A hybrid modeling framework for improved water quantity/quality simulation: using data-driven and process-based models

Abstract

Surface water and groundwater are crucial for environmental systems. However, the water quantity and quality have faced challenges in recent decades due to human activities and climate change. These challenges encompass water contamination from industrial chemicals in waterbodies and hydrological variations in watersheds, raising concerns about potential impacts on drinking water safety, ecosystems, and the economy. In response to these issues, the study proposes a hybrid modeling approach that combines process-based and data-driven methods, aiming to improve the accuracy and efficiency of simulating spatiotemporal water quantity and quality. Process-based models provide a detailed understanding of environmental processes but require extensive data and computation time. On the other hand, data-driven models are efficient for large datasets and excel in learning nonlinear relations; however, they can involve limitations in interpreting model predictions due to their black-box nature. The study bridges the gap between these two methods by coupling process-based models and utilizing process-based modeling outputs as datasets for training data-driven models. The proposed methodology includes a process-based modeling system coupling urban pipeline and hydrodynamic models to simulate chemical spill accidents in freshwater, analyzing the impact of pipeline transport on chemical concentration simulations (Chapter 3); a data-driven modeling framework using a convolutional neural network (CNN) to simulate chemical accidents, addressing computational challenges discussed in Chapter 3 (Chapter 4); a novel modeling framework that integrates CNN models with a fully distributed watershed model to estimate spatiotemporal conditions of groundwater and surface water, evaluating predictive performance against process-based models (Chapter 5); and an unstructured mesh-based modeling system using a graph neural network (GNN) to estimate spatiotemporal distributions of artificial sweeteners, overcoming limitations of traditional CNN models discussed in Chapter 5 (Chapter 6). In addition, the research conducts further analyses, such as simulations of various chemical spill cases, interpretation of deep learning model predictions using an explainable artificial intelligence (explainable AI) technique, optimization of hyperparameters, model architectures, and input designs for deep learning models, and examination of the influence of data transformation, dataset-splitting methods, and data assemblage on model performance. Our proposed models demonstrated satisfactory estimation performances through these analyses, providing rapid and computationally efficient simulations. While there is a need for further improvements in prediction performance, validation, and generalization of the models, our research findings can contribute to decision-making processes in water resource management, including pollution preparedness and response, risk assessment, and countermeasure formulation.