Weather Radar Echo Extrapolation Based on Convolutional LSTM Prediction Neural Network

We conducted research on weather radar echo extrapolation prediction based on convolutional LSTM neural network. It can be seen that in the input-target (X-T) time-series samples, the sample closest to the predicted result gradually changes backwards (in time) as the epoch increases. When the closest sample is the target sample, it indicates that the model has good predictive ability. Therefore, a loss function based solely on the similarity between the predicted result and the target sample is not sufficient to determine the training level of the prediction model. It is recommended to simultaneously examine the similarity between the predicted result with X- T time-series samples to determine the training level and termination condition of the prediction model. In addition, by comparing with the small ball testing prediction model, we speculate that due to the smaller detection range and faster image change of X-band weather radar, the training of echo extrapolation prediction neural network is faster than that of S-band weather radar.