Asymmetric generalization and failure mechanisms in zero-shot alpine runoff simulation: The role of hydrological memory heterogeneity

Study region Alpine headwaters of the Kaidu River (central Tianshan) and Keriya River (northern Kunlun), Northwest China. Study focus We evaluated seasonal performance and bidirectional zero-shot transfer of seven data-driven models and an ensemble with validation-optimized weights for current-month runoff simulation. Models used concurrent and antecedent hydrometeorological and land surface inputs without forecast lead time. Analyses covered seasonal diagnostics, a monthly climatology benchmark, lookback tests (3, 6, 9 months) for four deep learning models, and controlled comparisons of mean squared error (MSE) with hydrology-aware loss. Transfer used fixed source-basin configurations. New hydrological insights for the region Input windows varied by basin and architecture: all four Keriya networks favored 3 months versus 6–9 in Kaidu. These preferences reflect input sensitivity, not measured storage timescales. In Keriya, monthly climatology nearly matched the ensemble in Nash–Sutcliffe efficiency (NSE: 0.791 versus 0.804). Negative NSE in small, low-variance snowmelt and baseflow subsets coincided with modest absolute errors. Hydrology-aware loss improved NSE for Kaidu LSTM and DLinear but reduced it in six other comparisons. Transfer exposed another limitation: ensemble NSE was 0.722 from Keriya to Kaidu and 0.357 in reverse. During transfer to Kaidu, XGBoost exceeded the ensemble (NSE = 0.750), whereas TCN retained high correlation ( r = 0.930) but exaggerated simulated variability (ratio = 2.06; NSE = −0.417).