Transmission line icing prediction based on multi-source monitoring data

Abstract

With the development of smart grids, accurate and real-time transmission line icing prediction is crucial for power system safety. Traditional methods struggle to capture the complexity of icing formation and multi-factor interactions. To address this, we propose a Multi-source Spatiotemporal Attention Network (MSTAN) for icing thickness prediction. MSTAN integrates meteorological, line status, and environmental data to fully characterize influencing factors. A ConvLSTM module extracts spatiotemporal features, modeling both temporal dynamics and spatial correlations. An attention mechanism highlights key features during icing formation, enhancing sensitivity to critical inputs. A multi-source fusion layer is also designed to improve the complementarity of heterogeneous data. Experimental results show that MSTAN outperforms traditional LSTM models, reducing RMSE by 33.1% and increasing R² by 4.5%, demonstrating its effectiveness and potential for practical application.