Аңдатпа
Reliable streamflow forecasting supports a wide range of water-management decisions, from flood warning to long-term water resource planning. In snow-influenced basins, forecast performance depends on accurately representing seasonal snow accumulation and melting. Transfer learning and multitask learning have each been explored for deep-learning streamflow models, but their combined use for snowmelt-driven forecasting remains insufficiently understood. In this study, we propose a multitask transfer-learning framework that integrates snow water equivalent as an auxiliary prediction target during both large-sample pretraining and target-basin fine-tuning for Long Short-Term Memory streamflow prediction models. The framework is evaluated through four controlled training configurations designed to assess the effects of transfer learning and auxiliary supervision on overall and high-flow forecasting performance. Local single-task, local multitask, transfer-based single-task, and transfer-based multitask learning were compared across four input sequence lengths from 30 to 365 days. Models were evaluated on 30 held-out snow-influenced basins from CAMELS-US dataset and, as a single out-of-domain case study, on the Uba River basin in East Kazakhstan. Transfer learning raised median Nash-Sutcliffe efficiency from 0.65 under local single-task training to 0.77, and adding snow water equivalent supervision during transfer raised it further to 0.79, the corresponding absolute high-flow bias dropped from 56% to 33%, and to 26% with multitask supervision. Auxiliary snow water equivalent supervision produced significant improvements within the transfer-learning framework, with the clearest gains in high-flow bias and peak-timing accuracy. These results suggest that physically meaningful auxiliary targets act as representation guides during large-sample pretraining, improving snowmelt-driven high-flow forecasting.