Abstract:

The recurrent neural network (RNN) is applied to the thunderstorm prediction, and the prediction results in the next hour are given by using the historical reflectivity factor data of weather radar. The core of the network is spatiotemporal long short-term memory (ST-LSTM) unit, and a memory decoupling structure is added to separate temporal memory and spatial memory states. The training and testing datasets are constructed from thunderstorm data, which are sifted through the HKO-7 dataset of the Hong Kong Observatory. The results of ST-LSTM network with memory decoupling structure and without memory decoupling structure, are compared with the MIM (memory in memory) network and the traditional centroid tracking method. Numerical comparison of prediction score factor and individual case analysis results show that the prediction neural network is higher than the centroid tracking method in probability of detection success and critical success index, and the false alarm ratio is lower than the centroid tracking method. The prediction factor score of the network with memory decoupling structure is higher than that of the conventional ST-LSTM network and the MIM network, and the prediction effect of strong echo is better.

Key words: recurrent neural network (RNN), thunderstorm prediction, weather radar, deep learning