Forecasting change in groundwater storage: A multilayer perceptron model leveraging system dynamics simulation data

This study focuses on the Lower Rio Grande (LRG) region of Southern New Mexico, a semi-arid area facing significant water scarcity challenges. Groundwater serves as a critical resource for agriculture, domestic use, and industry in the region, necessitating robust management and predictive tools. The research develops a novel groundwater modeling approach by integrating System Dynamics (SD) and Machine Learning (ML) techniques. A validated SD model of the LRG provided simulation data, which was used to train an ML model employing a feedforward Multilayer Perceptron (MLP) as an artificial neural network (ANN) technique. This approach simplifies groundwater dynamics by concentrating on key variables identified through correlation analysis and simulation results. The model achieved a training RMSLE of 0.031 and R-squared of 0.77, with testing results showing RMSLE of 0.036 and R-squared of 0.71, demonstrating predictive reliability. The proposed model enables accurate forecasting of groundwater storage changes while reducing reliance on extensive data inputs. It pioneers using SD simulation data as a data augmentation method for MLP model training, enhancing predictive capabilities. This integrated methodology supports informed groundwater management and policy-making, offering a transferable framework for other hydrologically similar regions. • Introduces a hybrid System Dynamics–Deep Learning framework for groundwater forecasting. • Transfers socio-hydrological feedbacks from a validated SD model into a reduced-variable ANN. • Uses SD simulations as structured synthetic data to train deep learning under data scarcity. • Enables groundwater storage prediction without reliance on full SD or physical model execution. • Provides a deployable decision-support tool for regional groundwater management and policy.