Multi-region Probabilistic Load Forecasting with Graph Bayesian Transformer Network

Abstract

Accurate probabilistic load forecasting is essential for efficient energy management and the safety operation of power system. Existing load forecasting methods suffer from two limitations: 1) Inadequate utilization of feature; 2) insufficient modelling capability for fine-grained dependencies. To end these problems, a multi-region probabilistic load forecasting method based on graph Bayesian Transformer network is proposed. Specifically, the proposed forecasting framework consists of graph neural network and hybrid Bayesian Transformer connected in cascaded configuration. The former one is used to develop multigraph spatial-temporal features, which can enhance the feature learning ability and share the graph structure information to realize the joint forecasting of multi-region. The latter one is used to capture multi-scale information, which can improve the adaptability of model to complex dynamic data and forecasting accuracy. For validation, a series of compared experiments and ablation analysis are conducted under New England dataset. The experimental results demonstrate that the proposed method has good performance in foresting accuracy, and adaptability. In particular, compared to other comparative methods, the Continuous Ranked Probability Score (CRPS) is reduced 32.7%.