Improving radar echo extrapolation algorithms based on self-attention and gated recurrent neural networks

To enhance the spatiotemporal feature extraction ability of existing neural networks for radar echo sequences, establish a spatiotemporal sequence prediction model with better extrapolation performance, and carry out research on improving radar echo extrapolation algorithms. Based on the radar echo dataset jointly established by the Shenzhen Meteorological Bureau and the Hong Kong Observatory, at the data processing level, by improving the normalization method of radar echo image sequences, the prediction level of five commonly used spatiotemporal series prediction models for strong echoes has been improved. At the model algorithm level, two simultaneous self attention structures are introduced into the ST-LSTM structure to form a new cyclic gating unit, and these cyclic gating units are stacked to establish the ST-SARNN model. Selecting CSI and POD as accuracy evaluation indicators for model comparison analysis: (1) The improved normalization method has improved the prediction level of five commonly used spatiotemporal series prediction models in recent years for strong echoes. (2) The ST-SARNN model incorporating self attention has significantly better predictive performance than models such as ConvLSTM, PredRNN, and MIM for radar echoes. The improved normalization method can change the distribution of sample data and improve the model's extrapolation performance to a certain extent; The self attention structure can effectively mine the spatiotemporal features of radar echo sequences, thereby improving the extrapolation performance of neural networks.