Research on substation flood forecasting model with spatio-temporal attention radar echo feature identification

In order to prevent the impact of extreme rainstorms on power equipment in substations, accurate and reliable real-time precipitation estimates are essential for emergency flood control in substations and to ensure stable grid operation. Compared with ground precipitation measurements, weather radar echo-based monitoring can effectively utilize the latest available information for short-term prediction. In this paper, a deep learning model based on spatio-temporal attention mechanism is proposed for precipitation estimation. The model extracts the rainfall intensity resolution at spatial granularity in high-resolution radar echo maps using separable convolution operations in UNet deep networks, combined with an attention mechanism to capture the relevant features at temporal granularity. To evaluate the effectiveness of the proposed model, a comparison is made with other deep learning models in terms of probability of detection, false alarm rate, critical success index and Heidke skill score. Experimental results on precipitation data from a meteorological institute show that the proposed model has better performance in short-time precipitation forecasting compared with traditional methods.