A forecasting method for corrected numerical weather prediction precipitation based on modal decomposition and coupling of multiple intelligent algorithms

Numerical weather models often face significant challenges in achieving high prediction accuracy. To enhance the predictive performance of these models, a solution involving the integration of deep learning algorithms has been proposed. This paper introduces a machine learning approach for correcting the numerical weather forecast results from the Weather Research and Forecasting (WRF) model. Initially, the WRF model is used to simulate summer precipitation in the Jinsha River Basin. Subsequently, the adaptive noise-robust empirical mode decomposition (CEEMDAN) method is employed to decompose WRF simulation errors. These decomposed subsequences are then input into four machine learning algorithms and two metaheuristic optimization algorithms to predict the error sequences. Finally, the predicted error subsequences are merged and superimposed on the WRF simulation values to obtain the corrected precipitation. Research findings demonstrate that the integration of machine learning algorithms with WRF significantly improves prediction accuracy. The correlation coefficient of the optimal model increases by 158%, and Nash-Sutcliffe Efficiency (NSE) increases by 149% compared to before correction. This indicates that correcting the WRF model through deep learning methods effectively enhances precipitation forecasting accuracy.