Abstract: To address the insufficient prediction accuracy and hysteresis effect of nonlinear deformation of reservoir bank slopes under reservoir water level fluctuation, taking the Baijiabao landslide as the research object, this study proposes a deformation prediction model based on Physics-guided Neural Network (PGNN). Firstly, numerical simulation is adopted to reveal the seepage hysteresis and stability evolution mechanism of landslides induced by reservoir water level variation. The results show that within the annual reservoir operation cycle, the safety factor of the slope reaches the minimum value when the reservoir water level first rapidly drops to the low level, accompanied by the weakest external support and the strongest pore water pressure hysteresis effect, which corresponds to the most unfavorable stability state of the slope. On this basis, the PGNN takes the multi-layer perceptron as its baseline network. Combined with soft threshold gating and a multi-objective loss function, it realizes the decoupling and prediction of fundamental deformation and nonlinear step deformation based on the threshold of the safety factor. Compared with the baseline model, LSTM, and GRU, the average root mean square error of prediction at three monitoring points of the PGNN model is reduced by 61.7%, 27.3% and 38.4% respectively, the average coefficient of determination is improved to 0.84, and the mean bias error is close to zero. Meanwhile, the critical safety factor threshold triggering landslide initiation is determined via simultaneous inversion as 1.091, which can provide a quantitative reference for early warning judgment and risk management and control of reservoir landslides.