Forecasting drought across Canada using a hybrid machine learning approach: a national perspective for adaptations and early warning systems

Drought is a major natural hazard in Canada, with significant socio-economic impacts across the country’s diverse climatic and geographic regions. This study proposes a hybrid forecasting framework that combines fast adaptive empirical mode decomposition (FAEMD), an ensemble deep random vector functional link (edRVFL) network, and the energy valley optimizer (EVO) to model nonstationary hydrological time series. Using 74 years (1951–2024) of meteorological data from 24 Canadian meteorological stations, we first identified substantial spatial variability and nonstationary trends in historical drought conditions. The proposed FAEMD-edRVFL-EVO framework was then used to forecast the Standardized Precipitation Evapotranspiration Index (SPEI 12 ) at one-, two-, and three-month lead times and compared against standalone and hybrid implementations of extreme gradient boosting, random forest, and k -nearest neighbor algorithms. FAEMD-edRVFL-EVO consistently achieved the best performance, with median correlation coefficients of 0.999, 0.980, and 0.906 for one-, two-, and three-month forecasts, respectively. Performance remained strong in climatically complex regions, particularly in Western and Northern Canada. By effectively isolating nonstationary signals, this applied study demonstrates that the integrated pipeline provides a scalable, accurate, and computationally efficient practical solution for large-scale drought forecasting and early warning under increasing climatic uncertainty.