Transmission line icing forecasting based on characteristic graph structure

Icing of transmission lines has always been a pain point for grid companies. The economic and property losses caused by icing every winter are huge. How to make an effective prediction of transmission line icing is a difficult problem. Existing forecasting methods are often based on micro-meteorological and micro-topographic information. In the characteristic variables of micro-meteorology and micro-topography, there are often interdependencies and potential spatial correlations. However, existing icing prediction methods do not fully exploit the interactions among these characteristic variables. Therefore, this paper proposes a transmission line icing prediction model based on the feature map structure, which reveals the potential agnostic topological relationship between the feature variables by adaptively extracting the sparse adjacency matrix between the feature variables. In addition, while the dilated convolution can improve the receptive field, there is also a loss of information continuity due to the discontinuity of the convolution kernel of the dilated convolution. We propose a temporal capture module to improve the loss of information continuity through GRU and dilated convolution in parallel. End-to-end prediction is achieved by stacking a graph convolution module and a temporal capture module, and after conducting several experimental comparisons, the effective prediction of the proposed model is validated.