A multivariate decomposition-ensemble model for estimating long-term rainfall dynamics

This study aims to present a novel decomposition-ensemble model that uses multivariate data. Two algorithms, Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost), were used to develop a new model for rainfall reconstruction using daily meteorological data from 2003 to 2017 in Seoul, South Korea. First, the dataset was decomposed by two decomposition methods: singular spectrum analysis (SSA) and empirical mode decomposition (EMD). Second, the input time series was constructed as trend terms, fluctuating terms, and noise components in the SSA method and as intrinsic mode functions (IMFs) in the EMD method. Finally, these decomposed datasets were used as the input sets for the ensemble models for training, testing, and evaluation to reconstruct long-term daily rainfall. Performance statistics indicated that SSA integrated with LightGBM improved the efficiency over other combinations (EMD-LightGBM, SSA-XGBoost, EMD-XGBoost) by a lower root-mean-square error (RMSE) and higher Nash-Sutcliffe efficiency coefficient (NSE). After selecting the best combination, support vector machine recursive feature elimination (SVM-RFE) was applied to select the best decomposed dataset to improve the SSA-LightGBM performance. Finally, the performance of the model was compared to Random Forest (RF) algorithm for robustness analysis. The results showed that using the SSA and LightGBM models reconstructed long-term daily rainfall more accurately, especially when coupled with SVM-RFE, which obtained values for the square of the correlation coefficient (R-2), RMSE, NSE, and mean absolute error (MAE) of 0.92, 3.27, 0.91, and 0.99, respectively. In particular, the proposed decomposition-ensemble model (SSA-SVM-RFE-LightGBM) can be used to reconstruct long-term daily rainfall on a global scale.