Runoff prediction based on deep learning and interpretability on covariant relationship in the Yarlung Zangbo River Basin

Study region Yarlung Zangbo River Basin (YZRB). Study focus Site-specific LSTM models were developed to predict daily runoff in the YZRB using precipitation, air temperature, and relative humidity. SHAP attributions were compared across four seasons and between flood and non-flood periods to examine how the contributions of the same meteorological variables vary across stations and periods. New hydrological insights At all five stations, precipitation and relative humidity had larger mean absolute SHAP values during the flood season than the non-flood season. Temperature showed the opposite pattern at four stations, whereas Lhasa had larger temperature attributions during the flood season. At Lhasa, higher flood-season temperatures tended to have negative temperature SHAP values, while lower temperatures tended to have positive values. These differences show that temperature attributions cannot be interpreted uniformly across the YZRB and support examining runoff predictions by station and period. The attributions describe statistical associations learned by the models; possible links to meltwater supply and evapotranspiration require independent verification.