Abstract:

This study aims to explore the application of recurrent neural network models in nonlinear prediction and improve the accuracy of thunderstorm potential prediction. An LSTM model was constructed using Guiyang sounding station as an example, and the number of lightning occurrences within a 100 km and 12-hour range around the station was counted. Twenty-three sounding physical parameters with a correlation coefficient (r) of more than 0.8 with lightning activity were obtained and used as sample features for the prediction model. The impact of hyperparameter selection on the model was analyzed. An investigation was carried out into thunderstorm potential forecasting within a 0-24 hour timeframe and its performance was evaluated and compared. The research results indicate that constructing an LSTM network model with a learning rate of 0.000 1, batch sample size of 32, and input sequence length of 10 is beneficial for improving the model's generalization and accelerating its convergence. The model, by inputting 23 sounding physical parameters such as gale index and correction index of the previous 5 days, found that effect of 0-12 hour thunderstorm potential prediction is significantly better than that of 12-24 hour prediction. This model was used to verify the 0-12 hour test data and it was found that the area under curve(AUC) the receiver operating characteristic curve is close to 1, the hit rate is 93.4%, the false alarm rate is 17.4%, the critical success index is 78.1%, and the error rate is only 6.6%, which verifies the effectiveness of the model used.

Key words: long short-term memory (LSTM) recurrent neural network, atmospheric physical quantity parameters, correlation, lightning potential prediction, accuracy