Complete ensemble empirical mode decomposition hybridized with random forest and kernel ridge regression model for monthly rainfall forecasts

Persistent risks of extreme weather events including droughts and floods due to climate change require precise and timely rainfall forecasting. Yet, the naturally occurring non-stationarity entrenched within the rainfall time series lowers the model performances and is an ongoing research endeavour for practicing hydrologists and drought-risk evaluators. In this paper, an attempt is made to resolve the non-stationarity challenges faced by rainfall forecasting models via a complete ensemble empirical mode decomposition (CEEMD) combined with Random Forest (RF) and Kernel Ridge Regression (KRR) algorithms in designing a hybrid CEEMD-RF-KRR model in forecasting rainfall at Gilgit, Muzaffarabad, and Parachinar in Pakistan at monthly time scale. The rainfall time-series data are simultaneously factorized into respective intrinsic mode functions (IMFs) and a residual element using CEEMD. Once the significant lags of each IMF and the residual are identified, both are forecasted using the RF algorithm. Finally, the KRR model is adopted where the forecasted IMFs and the residual components are combined to generate the final forecasted rainfall. The CEEMD-RF-KRR model shows the best performances at all three sites, in comparison to the comparative models, with maximum values of correlation coefficient (0.97–0.99), Willmott's index (0.94–0.97), Nash-Sutcliffe coefficient (0.94–0.97) and Legates-McCabe's index (0.74–0.81). Furthermore, the CEEMD-RF-KRR model generated the most accurate results for Gilgit station considering the Legate-McCabe's index as base assessment criteria in addition to obtaining the lowest magnitudes of RMSE = 2.52 mm and MAE = 1.98 mm. The proposed hybrid CEEMD-RF-KRR model attained better rainfall forecasting accuracy which is imperative for agriculture, water resource management, and early drought/flooding warning systems.