Prediction of river streamflow time series is a necessity in hydrological studies and water resources management. Therefore, river flow prediction tools with reliable accuracy need to be developed. This study first proposed the Kolmogorov-Arnold network (KAN), and then tried to improve the baseline KAN model performance utilising convolutional KAN (CVKAN) technique. To test the accuracy of the developed methods, two river stations in the United States (USGS 09380000 and USGS 13269000) were chosen as a case study. In general, the CVKAN models performed better compared with KAN in river streamflow prediction. The optimal reductions in root mean square error (RMSE) were achieved as 7.978% at USGS 09380000 (CVKAN2 vs KAN2) and 3.813% at USGS 13269000 (CVKAN6 vs KAN6). The convolutional layers act as nonlinear local feature encoders, while the KAN component performs global functional approximation via learnable spline-based basis functions. This hybrid design enables simultaneous local feature abstraction and high-order functional representation, improving generalisation and nonlinear modelling capacity. To further evaluate the performance of KAN and CVKAN models, two machine learning (ML) schemes, including multi-layer perceptron (MLP) and extreme learning machine (ELM), as well as two deep learning (DL) techniques, namely long short-term memory (LSTM) and gated recurrent unit (GRU) were also implemented. The findings exhibited that the KAN and CVKAN models generally outperformed both the ML and DL frameworks in predicting river streamflow. It was also concluded that the ML-based MLP and ELM illustrated competitive results with KAN and CVKAN models. Shapley additive explanations (SHAP) were finally applied to determine the contributions of input predictors. The outcomes indicated that the first lag of river streamflow represented the highest contributions in river streamflow prediction.
Gharehbaghi et al. (Mon,) studied this question.