Redefining inland water temperature forecasting with advanced neural architectures

Water temperature is a key driver of ecological and biogeochemical processes in inland waters, and the reliable forecasting of this variable is essential for effective environmental management. This study evaluates a hybrid Kolmogorov–Arnold Network (KAN) and a Bidirectional Long Short-Term Memory (Bi-LSTM) model for forecasting river water temperature (RWT) and lake surface water temperature (LSWT) in the Pannonian Ecoregion (Hungary). The models were trained using air temperature and historical water temperature and tested across forecasting horizons from 1 to 7 days. Their performance was compared with two widely used empirical benchmarks, air2stream and air2water. Both deep learning models substantially outperform the benchmark approaches for 1-day-ahead forecasts and retain strong predictive skill across longer horizons. Although accuracy decreases with lead time, both architectures preserve high correlation and low bias up to 7 days ahead. Across both mean conditions and temperature extremes, the hybrid KAN and Bi-LSTM exhibit closely aligned performance, with the hybrid KAN providing comparable accuracy at a fraction of the computational cost. These results demonstrate that hybrid spline-based and recurrent neural networks provide an effective and efficient framework for inland water temperature forecasting, supporting both short-term operational applications and multi-day environmental monitoring. • Performance of KAN and Bi-LSTM in inland water temperature modeling was evaluated. • KAN and Bi-LSTM models significantly outperformed the benchmark models, air2stream and air2water. • The KAN and Bi-LSTM exhibit closely aligned performance. • Model performance declined with increasing forecasting horizons.