A study on spatiotemporal groundwater level forecasting by a hybridization of machine learning and physically-based models

Water resources are essential for human survival. Moreover, effective management and exploration of water resources can bring more opportunities for achieving sustainable development goals (SDGs), i.e., SDG 11 and SDG 12. Therefore, developing a soft computing groundwater forecast model enables us to manage groundwater resources and design a long-term sustainable water management plan. This study proposed a novel model (LSTM-HBV) that hybrides a machine learning model (LSTM) and a physically-based model (HBV-light model) to make accurate spatiotemporal groundwater level forecasts of the future three months for the Jhuoshuei River basin of Taiwan. The ten-day groundwater level, rainfall, and river flow factor datasets were collected from monitoring stations in the Jhuoshuei River basin during 2000 and 2019. A total of 767 ten-day datasets were collected and allocated into training (491, 64%), validation (123, 16%), and testing (153, 20%) stages. We aim to extract essential features to improve the forecast accuracy of the LSTM-HBV model. Besides that, the computational time of LSTM-HBV is shorter than that of the physically-based model, which makes the proposed LSTM-HBV model more applicable. Furthermore, the interpretability of LSTM-HBV allows us to understand the causal relationship between input and output, which breaks the limitations of black-box characteristics of machine learning techniques and hits a new milestone. Therefore, the understanding of the interactive mechanism between surface water and groundwater can assist in groundwater recharge by reservoir operation, which can make groundwater a sustainable water resource.