Annual Peak Runoff Forecasting Using Two-Stage Input Variable Selection-Aided k-Nearest-Neighbors Ensemble

Annual river forecasting is a useful tool for practical flood management. However, since it involves a problem with a small dataset, deep learning models with complex network structures are generally not applicable. In this study, two-stage input variable selection (IVS)-aided k-nearest-neighbors (kNN) ensemble models are developed for annual peak flow (APF) forecasting at the Lechangxia Reservoir in China. In the first stage, three correlation-based metrics (Pearson correlation coefficient, Spearman correlation coefficient, and mutual information index) are used to rank tele-connected indicators according to their linear and nonlinear relationships with historical APF. The top indicators identified through this filter method are forwarded to the second stage, where a leave-one-out cross-validation-based exhaustive search systematically evaluates all possible combinations of the retained inputs. The optimal kNN member uses the two-stage IVS method based on the Spearman correlation coefficient. Multiple kNN models are subsequently developed using distinct input subsets, and these models are aggregated through a simple average method to create the final ensemble forecast. The optimal ensemble model improves the validation R and RMSE of the optimal member model, by 17.5% and 16.1%, respectively. This study highlights the effectiveness of improving long-term river forecasting performances through integrating pre-processing (two-stage input variable selection) and post-processing (multi-strategy kNN ensemble) methods.