Abstract: The water cut curve of production wells in water-flooded reservoirs can reflect water front advancement, injection-production connectivity, and remaining oil production status, serving as one of the core response indicators for reservoir dynamic analysis and development plan adjustment. To balance the physical consistency of numerical simulation and the rapid prediction capability of machine learning models, this paper proposes a hybrid Quantum Kolmogorov-Arnold and Long Short-Term Memory network (QKAN-LSTM) method for predicting water cut of reservoir production wells. Specifically, the Long Short-Term Memory (LSTM) network is adopted to extract temporal dependencies embedded in injection-production historical data, while the Quantum Kolmogorov-Arnold Network (QKAN) is utilized to characterize the nonlinear mapping relationships among reservoir physical parameters, injection-production regimes and water cut responses. First, the MRST simulator is employed to generate physically consistent labeled datasets for two-phase oil-water flow, and two numerical test cases are designed to verify the model performance. Case 1 covers 2, 000 sample groups with varied reservoir parameters and injection-production schedules to examine the cross-parameter generalization ability of the proposed model. Case 2 adopts a modified fixed inverted nine-spot well pattern, where supervised training and post-stage prediction are carried out on 1, 000 sample groups of random injection-production regimes, with three typical production well time series selected to demonstrate the prediction performance. The results indicate that the mean absolute error (MAE) on the validation set of Case 1 is approximately 0.0016, and the predicted values for unseen parameter combinations are in good agreement with MRST simulation outputs. Typical samples from Case 2 prove that the QKAN-LSTM model can accurately capture water cut evolution in both fitting and forecasting stages under a fixed well pattern and random injection-production constraints. The above findings demonstrate that the proposed model is capable of learning dynamic water cut responses within the scope of the simplified physical model and parameter settings adopted in this work, providing an expandable surrogate modeling paradigm for rapid evaluation of water-flooding development schemes.