DRSTF: A hybrid-approach framework for reservoir water temperature forecasting considering operation response

Abstract

Reservoir water temperature forecasting typically relies on meteorological factors as inputs and often overlooks the influence of reservoir operation. Moreover, the accuracy of reservoir water temperature forecasting and dispatch management is limited by poor-quality samples and the effective capturing of relevant information. In this study, we propose a Dimensionality Reduction and Simulation Training Framework (DRSTF) that generates training data using CE-QUAL-W2 (W2). Using the maximal information coefficient and an autoencoder, the framework reconstructs significant patterns of reservoir water temperature influenced by meteorological and operational factors from historical data (7 days preceding the target date). This process guides machine learning models for forecasting reservoir water temperatures. Three schemes (simulation accuracy, training scale, and reservoir operation scenario) were built to analyze the performance of the DRSTF. The results show that: (1) DRSTF performance is slightly affected by the accuracy of the simulation samples, and simulation training can attain forecasting precision comparable to observation training; (2) DRSTF exhibits a low training size requirement; compared to training with all samples (731 days), the mean absolute error (MAE) value when training on a limited number of samples (73 days) increases by only 10.8 %; and (3) DRSTF successfully forecasts the water temperature distribution, with the MAE reaching 0.266 °C, influenced by flood season reservoir scheduling. The proposed DRSTF can overcome constraints related to the quality, quantity, and dimensionality of hydrological and meteorological data, thereby achieving efficient forecasting of reservoir water temperatures, while accounting for the influence of reservoir operation.