A comparative study of long-term precipitation forecasting models based on CEEMDAN decomposition

ABSTRACT Accurate monthly precipitation forecasting is of great significance for the rational exploitation of water resources and the prediction and prevention of floods. To improve the prediction accuracy of precipitation models, this study utilizes monthly precipitation data from January 2000 to December 2021 and constructs three coupled models – Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN)–Long Short-Term Memory (LSTM), CEEMDAN–Support Vector Regression (SVR), and CEEMDAN–Back Propagation (BP) Neural Network – based on the CEEMDAN decomposition method. These models are applied to Anyang City to compare their predictive performance. The results show that the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) of the CEEMDAN–LSTM model are 8.17, 5.66, and 0.9966, respectively. The RMSE, MAE, and R2 values for the CEEMDAN–SVR model are 22.29, 16.14, and 0.9745, respectively, while the CEEMDAN–BP model shows RMSE, MAE, and R2 values of 52.07, 34.44, and 0.8610, respectively. These results indicate that the CEEMDAN–LSTM model surpasses the CEEMDAN–SVR and CEEMDAN–BP models in terms of prediction accuracy, demonstrating its excellent predictive performance and strong potential for application in regional hydrological forecasting.