Performance comparison of physical process-based and data-driven models: a case study on the Edwards Aquifer, USA
Download Full Text
Physical process-based groundwater flow models are the major tools for studying fluid-flow behavior and for simulating the hydrological responses of water levels and spring discharge to human- and/or nature-induced triggers such as pumping and recharge. Such models are built with deep understanding of the physical processes and are based on geological models, developed by integrating data from geology, geophysics, and geochemistry. However, data-driven models can be built with limited data, eliminating the need for a detailed understanding of the physics. In this research, a data-driven model is built for simulating hydraulic responses (both groundwater levels and spring discharges) in a complex groundwater flow system of the Edwards Aquifer in Central Texas, USA, with the recurrent neural network (RNN) technique. The model is first trained and validated with the observation data of four targets—water levels from two index wells, and spring flow rates from San Marcos and Comal springs—from 2001 through 2015. The model is then used to predict the hydrological responses for the drought of record (1947–1958). The performance of the RNN model for the training, validation and prediction period is then quantitatively compared to that of the physical process-based MODFLOW model in terms of four statistical measures. The statistical measures suggest that the RNN model performs almost as well as the MODFLOW model. With further improvements, a data-driven model may be a surrogate to (or integrate with) a physical process-based model for simulating hydrological responses in the Edwards Aquifer.
Numerical modeling, Groundwater management, Physical process-based model, Recurrent neural network, USA
Digital Object Identifier (DOI)
Zhang, Andi; Winterle, James; and Yang, Changbing, "Performance comparison of physical process-based and data-driven models: a case study on the Edwards Aquifer, USA" (2020). KIP Articles. 6386.