Short-Term Multi-Energy Load Forecasting Based on Multiscale Spatial-Temporal ChebyNet

The accurate prediction of multi-energy loads is of crucial importance for the optimisation, scheduling and economic operation of integrated energy systems. However, current multi-energy load forecasting methods still suffer from insufficient capture of the complex coupling between multi-energy loads and meteorological features, as well as inadequate extraction of short-term characteristics and periodic patterns. In response to the aforementioned issues, we propose a Multiscale Spatial-Temporal ChebyNet model that learns directed signed graph structures from both dynamic and static views, aggregates node information over them through a role-lifted Chebyshev spectral filter, and uses a multiscale forward attention mechanism and gated dilated convolutions to capture short-term and periodic temporal dependencies. Experiments on two public datasets show that MST-ChebyNet outperforms ten baseline networks in forecasting electricity, cooling, and heating loads. Compared with the representative adaptive graph based spatiotemporal models Graph WaveNet and MTGNN, it reduces load MAPE by 0.54–1.40% on the Arizona dataset and by 0.59–1.53% on the Eagle dataset.