Surface water and groundwater, integral to the hydrological cycle, engage in complex hydraulic interactions and frequent transformations. Isolating surface water and groundwater systems in individual studies often fails to capture and analyse their interrelationships, limiting the comprehensive understanding of regional water resources. Additionally, conventional physics-based coupled models encounter challenges arising from the complexities and non-linearity of interactions, impeding their accuracy in simulation results.
To address this challenge, this thesis proposes a novel framework that integrates artificial intelligence and physics-based coupled models to simulate variations in surface water and groundwater, establishing a foundation for integrated water resource management. Specifically, the study develops a boundary-coupled framework to model interactions between surface water and groundwater. In this framework, a data-driven deep learning model is employed to simulate surface water flow. Additionally, physics-based analytical models are used to describe groundwater movement in riparian zones, while simplifying river behaviour to a Dirichlet boundary condition to assimilate data from the surface water model. Subsequently, the simulated values from analytical solutions serve as the source data, while groundwater observation data is employed as the target data. A transfer learning model is then be utilized to learn the features of the source data and, in conjunction with the target dataset, facilitate the prediction and regression of groundwater. Finally, the framework is applied at the watershed scale to predict and model catchment-scale surface water flow and groundwater head.
In this framework, the thesis assesses the influence of various input variables on surface water prediction, explores the effect of groundwater layer heterogeneity, and validates the effectiveness of the deep transfer learning approach, particularly in catchment-scale predictions. The main conclusions are as follows:
1. The selection of model inputs greatly influences accuracy. The PCA method effectively enhances the precision of the deep RNN model, especially in scenarios with numerous input variables. It achieves this by distilling essential information, categorizing original data into several comprehensive variables.
2. The two-layer structure significantly influences groundwater flow responses to hydrological events. During recharge events with a less permeable upper layer, lateral discharge to the river is hindered, directing more groundwater downward into the more permeable lower layer. Conversely, when the upper layer is more permeable, greater lateral flow into the river occurs, with less downward flow into the less permeable lower layer. During a flood event with a less permeable upper layer, river water predominantly infiltrates the more permeable lower layer initially, then flows upward into the upper layer, creating a vertical flow. The direction of this flow reverses during the recession period. However, this phenomenon is not evident when the upper layer is more permeable than the lower layer.
3. The transfer learning method can enhance the capacity of analytical solutions for heterogeneous aquifers. By integrating analytical knowledge with the neural network, the analytical solution-transfer learning method significantly improves hydraulic head prediction accuracy. Even for very sparse training data, the analytical solution-transfer learning method still performs more satisfactorily than the traditional deep learning method.
4. The analytical solution-transfer learning method is also effective at the catchment scale. The analytical solution-transfer learning method can obtain more accuracy and robust results than traditional deep learning methods with the same training dataset.
