Parallel Spatial-Temporal Graph Attention Network for Short Term Multi-Sequence Load Forecasting

Abstract

Short-term load forecasting is a crucial task within the power system. However, existing studies have overlooked the spatial-temporal relationships between multiple series loads. Accounting for this spatial-temporal adjacency can lead to more accurate forecasting in certain scenarios. In this paper, we propose a short-term load forecasting model named Parallel Spatial-Temporal Graph Attention Network (PST-GAT). The method encodes the multi-sequence loads as nodes and constructs the adjacency matrix using the Dynamic Time Warping (DTW) technique to form a fully connected graph of the load data. Combining the sliding window concept, the constructed fully connected graph is partitioned into a series of subgraphs, and the Graph Attention Network (GAT) performs feature extraction on each subgraph individually. To realize learning at multiple scales, PST-GAT adopts a parallel multi-branching approach, where each branch splits the subgraphs with varying lengths. Finally, the feature vectors extracted by each branch are concatenated, and forecasting is accomplished using a fully connected layer. Moreover, the unique structural design of PST-GAT allows for the simultaneous prediction of multiple sequence loadings. Experimental results based on real-world load data validate the superior prediction accuracy of this method compared to existing algorithms.