Study region: The Daheihe River Basin, located in the arid region of northwestern China, experiences strong hydrological non-stationarity. Climate variability and intensive human interventions jointly alter runoff regimes. Limited engineering operation data further increase prediction uncertainty. This basin serves as a representative case for testing adaptive and interpretable runoff prediction approaches. Study focus: We propose an interpretable runoff prediction framework integrating physics-guided signal extraction and temporally adaptive feature selection. A calibrated semi-distributed hydrological model isolates human impacts and builds a learnable regulation signal. An ExpandingWindow Recursive Feature Elimination with Cross-Validation (EW-RFECV) strategy identifies key drivers as the training window expands annually. This achieves time-adaptive recognition of basin regulation dynamics. Uncertainties from both physical and machine-learning components are quantified to verify dual-component robustness. In testing, the downstream section achieved NSE = 0.68 and KGE = 0.84, exceeding a tuned LSTM by 0.56 and 0.76. Performance degradation remained below 8 % under +/- 10 % parameter and +/- 5 % input perturbations. New hydrological insights for the region: SHAP analysis revealed annual memory effects and threshold-based regulation behavior. When naturalized runoff exceeded 2.15 m3 /s, the system shifted from discharge to storage, consistent with Longsheng Reservoir observations. These findings confirm that the framework effectively captures hydrological-engineering coupling under data-scarce conditions. It provides a practical pathway toward interpretable runoff prediction and adaptive water management in arid basins.