Multi-scale temporal representation with sparse dynamic graph learning for district heat load forecasting

The diverse temporal patterns of heat demand in district heating systems (DHS), pose significant challenges to achieving accurate forecasting. Graph neural networks (GNNs) exhibit promise in capturing these spatio-temporal dependencies, but existing models are limited by focusing on single-time-scale temporal patterns. While multi-scale graph representations in GNNs could introduce more timing-dependent features, its high computational cost constrains wider real-world application. Addressing this, we propose a Multi-scale Sparse Dynamic Graph Neural Network (MSDGN) for district heat load forecasting. MSDGN uses a multi-scale dynamic graph structure to learn various temporal dependencies without pre-defined priors. It includes a temporal attention mechanism, which reduces computational costs by sparsifying the graph’s edges. Additionally, MSDGN also integrates a spatio-temporal enhancement module and a residual fusion module, efficiently extracting features across scales and including recent short-term trends. Comprehensive experiments with real-world data from 3021 heat meters demonstrate MSDGN’s effectiveness and superiority over other state-of-the-art methods across various settings.