A comparative study among machine learning and numerical models for simulating groundwater dynamics in the Heihe River Basin, northwestern China.

Groundwater is unique resource for agriculture, domestic use, industry and environment in the Heihe River Basin, northwestern China. Numerical models are effective approaches to simulate and analyze the groundwater dynamics under changeable conditions and have been widely used all over the world. In this paper, the groundwater dynamics of the middle reaches of the Heihe River Basin was simulated using one numerical model and three machine learning algorithms (multi-layer perceptron (MLP); radial basis function network (RBF); support vector machine (SVM)). Historical groundwater levels and streamflow rates were used to calibrate/train and verify the different methods. The root mean square error and R were used to evaluate the accuracy of the simulation/training and verification results. The results showed that the accuracy of machine learning models was significantly better than that of numerical model in both stages. The SVM and RBF performed the best in training and verification stages, respectively. However, it should be noted that the generalization ability of numerical model is superior to the machine learning models because of the inclusion of physical mechanism. This study provides a feasible and accurate approach for simulating groundwater dynamics and a reference for model selection.