A machine learning framework for multi-step-ahead prediction of groundwater levels in agricultural regions with high reliance on groundwater irrigation

This study presents a machine learning framework for multi-step-ahead prediction of groundwater levels in agricultural regions heavily reliant on groundwater irrigation. The framework utilizes a comprehensive set of predictive factors, including meteorological, hydrological, and human activity data. An optimal combination of input variables and their temporal delays was determined using a novel selection method. To address overfitting, a mathematical model for hyperparameter optimization was developed, leveraging sample subset cross-validation and an improved differential evolution algorithm. Numerical experiments on the YingGuo region in the Huaihe River Basin demonstrated that the hyperparameter optimization resulted in an 11.6%–38.5% increase in the Nash-Sutcliffe Efficiency (NSE) indicator. Additionally, fine-tuned temporal scales, from monthly to five-day resolution, significantly improved predictive performance, with NSE increasing from 0.629 to 0.952 (33.9% enhancement). However, longer forecasting horizons led to a 29.4% reduction in NSE. The study also implemented a multi-core parallel computing framework, which achieved a 15.35-fold improvement in computational efficiency while maintaining predictive precision. The integration of external factors enhanced the predictive performance across various observation wells. These findings contribute to a better understanding of groundwater dynamics and highlight the potential of machine learning models in improving groundwater depth predictions in agricultural regions with high reliance on groundwater irrigation.