Dynamic Spatial-Temporal Load Forecasting for Multi-Regional Power Systems via Graph Attention Networks

Accurate multi-region load forecasting (MRLF) can effectively ensure the stable operation of the power system and is essential for the economic dispatch of the power system. However, with the gradual deepening of the dynamic spatial dependence among multi-region loads and the unmeasured nature of loads, the challenge is posed to MRLF. To address this challenge, a graph attention learning-based load forecasting model for multi-region considering dynamic spatial correlations aggregation is presented in this paper. Firstly, the regional load series is decomposed into a trend component and a fluctuation component through the discrete wavelet transform. Secondly, a generative adversarial network with the concept of zero-sum game is proposed for adversarial training of spatio-temporal prediction models. Furthermore, reasonable future multi-regional load forecasting values are obtained through a aggregation module that aggregate the trend component and the fluctuation component. Finally, the multi-regional load data of the New York Independent System Operator (NYISO) is used as a case. Compared with the evaluation metrics of mainstream models, the model presented in this paper is effective and superior in MRLF.