Combining time varying filtering based empirical mode decomposition and machine learning to predict precipitation from nonlinear series

Abstract

In recent years, due to the influence of human activities and changing climate conditions, precipitation time series have had increasing non-stationarity and randomness. Therefore, the scientific and accurate prediction of precipitation has become increasingly difficult. To improve precipitation prediction accuracy, nonlinear precipitation data was decomposed into several subcomponents via three decomposition methods (time varying filtering based empirical mode decomposition (TVF-EMD), wavelet transform (WT), and complementary ensemble empirical mode decomposition (CEEMD)). Then, the Elman neural network (ENN) was used to construct TVF-EMD-ENN, WT-ENN, and CEEMD-ENN models to predict the subcomponents of precipitation. Finally, these three models were applied to the monthly and daily precipitation series from different meteorological stations. The results showed that TVF-EMD-ENN made relatively stable predictions for monthly and daily precipitation, and was better than WT-ENN, CEEMD-ENN, and other models. Predictions for Turpan station were generally poor, likely due to the discontinuous resonance periods and abnormal resonance periods between precipitation and climate indices. Overall, the predictions by TVF-EMD-ENN were better, which was clearly reflected in the variance contribution rates (VCR) of high-frequency subcomponents, which were lower than those of WT and CEEMD. Parameter sensitivities of TVF-EMD indicated that the bandwidth threshold (BT) and the b-spline order (BSO) had a strong influence on model performance, and that BT and BSO should remain within 0.15–0.50 and 15–65, respectively. This study has provided a new approach to precipitation prediction that is better equipped to handle non-stationarity and randomness.