Deep Learning for Long-term Maximum Water Level Prediction in the Iset River

This paper presents a numerical analysis of applying deep learning to develop a long-term method for maximum water level forecasting at several gauging stations on the Iset River (Kataysk, Shadrinsk, Mekhonskoe). Two neural network architectures were analyzed in detail using the same set of initial hydrometeorological data. The results show that the N-HiTS architecture improves forecast accuracy on the validation set compared to the previously used TFT architecture. Cross-validation estimates of classical statistical criteria confirm the method's applicability in operational practice for all three gauging stations. Additionally, groundwater level data from observation wells of the Rosnedra (Federal Agency on Subsoil Use) system were analyzed. The results indicate that in some cases, incorporating such data can significantly improve forecast accuracy.