Evaluation of the effectiveness of Kolmogorov–Arnold networks for SSC prediction on the Mohawk River

Abstract The prediction of suspended sediment concentration (SSC) is of critical importance for the management of aquatic ecosystems, flood control, and erosion assessments. However, the continuous measurement of SSC is technically and logistically challenging. Consequently, reliable prediction models are of great value in replacing or supplementing field measurements. In this study, a range of regression models, as well as artificial neural networks (ANNs) and Kolmogorov–Arnolds networks (KANs) models, were employed to predict SSC data using USGS #01357500 Mohawk River at Cohoes, NY data. The data underwent a process of normalisation and scaling, after which they were divided into training, validation, and test subsets. The ANN and KAN models were trained using the following optimisation algorithms: stochastic gradient descent (SGD), ADAM, and Broyden–Fletcher–Goldfarb–Shanno (LBFGS). In the context of time series prediction, the incorporation of one, two, and three time-step lagged features within the data set is essential for capturing the impact of past values. The findings demonstrate that, in addition to the efficacy of the ANN model, the KAN model yielded the most accurate SSC prediction, with regard to both general data and peak values. The LBFGS algorithm, a prominent tool in the field of optimisation, yielded particularly successful outcomes when employed in the enhancement of the models. The incorporation of lagged values within the data set has been observed to exert a favourable influence on the prediction performance. It is important to note that processing the data in chronological order did not cause any decline in model performance. The findings indicate that the KAN method is a reliable and applicable model for SSC prediction.