基于动态自适应图神经网络的电动汽车充电负荷预测

电动汽车充电站负荷波动的不确定性与长时间预测任务给提升充电负荷预测精度带来巨大的挑战。文中提出一种基于动态自适应图神经网络的电动汽车充电负荷预测算法。首先，构建了一个充电负荷信息时空关联特征提取层，将多头注意力机制与自适应相关图结合生成具有时空关联性的综合特征表达式，以捕获充电站负荷的波动性；然后，将提取的特征输入到时空卷积层，捕获时间和空间之间的耦合关系；最后，通过切比雪夫多项式图卷积与多尺度时间卷积提升模型耦合长时间序列之间的能力。以Palo Alto数据集为例，与现有方法相比，所提算法在4种波动情况下的平均预测误差大幅降低。

Electric Vehicle Charging Load Prediction Based on Dynamic Adaptive Graph Neural Network

The uncertainty and long-term prediction of the load fluctuation of electric vehicle (EV) charging stations pose significant challenges to accurately predict the charging load. An EV charging load prediction based on dynamic adaptive graph neural network is proposed. Firstly, a spatiotemporal correlation feature extraction layer for charging load information is constructed. By combining multi-head attention mechanism with adaptive relevance graph, a comprehensive feature representation with spatiotemporal correlation is generated to capture the load fluctuation of EV charging station. Then, the extracted features are input into a spatiotemporal convolutional layer to capture the coupling relationship between time and space. The ability of the model to couple long time series is enhanced by Chebyshev polynomial graph convolution and multi-scale temporal convolution. The effectiveness of the algorithm has been verified using two real datasets. Taking the Palo Alto dataset as an example, compared with existing methods, the average prediction error of this algorithm under 4 volatile conditions is reduced sharply.