基于可解释人工智能的空间电力负荷预测方法OA

To address the issues of insufficient historical data for spatial power load forecasting in urban distribution networks and the lack of explainability of existing artificial intelligence models,this paper proposes a spatial power load forecasting method based on explainable artificial intelligence.First,a power geographic information system is established to generate Type Ⅰ cells and Type Ⅱ cells,and a time series generative adversarial network(TimeGAN)is employed to construct a data augmentation model for few-shot scenarios.Second,a spatio-temporal information graph of Type Ⅱ cells is built,and the spatial attention mechanism of a graph attention network(GAT)is used to extract the spatial features of each Type Ⅱ cell load.A heatmap of attention weights is plotted to visualize the degree of attention each Type Ⅱ cell receives during information aggregation,thereby providing explainability of the forecasting model in the spatial dimension.Then,the temporal self-attention mechanism of iTransformer is utilized to capture the temporal features in the load sequences of the cells.The temporal attention weights are extracted to identify the time steps that play key roles in the forecasting task,offering explainable support for the model's decision-making basis in the temporal modeling process.Finally,the spatial features of Type Ⅱ cell loads output by the GAT and the temporal features output by the iTransformer are fed into the fully connected layer of the iTransformer.Meanwhile,a parallel one-dimensional convolutional neural network(1D-CNN)module is introduced,and skip connections are used to enhance the model's ability to represent and utilize spatio-temporal features at the output stage.Engineering case studies demonstrate that the proposed forecasting method achieves higher accuracy than traditional forecasting methods.