Hybrid machine learning for predicting groundwater level: A comparison of boosting algorithms with neural networks

This study proposes a novel hybrid machine learning framework that integrates gradient boosting (XGBoost, LGBM) and neural network models (LSTM, MLP) with Basin Hopping Optimization (BHO) to improve groundwater level forecasting. The approach simultaneously optimizes input lag times and model hyperparameters, addressing a key limitation in previous studies. Four hybrid models (XGBoost-BHO, LGBM-BHO, LSTM-BHO, MLP-BHO) are evaluated for daily one-to seven-day-ahead predictions, incorporating meteorological inputs. Results showed that all models achieved high predictive accuracy (R 2 > 0.98), with LSTM-BHO yielding the lowest MAE and RMSE across both boreholes. Boosting models, particularly XGBoost-BHO, demonstrated strong short-term performance with narrow residual distributions and significantly lower computation time. These findings highlight the effectiveness of combining machine learning and metaheuristic optimization for robust groundwater forecasting. • A new approach for the prediction of groundwater level (GWL) is proposed. • The study provides insights for hybrid modelling of GWL forecasting for up to 7 days. • The study provides multivariate forecasting utilizing metrological data. • Joint optimization of lag and model parameters is perform using Basin Hopping • A Comparison of 4 hybrid models including boosting and neural network is provided.