A Regional Short-Term Load Forecasting Approach Based on Multi-Channel Fusion and Adaptive Graph Neural Network

Accurate short-term load forecasting is crucial for the operation of regional power grid systems. However, different types load within a region usually exhibit different variation patterns in short-term time scale. Existing forecasting methods mostly ignore the differentiated load characteristics and the interrelationships among various loads, solely focus on the overall trend of changes. Graph neural network (GNN) provides a way to consider the correlation of various types of loads when establishing load forecasting model, but it still has two major problems: 1) the predefined adjacency matrix is difficult to fully and accurately represent the correlation of different industry loads; 2) the graph convolution operation is difficult to take into account the different time series variation patterns of various loads. To this end, a regional load forecasting approach based on multi-channel fusion and adaptive graph neural networks (MC-ADAGNN) is proposed, which is designed for regional power grids with various load types. Firstly, multiple feature extraction channels including Skeleton-Foundation, Dynamic-Interaction, and Core-Stability are fused to build a graph construction layer, which can adaptively generate the adjacency matrix to extract the correlative features of different types of load nodes. Secondly, considering the different scales periodic variation of different loads, a convolution module including multi-hop propagation layer and inception-dilated convolution is constructed, speeding up the information transfer efficiency and adaptively selecting the convolution kernel size to learn multi-scale features. Finally, the features output by the convolution module are passed through the output layer to give the load forecasting results of each category, and they are summed to obtain the regional load. Simulation results based on the actual load data show that the MC-ADAGNN model has the best performance compared with other benchmarks, verifying the validity of the proposed approach.