The rapid growth of electric vehicle ownership and their integration into power grids have introduced significant challenges for the planning and operation of modern energy systems. As a key player in the global transition to sustainable transportation, China EV market expansion necessitates accurate forecasting of charging demands to ensure grid stability and efficient resource allocation. In this study, I propose a data-driven approach for spatio-temporal load forecasting of electric vehicle charging demand based on a hybrid gated recurrent unit and long-short-term memory neural network model. This methodology addresses the limitations of traditional model-driven approaches by leveraging user characteristic data to predict travel patterns, which are then mapped onto actual road networks to estimate node-specific charging loads.
The foundation of this research lies in processing and analyzing large-scale travel survey data to establish correlations between user attributes and mobility behaviors. The National Household Travel Survey dataset serves as the primary source, containing diverse user state information that reflects typical driving patterns. After preprocessing to handle missing values, outliers, and redundant features, the dataset is normalized to ensure comparability across different scales and dimensions. The normalization process follows the formula:
$$ x_{norm} = \frac{x_0 – x_{min}}{x_{max} – x_{min}} $$
$$ x_{scaled} = x_{norm}(x_{max} – x_{min}) + x_{min} $$
where $x_{norm}$ represents the standardized value, $x_0$ is the original feature data, and $x_{min}$ and $x_{max}$ denote the minimum and maximum values in the dataset, respectively. This preprocessing step ensures that the input features are suitable for neural network training, enhancing model convergence and prediction accuracy.
The core of the forecasting framework is the GRU-LSTM hybrid neural network, which combines the computational efficiency of GRU with the precision of LSTM in capturing long-term dependencies. The GRU component utilizes update and reset gates to manage information flow, with the update gate $g_t$ and reset gate $r_t$ computed as:
$$ g_t = \sigma(W_g [h_{t-1}, x_t] + b_g) $$
$$ r_t = \sigma(W_r [h_{t-1}, x_t] + b_r) $$
where $\sigma$ is the sigmoid activation function, $W_g$ and $W_r$ are weight matrices, $h_{t-1}$ is the previous hidden state, $x_t$ is the current input, and $b_g$ and $b_r$ are bias vectors. The candidate hidden state $y_t$ is then derived as:
$$ y_t = \tanh(W_y [r_t \odot h_{t-1}, x_t] + b_y) $$
and the current output $h_t$ is updated through:
$$ h_t = (1 – g_t) h_{t-1} + g_t y_t $$
The LSTM layer complements this with its ability to maintain cell state over time, employing forget, input, and output gates. The forget gate $f_t$ determines which information to discard:
$$ f_t = \sigma(W_f [h_{t-1}, x_t] + b_f) $$
while the input gate regulates new information storage through $e_t$ and $z_t$:
$$ e_t = \sigma(W_e [h_{t-1}, x_t] + b_e) $$
$$ z_t = \tanh(W_z [h_{t-1}, x_t] + b_z) $$
The cell state $C_t$ is updated as:
$$ C_t = f_t C_{t-1} + e_t z_t $$
and the output gate $o_t$ along with the current hidden state $h_t$ are given by:
$$ o_t = \sigma(W_o [h_{t-1}, x_t] + b_o) $$
$$ h_t = o_t \tanh(C_t) $$
By integrating these layers, the hybrid model effectively learns the nonlinear relationships between user characteristics and travel parameters, such as trip start times and durations. The architecture includes dropout layers for regularization and a fully connected layer for final output generation, ensuring robust predictions without overfitting.
To translate predicted travel data into spatial load distributions, I develop a road network model based on an actual urban area, simulating electric vehicle movements and charging behaviors. The network consists of multiple nodes connected by edges, with path planning performed using Dijkstra’s algorithm to compute shortest routes based on weighted costs. The driving cost $U_j$ for edge $j$ is calculated as:
$$ U_j = \sum_j \left( \frac{d_j}{v_j} \right) w_j $$
where $d_j$ is the travel distance in kilometers, $v_j$ is the velocity in km/h, and $w_j$ represents edge weights accounting for factors like time and congestion. This approach enables realistic simulation of electric vehicle routes, which is crucial for accurate load estimation.
