Abstract: A data-driven water level forecasting method is constructed using Kolmogorov-Arnold Networks (KAN), which decomposes the complex relationships among hydrological variables into a linear combination of univariate functions, enabling accurate capture of the trends in hydrological data. The method has been applied to water level forecasting based on discharge and water level data from the Lianhuatang and Shashi stations in the middle and lower Yangtze River. Results show the KAN model has a seven-day mean absolute error of 0.187 m at Lianhuatang and 0.109 m at Shashi. In the case of Lianhuatang, it improves forecasting accuracy by 20.1%, 45.0%, 16.5%, and 13.0% compared to traditional Multi-Layer Perceptron, Long Short-Term Memory network, Gated Recurrent Unit network, and Transformer models, respectively. To deepen our understanding of this model further, sensitivity analysis and simplification tests are conducted. Results indicate its short-term upstream discharge forecasting significantly affects the predicted downstream water levels. Equipped with a minimal number of parameters, it achieves effectively the relationship between upstream discharge and downstream water level changes, demonstrating remarkable interpretability.

Key words: water level forecasting, Kolmogorov-Arnold networks, machine learning, middle reaches of the Yangtze River, interpretability