechanisms. Through vehicle-grid interaction, electric vehicle scheduling can support peak shaving, valley filling, voltage regulation, and renewable energy consumption. These capabilities are presented as important support for a new type of power system.
The research also notes that electric vehicle scheduling involves multiple stakeholders. Optimization models for ordered electric vehicle charging generally consider the distribution network, users, and charging stations. Some studies focus on minimizing grid load peak-valley difference and total user charging cost. Others consider user interests alone, such as minimizing total charging and discharging cost and maximizing remaining electric vehicle battery energy after charging and discharging. Additional research includes charging station interests and seeks balanced benefits among the distribution network, charging stations, and users. The study states that because electric vehicle charging load has spatio-temporal distribution characteristics, scheduling objectives should be built on distribution network security and stability, and optimization should combine transport network and distribution network analysis. Algorithm design should also avoid premature convergence and local optima in traditional methods.
9. Conclusion and Future Research Directions
The research concludes that electric vehicle charging load spatio-temporal distribution characteristics have a growing influence on grid operation and energy scheduling. In the time dimension, periodicity affects power system security and stability. In the space dimension, significant differences affect grid capacity and charging station planning. Prediction and scheduling should therefore not rely on single-variable models but should fully exploit electric vehicle charging load spatio-temporal features. The proposed graph convolutional network and Transformer model achieves improved prediction performance by extracting spatial features among electric vehicle charging stations and modeling temporal dependencies. The proposed scheduling strategy improves the particle swarm optimization algorithm with spatio-temporal adaptive weights, chaotic mapping, and Cauchy mutation to enhance global search capability and convergence speed in complex spatio-temporal scenarios.
The study reports that the combined prediction model achieves advantages in root mean square error, mean absolute percentage error, mean absolute error, and coefficient of determination. The improved scheduling strategy reduces the peak-valley difference by 59 percent, lowers network loss, reduces the number of voltage violation nodes, and saves 11 percent of distribution network total cost in the reported scenario comparison. The research states that these results verify the effectiveness of the proposed electric vehicle charging load prediction and scheduling strategy. The work also identifies future research directions. One direction is to include other mobile energy storage resources, such as hydrogen electric vehicles and mobile energy storage vehicles, in electric vehicle charging load prediction and scheduling. Another direction is to incorporate real-time transport network data and use multi-source data fusion and dynamic modeling to improve spatio-temporal prediction resolution and timeliness. A further direction is to expand the scheduling model from grid-side economic optimization to multi-stakeholder coordination, including electric vehicle user charging cost minimization and quantitative assessment of battery life degradation, so as to build a mutually beneficial scheduling model for electric vehicle integration.