[Objective] Water level variations in tidal river reaches are highly complex and influenced by both upstream runoff and downstream tides. Accurate prediction of these variations is crucial for flood control, disaster reduction and ecological security. Taking the Nanjing Hydrological Experimental Station in the tidal reach of the Yangtze River as a case study, this paper aims to develop a model that can capture the nonlinear, dynamic and complex fluctuation characteristics of water levels and achieve real-time prediction. [Methods] A hybrid CNN-LSTM-KAN model is proposed, which combines a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and a Kolmogorov-Arnold Network (KAN). First, the CNN is used to extract spatial features from the water level sequence. Then, the LSTM captures the dynamic temporal dependencies of water level changes. Finally, KAN is introduced to further enhance the model’s ability to represent nonlinear and dynamic characteristics. [Results] The hybrid model shows strong real-time prediction performance on the measured data from the Nanjing station. It outperforms conventional deep learning methods such as optimized LSTM and CNN-LSTM in terms of prediction accuracy, peak capture capability and robustness to input data. On the test set, the model achieves a root mean square error (RMSE) of 0.065 4 m, a mean absolute error (MAE) of 0.042 9 m, a mean absolute percentage error (MAPE) of 0.023 5, and a Nash-Sutcliffe efficiency coefficient (NSE) of 0.995 1. [Conclusions] The proposed CNN-LSTM-KAN hybrid model has a simple structure and is easy to implement. It provides strong technical support for flood and drought disaster prevention and sustainable socio-economic development in the tidal reaches of the lower Yangtze River.