Improving Forecasting Accuracy of Multi-Scale Groundwater Level Fluctuations Using a Heterogeneous Ensemble of Machine Learning Algorithms

Abstract

Accurate groundwater level (GWL) forecasts are crucial for the efficient utilization, strategic long-term planning, and sustainable management of finite groundwater resources. These resources have a substantial impact on decisions related to irrigation planning, crop selection, and water supply. This study evaluates data-driven models using different machine learning algorithms to forecast GWL fluctuations for one, two, and three weeks ahead in Bangladesh’s Godagari upazila. To address the accuracy limitations inherent in individual forecasting models, a Bayesian model averaging (BMA)-based heterogeneous ensemble of forecasting models was proposed. The dataset encompasses 1807 weekly GWL readings (February 1984 to September 2018) from four wells, divided into training (70%), validation (15%), and testing (15%) subsets. Both standalone models and ensembles employed a Minimum Redundancy Maximum Relevance (MRMR) algorithm to select the most influential lag times among candidate GWL lags up to 15 weeks. Statistical metrics and visual aids were used to evaluate the standalone and ensemble GWL forecasts. The results consistently favor the heterogeneous BMA ensemble, excelling over standalone models for multi-step ahead forecasts across time horizons. For instance, at GT8134017, the BMA approach yielded values like R (0.93), NRMSE (0.09), MAE (0.50 m), IOA (0.96), NS (0.87), and a-20 index (0.94) for one-week-ahead forecasts. Despite a slight decline in performance with an increasing forecast horizon, evaluation indices confirmed the superior BMA ensemble performance. This ensemble also outperformed standalone models for other observation wells. Thus, the BMA-based heterogeneous ensemble emerges as a promising strategy to bolster multi-step ahead GWL forecasts within this area and beyond.