Deep learning as a refining tool for hydrodynamic model in predicting lake temperature profile: A case study in Ogouchi reservoir

The advent of big data has opened doors to valuable insights into water resources, harnessing its power to develop a better paradigm, made possible by a combination of process-based and data-driven models. This coupling has the potential to increase prediction accuracy while maintaining the underlying mathematical cause-and-effect relationships. In this study, a hybrid model, a combination of a simple hydrodynamic model (a baseline model), which contains coarse grid resolution and has uncertainties of boundary conditions, and a simple Long Short-Term Memory (LSTM) model, was developed to improve the prediction of vertical temperature profile (VTP) in the Ogouchi reservoir based on a data assimilation technique. Comparisons were conducted among the two simple models, a fully calibrated hydrodynamic model (a reference model), and the hybrid model in predicting the weekly, biweekly (every two weeks), and triweekly (every three weeks) VTP in the Ogouchi reservoir during the summer periods. While both the simple models presented errors at certain depths, the hybrid model successfully mitigated these systematic biases and yielded performance comparable to the reference model. Additionally, the hybrid model significantly reduced computational costs and calibration efforts in contrast to the reference model. As such, this new paradigm not only integrates physical process outputs but also corrects specific discrepancies that could be time-consuming to calibrate in a standalone hydrodynamic model. • Long Short-Term memory model predicts temperature well at the surface layers. • Hydrodynamic model captures deep-water temperatures well despite simplified boundaries. • Hybrid technique improves weekly, bi-weekly, and tri-weeklytemperature predictions. • Quantifying the impact of meteorological variables explains LSTM superiority in surface layer predictions.