Interpretable streamflow prediction in headwater area of the Yellow River using a climate-adaptive transformer and...

Water scarcity remains the central challenge in the Yellow River Basin. Its headwaters, a key source of streamflow, plays a critical role in regional water supply. To address hydrological uncertainty under climate change, this study proposes a Climate-Adaptive Transformer (CAT) model for streamflow prediction. The proposed framework demonstrates four key advantages: (i) achieves high predictive accuracy using only temperature and precipitation as meteorological inputs; (ii) physically consistent and stable outputs under extreme climate scenarios, including +5 °C warming and doubled precipitation; (iii) robust future streamflow projections across four Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5) based on seven General Circulation Models (GCMs); and (iv) provides region-specific interpretability through Shapley Additive exPlanations (SHAP) analysis, supported by independent soil moisture validation. The model achieved a Nash-Sutcliffe Efficiency (NSE) of 0.8 across the training, validation, and testing sets, consistently at both daily and monthly time scales. The average of the seven GCMs projects an overall increase in annual streamflow at the Tangnaihai station through 2100. Low-to mid-emission scenarios (SSP1-2.6 and SSP2-4.5) yield larger and more stable streamflow increases, while high-emission scenarios (SSP3-7.0, SSP5-8.5) show greater variability and can decline below historical levels. Projected annual streamflow under the four SSPs is 22.14, 21.59, 20.54, and 21.51 × 10⁹ m³ yr^-1, respectively. SHAP analysis identifies minimum temperature as the dominant driver of streamflow variation, particularly in cold, low-elevation areas like Ruoergai. The CAT model offers a practical solution for streamflow prediction and water resource planning in climate-sensitive, data-limited regions.