Abstract
To improve the accuracy of extreme precipitation prediction and based on CEEMDAN-decomposed LSTM and BP neural networks, this paper proposes a hybrid model. Traditional models like ARIMA, LSTM, and BP neural networks can capture precipitation trends but often exhibit large errors when predicting extreme precipitation events due to high data volatility and uneven distribution. Applying the Adaptive Noise Complementary Ensemble Empirical Mode Decomposition (CEEMDAN) algorithm for precipitating data in Yinchuan can solve this problem, reducing volatility and achieving a more balanced distribution, which simplified the modeling process. The BP and LSTM models were then used to predict each decomposed component individually, and the results were aggregated for overall precipitation forecasts. Experimental results show that the CEEMDAN-BP and CEEMDAN-LSTM models outperform single models in predicting extreme precipitation, with predictions closely matching actual observed values. This study offers new insights and methods for precipitation forecasting, and future work may expand the algorithm’s application to more datasets, incorporating additional meteorological factors for model optimization.