The changes in climate and watershed properties are reshaping hydrological processes worldwide at multiple scales, leading to a transition from steady to non-stationary states. The traditional Budyko framework commonly relies on a steady-state assumption and neglects terrestrial water storage changes, thereby exhibiting notable limitations at finer temporal scales. Hence, a novel runoff elasticity attribution model was developed by integrating a modified Budyko framework (incorporating parameter lambda for watershed properties and y0 for the net effect of terrestrial water storage dynamics), Random Forest, and a CNN-LSTM-Attention neural network. Then, the spatial differences in seasonal runoff variations were revealed by the proposed runoff elasticity attribution model in the Yellow River Basin during 1960-2020. The results show that the CNN-LSTM-Attention model can effectively interpret the parameters lambda and y0, with determination coefficients (R2) exceeding 0.9 between simulated and calibrated values. Precipitation and terrestrial water storage exert positive controlling effects on runoff, whereas potential evapotranspiration and underlying surface changes suppress runoff generation. In the source region of the Yellow River Basin, climatic factors (precipitation and potential evapotranspiration) contribute 33.9% to the overall runoff variation. The influence of underlying surface changes on runoff gradually increases along the downstream direction, accounting for 68.6%, 80.0%, and 87.6% of the variation in the upper, middle, and lower reaches, respectively. Seasonal runoff in arid regions shows greater sensitivity to the net effect of terrestrial water storage dynamics, accounting for 9.0% of the runoff variation in the upper Yellow River. A new methodological framework is provided in this study for diagnostic attribution and adaptive management of watershed water resources under evolving environmental conditions.