Integrating Multi Spatial-Temporal Information with Globally Attentive Gated Convolutional Network for Transmission Line Icing Tension Prediction

Accurate transmission line tension prediction is crucial for avoiding power grid suffering from serious lines ice coating, which is a very challenging problem affected by multiple complex reasons, e.g., the massive sharp changes accompanying the occurrence of ice cover, and the non-negligible association with multiple influencing factors (likes local geographic information and future weather forecasts). In addition, spatial-temporal correlation among different time intervals and different factors leads to their different contribution degrees to prediction. To address these issues, we propose a Globally Attentive Gated Convolutional Network (GAGCN) to integrate multiple sources of Spatial-Temporal information for transmission line icing tension prediction. Specifically, we first comprehensively fuse historical observations and external static factors for simultaneous modelling, second the Global Time-series Attention module is applied to assign different attention to each time step and each feature of the input sequence from a global perspective. Then, a Temporal Gated Convolutional Module is subsequently set up to capture the temporal dependence with various subperiods. Finally, future information is fed into the prediction component along with historical information, where the shape and temporal criteria are incorporated into the training objective by a hybrid-loss function so as to capture sharp changes more accurately and timely. Experiments on two real-world datasets demonstrate the superior performance of our GAGCN beyond state-of-the-art methods.