Neural Computing and Applications Программное обеспечение

Heterogeneous Spatiotemporal Multi-Source Data Fusion for Short-Term Transmission Line Icing Forecasting

Xiong, Zhaofeng, Zhang, Wei, Tan, Ling, Chen, Wenjia
Springer Nature · 2026-11-21 · DOI 10.1007/s00521-026-12564-4

Аннотация

Images of icing on power-transmission lines are widely used for icing forecasting because of their intuitive and convenient nature. Existing studies suffer from limited data and shortcomings in fusion modeling, bringing constrained accuracy and homogeneous outputs. To address these issues, we present a heterogeneous spatiotemporal multi-source data fusion mechanism and propose MUC-LSTM (Multi-source Unified Cross-scale Convolutional Long Short Term Memory), a short-term icing forecasting model for transmission lines that enables precise multi-dimensional prediction of future icing images, icing type, and icing thickness. MUC-LSTM comprises a Multi-Source spatio-temporal Fusion (MSSF) module, a U-shaped Cross-modal Long Short-Term Memory (UC-LSTM) image-prediction module, and a Multi-Loss Fusion (MLF) optimization module. It employs a spatio-temporal-channel collaborative attention mechanism to focus on conductor icing dynamics and integrates super-resolution upsampling to enhance fine-grained feature recovery. To alleviate gradient conflicts among the multi-dimensional tasks, we introduce an inter-coupled, closed-loop multi-task learning scheme together with multiple loss functions, which effectively improves forecasting accuracy. To mitigate background interference in icing images, MUC-LSTM further uses spatio-temporal-channel weights to guide sub-pixel upsampling and residual detail compensation, thereby restoring high-resolution spatial details and markedly improving detail reconstruction. Experiments on the IceCast-Grid dataset show that the method substantially improves short-term icing forecasting accuracy: compared with other mainstream models, PSNR(Peak Signal to Noise Ratio) increases by at least 0.8141 and MSE(Mean-Square Error) decreases by at least 1.76. These results indicate that multi-source fusion can accurately and efficiently accomplish icing forecasting across different temporal scales, providing reliable decision support for grid icing disaster prevention and control.

Читать у издателя →