A New Multiscale Deep Learning Model for Daily Runoff Prediction in Snow-Influenced Alpine Catchments

Abstract

Accurate daily runoff prediction is fundamental to understanding runoff dynamics and supporting water resources management. However, runoff prediction in snow-influenced alpine catchments remains challenging because complex hydrometeorological and cryospheric interactions generate strongly seasonal and nonstationary runoff dynamics. To address these challenges, we proposed IWOA-VCBA, a multiscale deep learning model integrating variational mode decomposition (VMD), a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), an attention mechanism, and an improved whale optimization algorithm (IWOA), using the Lienz catchment in East Tyrol, Austria, as a case study. VMD constructed multiscale features from meteorological variables and antecedent runoff, CNN-BiLSTM-Attention learned local feature interactions, lagged dependencies, and key temporal information, and IWOA optimized key network hyperparameters. Model performance was evaluated for overall runoff, peak flows, snowmelt-season runoff, and direct cross-catchment model transfer to the Ziller, Möll, and Salzach catchments without target-catchment retraining or hyperparameter re-optimization. In Lienz, IWOA-VCBA achieved KGE and NSE values of 0.969 and 0.947, respectively, yielded the lowest peak-magnitude and peak-timing errors among the evaluated models, and maintained strong snowmelt-season performance, with KGE and NSE values of 0.961 and 0.945. In the transfer experiments, mean KGE and NSE values of 0.908 and 0.841 were obtained across the three target catchments, indicating useful predictive skill under direct transfer. These results demonstrate the potential of IWOA-VCBA for accurate daily runoff prediction and direct cross-catchment application in snow-influenced alpine catchments.