A hybrid analytical and machine learning framework for groundwater resources management

In Hydrogeology, numerical models are presented as essential tools for integrating, understanding, and predicting groundwater processes. However, these models face significant challenges: on the one hand, boundary conditions and hydraulic parameters are often subject to a large degree of uncertainty, and, on the other hand, numerical models usually require advanced solving and calibration techniques that generally imply long runtimes. Recently, innovative machine learning models have emerged as a promising alternative to address these issues and thus, the application of artificial intelligence in hydrology has increased significantly. In this study we present a hybrid model designed to predict groundwater heads in response to pumping. This model generates an initial analytical approximation of groundwater heads which is later enhanced by a machine learning framework based on recurrent neural networks. A real application of a pumping field for urban supply in Spain is presented as an illustration of the practical application of the presented methodology. Following model training and validation, we have also integrated a genetic algorithm to optimise flow rates, aiming to minimise energy consumption and/or head drawdowns. The results reveal that our hybrid approach achieves highly accurate head predictions with normalised absolute mean error lower than 4% which implies that the model reproduces properly the head measurements. Additionally, the optimisation algorithm successfully reduces energy consumption by 25%. This methodology represents a groundbreaking approach to quantify the effects of intense pumping and to facilitate long-term management of groundwater resources.