The Application of Recurrent Neural Network to Nowcasting

Radar extrapolation is an important means in nowcasting. The radar extrapolation methods widely used in China include COTREC and Optical Flow, by which two consecutive echoes are used to diagnose the advection velocity within rain analyses, involving the solution of Lagrangian persistence equation. A new method RNN(recurrent neural network) is applied in nowcasting. Using PredRNN(predictive RNN), by modeling historical radar data, the prediction of radar echo in the next hour is given. PredRNN consists of ST-LSTM unit, which is an improvement of LSTM. One advantage of using PredRNN is the operation of the state accumulation and the hidden layer output is replaced by convolution. Therefore, the neurons not only can get timing relationships, but also extract spatial features like convolutional layers. Another advantage is the addition of new spatial memory, which can enhance the transportation of the spatial feature information in different layers. In order to test the model performance, two radars of Daxing District of Beijing and Guangzhou are analyzed. The radar echo is pre-processed through quality controlling to remove isolated echo, abnormal echo, invalid radial and echo below 15 dBZ and ground echo, and then the combined reflectivity (CR) is made by 0-5 layers of data. To examine the applicability of the PredRNN, a contrast experiment is designed between PredRNN and COTREC, including an independent verification over months of each radar and two severe convective cases analysis. The test is carried out by point by point in three different reflectivity thresholds:20 dBZ, 30 dBZ and 50 dBZ. Indexes of verification are CSI, POD and FAR. The time range of the test is 0-1 h by 6 min. Results show that PredRNN has better forecast performance in all the verification items especially in 20 dBZ and 30 dBZ, when the CSI can be raised by 0.15-0.30, POD can be raised by 0.15-0.25, and FAR can be reduced by 0.15-0.20. This effect of improvement enhances with time. Although forecast performances of both PredRNN and COTREC fall with time, the performance of PredRNN method descends more slowly. The forecast performances of both PredRNN and COTREC fall with the increase of the combined reflectivity factor strength, which shows the insufficient of prediction ability for the region with intensity over 50 dBZ. Two cases show that the PredRNN method has predictive ability for the change of reflectivity factor intensity. In summary, PredRNN is suitable for nowcasting, and its forecast performance is much better than COTREC.