Physics-enhanced machine learning models for streamflow discharge forecasting

Accurate river discharge forecasts for short to intermediate time intervals are crucial for decision-making related to flood mitigation, the seamless operation of inland waterways management, and optimal dredging. River routing models that are physics based, such as RAPID ('routing application for parallel computation of discharge') or its variants, are used to forecast river discharge. These physics-based models make numerous assumptions, including linear process modeling, accounting for only adjacent river inflows, and requiring brute force calibration of hydrological input parameters. As a consequence of these assumptions and the missing information that describes the complex dynamics of rivers and their interaction with hydrology and topography, RAPID leads to noisy forecasts that may, at times, substantially deviate from the true gauged values. In this article, we propose hybrid river discharge forecast models that integrate physics-based RAPID simulation model with advanced data-driven machine learning (ML) models. They leverage runoff data of the watershed in the entire basin, consider the physics-based RAPID model, take into account the variability in predictions made by the physics-based model relative to the true gauged discharge values, and are built on state-of-the-art ML models with different complexities. We deploy two different algorithms to build these hybrid models, namely, delta learning and data augmentation. The results of a case study indicate that a hybrid model for discharge predictions outperforms RAPID in terms of overall performance. The prediction accuracy for various rivers in the case study can be improved by a factor of four to seven.