Robust drought forecasting in Eastern Canada: Leveraging EMD-TVF and ensemble deep RVFL for SPEI index forecasting

Drought stands as a highly perilous natural catastrophe that impacts numerous facets of human existence. Drought data is nonstationary and noisy, posing challenges for accurate forecasting. This study proposes a novel hybrid framework integrating TVF-EMD preprocessing, LASSO feature selection and Ensemble Deep RVFL modeling for improved multistep ahead drought prediction. Using decomposed SPEI values, six machinelearning techniques (Support Vector Regression (SVR), Simple RVFL, Ensemble Deep RVFL, and Recurrent Neural Network (RNN), XGBoost, Random Forest (RF)) were applied to forecast the SPEI 12 12 drought index. The present study involved forecasting drought in two Canadian stations located in the eastern region (Charlottetown in Prince Edward Island and Fredericton in New Brunswick), where agriculture is rainfed and mostly affected by drought. The statistical period of 1980–2022 was considered for analysis. Following the decomposition of drought data with TVF-EMD, lagged data was generated using the TVF-EMD results. Training time was decreased by utilizing the Lasso regression feature selection algorithm to select effective inputs. Various statistical measures, including the root mean square error (RMSE) and correlation coefficient (R), were employed to assess the precision of the models. The research findings indicated that the TVF-ED-RVFL model achieved the highest level of precision in forecasting multistep ahead (1,3,6 and 12) SPEI 12 drought index for both Charlottetown and Fredericton stations. During testing, the TVF-ED-RVFL model predicted 1-month SPEI 12 for Charlottetown (R = 0.9995, RMSE = 0.0352) and Fredericton (R = 0.9974, RMSE = 0.0560). For multistep ahead forecasting, the Rvalues range from 0.9924 for 3-months ahead to 0.9242 for 12-months ahead in Charlottetown and range from 0.9846 for 3-months ahead to 0.8293 for 12-months ahead in Fredericton. By increasing the forecasting horizon, the accuracy of models decreased. The present study's outcomes can contribute to enhancing water management practices during periods of drought.