Short-term load forecasting method for integrated energy systems based on graph neural network and multi-task balance

Electricity load, heat load and cold load in integrated energy systems are related to each other. Inspired by Graph Neural Networks (GNN), which can capture the topological structure of graph data, this paper proposes a novel multi-energy forecasting framework with multi-level task-sharing matrices (MTMs), which connects multiple tasks to capture the coupling characteristics between multiple loads. The proposed framework overwhelms most traditional multi-task approaches because traditional models usually have shared parameters in only one stage, which reserves representation space for temporal features of individual loads but limits their ability to capture multi-task correlations across multiple stages. Specifically, the proposed framework can capture coupling features across various stages of the entire model by MTMs, significantly enhancing the connections among multiple tasks while preserving adequate representation space for temporal features. The combination of graph attention and graph convolution can further expand the representational space of coupling features, so that multi-head attention can focus on capturing the temporal characteristics of individual loads. In addition, a gradient-based multi-task balance method is proposed to adapt to the above framework, which can balance different tasks by normalizing the weights by MTMs. Case studies demonstrate that the proposed framework has superior forecasting performance for multi-energy loads.