A multi-source deep learning framework named MUC-LSTM substantially improves short-term transmission line icing forecasting accuracy by simultaneously generating multi-dimensional predictions of future icing images, icing type, and icing thickness, according to research published in Neural Computing and Applications.
Visual imagery of power-transmission line icing is widely utilized for forecasting due to its intuitive and convenient characteristics. Nonetheless, existing operational studies encounter notable limitations in training data volume and structural shortcomings in fusion modeling, which constrain predictive accuracy and produce homogeneous outputs.
To address these constraints, the study presents a heterogeneous spatiotemporal multi-source data fusion mechanism configured within a Multi-source Unified Cross-scale Convolutional Long Short Term Memory architecture. The model integrates three primary components: a Multi-Source spatio-temporal Fusion module, a U-shaped Cross-modal Long Short-Term Memory image-prediction module, and a Multi-Loss Fusion optimization module.
The system incorporates a spatio-temporal-channel collaborative attention mechanism specifically calibrated to track conductor icing dynamics over time. To strengthen fine-grained feature recovery, the model embeds super-resolution upsampling alongside spatio-temporal-channel weights that guide sub-pixel upsampling and residual detail compensation, mitigating background interference in camera imagery and restoring fine spatial details.
To alleviate gradient conflicts that typically arise during concurrent multi-dimensional tasks, the authors implemented an inter-coupled, closed-loop multi-task learning scheme paired with multiple loss functions.
Experimental validation on the IceCast-Grid benchmark dataset demonstrated that the proposed method markedly outperforms mainstream models. In comparative evaluations, Peak Signal to Noise Ratio increased by at least 0.8141 and Mean-Square Error decreased by at least 1.76, providing robust evidence that multi-source fusion can support short-term disaster prevention and control across electrical transmission grids.
The reported evaluation metrics confirmed that multi-source data fusion maintains consistent forecasting capability across different temporal horizons.
Subject of Research: Short-term transmission line icing forecasting using spatiotemporal multi-source data fusion
Article Title: Heterogeneous Spatiotemporal Multi-Source Data Fusion for Short-Term Transmission Line Icing Forecasting
Article References: Xiong, Z., Zhang, W., Tan, L., & Chen, W. (2026). Heterogeneous Spatiotemporal Multi-Source Data Fusion for Short-Term Transmission Line Icing Forecasting. Neural Computing and Applications, 38(22), Article 906. https://doi.org/10.1007/s00521-026-12564-4
Image Credits: AI Generated
DOI: 10.1007/s00521-026-12564-4
Keywords: transmission line icing, icing forecasting, spatiotemporal data fusion, multi-source data fusion, MUC-LSTM, Multi-Source spatio-temporal Fusion, UC-LSTM, Multi-Loss Fusion, IceCast-Grid dataset, Peak Signal to Noise Ratio, Mean-Square Error, conductor icing dynamics, collaborative attention, sub-pixel upsampling, residual detail compensation, multi-task learning, grid disaster prevention
News Source: Denise Maddox. (November 21, 2026). Multi-Source Fusion Model Improves Short-Term Power Line Icing Forecast Accuracy. Scienmag.



