An intelligent weighted adaptive regression process for reliable and uncertainty aware drought forecasting

Accurate drought forecasting enables the management and reduction of impacts on water resources and farming systems. Hence, this research introduces a hybrid machine learning model called the weighted adaptive regression process (WARP), based on Gaussian process regression, generalized ridge regression, an adaptive mechanism to dynamically update the kernel matrix, and a feature relevance mechanism to assign weights to input variables. The model is optimized using the enhanced multi-operator Runge-Kutta algorithm. Moreover, a novel feature selection method based on kernel ridge regression and permutation importance is developed, while an enhanced multi-variate variational mode decomposition is implemented to decompose the input variables. The WARP applied to forecast SPEI at three stations in Iran outperformed LSSVM, DRVFL, and CFNN with correlation coefficient of up to 0.99 and MaxAE of 0.49. The uncertainty analysis demonstrated lower uncertainty and narrower prediction intervals confirming WARP as a reliable and promising drought forecasting approach.