Improving real-time forecasting of water quality indicators with combination of process-based models and data assimilation technique

Abstract

Water quality indicators can be used to characterize the status and quantify and qualify the change of aquatic ecosystems under different disturbance regimes. Although many studies have been done to develop and assess indicators and discuss interactions among them, few studies have focused on how to improve the predicted indicators and explore their variations in receiving water bodies. Accurate and effective predictions of ecological indictors are critical to better understand changes of water quality in aquatic ecosystems, especially for the real-time forecasting. Process-based water quality models can predict the spatiotemporal variations of the water quality indicators and provide useful information for policy-makers on sound management of water resources. Given their inherent constraints, however, such process models alone cannot actually guarantee perfect results since water quality models generally have a large number of parameters and involve many processes which are too complex to be efficiently calibrated. To overcome these limitations and explore a fast and efficient forecasting method for the change of water quality indictors, we proposed a new framework which combines the process-based models and data assimilation technique. Unlike most traditional approaches in which only the model parameters or initial conditions are updated or corrected and the models are run online, this framework allows the information extracted from observations and outputs of process models to be directly used in a data-driven local/modified local model. The results from the data-driven model are then assimilated into the original process model to further improve its forecasting ability. This approach can be efficiently run offline to directly correct and update the output of water quality models. We applied this framework in a real case study in Singapore. Two of the water quality indicators, namely salinity and oxygen were selected and tested against the observations, suggesting that a good performance of improving the model results and reducing computation time can be obtained. This approach is simple and efficient, especially suitable for real-time forecasting systems. Thus, it can enhance forecasting of water quality indictors and thereby facilitate the effective management of water resources.