The electric vehicle load calculation model incorporates battery dynamics and charging triggers. When the state of charge drops below 20%, a charging event is initiated. The SOC at time $t$ is updated as:
$$ SOC_t = SOC_{t-1} – \frac{Q d_j}{S_{max}} $$
where $Q$ is the energy consumption per kilometer in kWh/km, and $S_{max}$ is the battery capacity in kWh. The travel time $T_t$ accumulates as:
$$ T_t = T_{t-1} + \frac{d_j}{v_j} \times 60 $$
and the charging time $T_C$ required to reach a target SOC is:
$$ T_C = \frac{(SOC_{set} – SOC_{cur}) S_{max} \times 60}{P_C} $$
with $SOC_{set}$ set to 95% and $P_C$ denoting charging power in kW. The load matrix records temporal and spatial load distributions, with the total daily load $L_{sum}$ across all nodes computed as:
$$ L_{sum} = \sum_{i=1}^{1440} L(i, \cdot) $$
where $L(i, \cdot)$ represents the load at all nodes during minute $i$. This formulation allows for comprehensive assessment of grid impacts due to electric vehicle charging activities.
In the experimental analysis, I utilize the preprocessed NHTS dataset, containing 8,493 valid records after cleaning and normalization. The data is split into training and testing sets in an 80:20 ratio, and the model is implemented with a GRU layer of 50 hidden units and an LSTM layer of 32 units, using Adam optimization and mini-batch gradient descent. Key parameters for the electric vehicle load simulation include a battery capacity of 80 kWh, energy consumption of 0.296 kWh/km, and a charging power of 7 kW for standard cases. The China EV context is emphasized through the selection of parameters reflective of typical vehicles in the region.
The performance of the GRU-LSTM model is evaluated against standalone GRU and LSTM models using multiple metrics: root mean square error (RMSE), coefficient of determination ($R^2$), mean absolute error (MAE), and mean bias error (MBE). The formulas for these metrics are:
$$ E_{RMS} = \sqrt{\frac{1}{n} \sum_{i=1}^n (y_i – \hat{y}_i)^2} $$
$$ R^2 = \frac{\sum_{i=1}^n (\hat{y}_i – \bar{y})^2}{\sum_{i=1}^n (y_i – \bar{y})^2} $$
$$ E_{MA} = \frac{1}{n} \sum_{i=1}^n |y_i – \hat{y}_i| $$
$$ E_{MB} = \frac{1}{n} \sum_{i=1}^n (y_i – \hat{y}_i) $$
where $n$ is the number of samples, $y_i$ is the actual value, $\hat{y}_i$ is the predicted value, and $\bar{y}$ is the mean of actual values. The comparative results demonstrate the superiority of the hybrid approach in both accuracy and efficiency.
| Model | Data Type | RMSE | R² | MAE | MBE | Time (s) |
|---|---|---|---|---|---|---|
| GRU | Trip Start | 20.68 | 0.994 | 15.80 | 1.22 | 80.0 |
| GRU | Duration | 10.25 | 0.996 | 7.29 | -0.81 | 41.0 |
| LSTM | Trip Start | 19.52 | 0.994 | 15.69 | 0.75 | 94.0 |
| LSTM | Duration | 8.85 | 0.996 | 6.29 | -0.17 | 43.0 |
| GRU-LSTM | Trip Start | 18.05 | 0.995 | 14.71 | -0.07 | 103.0 |
| GRU-LSTM | Duration | 8.33 | 0.997 | 6.08 | 0.09 | 52.0 |
The results indicate that the GRU-LSTM model achieves lower errors and higher $R^2$ values across both trip start times and durations, validating its enhanced predictive capability. The slightly longer computation time is justified by the improvement in accuracy, making it suitable for practical applications in electric vehicle load forecasting.
Applying the predicted travel data to the road network model reveals distinct spatial and temporal load patterns. The spatial distribution shows uneven load concentrations, with certain nodes experiencing higher demands due to their locations in high-traffic areas. For instance, nodes situated near residential zones or transportation hubs exhibit peak loads exceeding 600 kW, while others remain below 400 kW. This heterogeneity underscores the importance of granular load forecasting for infrastructure planning. The temporal profile displays a pronounced peak around midday, with lows during early morning and late evening hours, aligning typical electric vehicle usage patterns in urban China EV scenarios.
The integration of data-driven travel prediction with physical network models enables precise estimation of charging loads at individual nodes, facilitating targeted expansion of charging infrastructure. This approach overcome the limitations of station-level forecasts by incorporating user behavior dynamics, thereby enhancing the relevance for grid operators and policymakers. The methodology’s scalability allows for adaptation to various regions, supporting the sustainable growth of electric vehicle adoption in China and beyond.
In conclusion, the GRU-LSTM hybrid model presents an effective solution for spatio-temporal electric vehicle load forecasting, leveraging advanced neural networks to capture complex user-travel relationships. The framework’s ability to generate accurate node-level predictions supports optimized charging station placement and capacity planning, contributing to the stability and efficiency of power systems amid increasing electric vehicle penetration. Future work could explore real-time data integration and multi-modal transportation influences to further refine forecasting precision for the evolving China EV landscape.