Multi‐load forecasting for integrated energy systems based on GAT‐MTL

Abstract Accurately forecasting multi‐load in integrated energy systems is crucial for enhancing overall energy efficiency and achieving greater economic benefits. However, a strong spatial correlation exists between load and meteorology. Graph attention network, a kind of neural network model dealing with graph data, can greatly analyse the intrinsic relationship between multi‐load and meteorological factors. To this end, a multi‐load prediction model with graph attention network and multi‐task learning is proposed. Firstly, considering the complex influencing factors of multi‐load, the authors propose a method combining Spearman correlation and causality tests to select feature variables for the model. Subsequently, an attention mechanism is introduced to assign varying degrees of attention to nodes in the graph data, thereby enhancing the model's ability to learn representation. Secondly, graph attention network is employed to construct the shared layer of multi‐task learning, with the Bayesian optimization algorithm utilized to optimize hyperparameters and fully explore the interdependencies among cooling, heating, and electricity loads. Finally, a real case study is used to demonstrate that the model can effectively improve the accuracy of multi‐load forecasting.