Contrastive self-supervised learning for streamflow forecasting: a multi-station comparison with supervised models

Daily streamflow forecasting is complicated by the fact that the predictive value of recent observations shifts considerably depending on the hydrological character of the catchment. This study develops a regime-aware, window-adaptive framework to compare conventional supervised learning algorithms — XGBoost, Random Forest, LSTM, and GRU — against four contrastive self-supervised (SSL) approaches: SimCLR, TS2Vec, CPC, and TempContrast. Experiments were conducted at four gauging stations in Türkiye, each representing a distinct hydroclimatic regime. The SSL stage employs a shared LSTM encoder pretrained with hydrology-informed augmentations and is evaluated across window lengths of 7, 14, 30, and 60 days, allowing a direct assessment of how context length interacts with catchment memory. A bias-regularised loss function is applied during fine-tuning to reduce systematic prediction errors — an aspect that has received limited attention in prior SSL-based hydrological studies. Model performance was quantified using R2, NSE, KGE, Willmott’s d, RMSE, MAE, MBE, and PBIAS. No single model emerged as universally superior. At the strongly seasonal Demirkapı station, supervised models reached Test R2 = 0.96–0.97, while SSL peaked near 0.91. At the low-flow, repetitive Beyşehir station, short-window SSL outperformed all supervised baselines, achieving Test R2 ≈ 0.93–0.94 against 0.89–0.90. At the noisier Akçay and Karadere stations, short-lag supervised models proved more reliable. These results point to a consistent pattern: the effectiveness of any given approach depends heavily on the memory structure and signal regularity of the basin in question, rather than on algorithmic complexity alone. © 2026 The Authors.