2026 Volume 7 Issue 3 Pages 694-700
In this study, we introduced an RNN incorporating HiPPO blocks (HiPPO+RNN) as a new method for predicting river temperature and evaluated its performance the Takatsu River in Shimane Prefecture. Specifically, we optimized the basis functions and their number (HiPPO size) in the HiPPO+RNN model and compared it with an existing river temperature prediction model using LSTM. The results indicated that the choice of basis function and HiPPO size significantly affected prediction accuracy. For river temperature prediction, which involves multiple temporal scales ranging from seasonal to daily variations, the model using the Laguerre basis with a HiPPO size of 20 achieved the highest accuracy. Furthermore, comparison with the LSTM-based river temperature prediction model revealed an improvement in prediction accuracy, particularly during flood periods from June to August, as the maximum RMSE decreased from approximately 3 °C for the LSTM model to less than 2 °C for the HiPPO+RNN model. These results demonstrate that HiPPO+RNN enables accurate basin-wide river temperature prediction.