Comparing Physics-Based, Conceptual and Machine-Learning Models to Predict Groundwater Levels by BMA.

Area of Science:

• Hydrogeology and Water Resource Management
• Computational Hydrology
• Machine Learning in Environmental Science

Background:

• Groundwater level data are crucial for aquifer management and operational forecasting.
• Various modeling approaches exist, ranging from physics-based to data-driven methods.
• Assessing the performance of these diverse models is essential for effective water resource management.

Purpose of the Study:

• To compare the predictive performance of a 3D groundwater flow model (MODFLOW) against eigen-, transfer-function, and machine learning models (ML) for groundwater level forecasting.
• To evaluate if physics-based models offer advantages in integrated predictions across multiple wells compared to individually optimized data-driven models.
• To determine the suitability of different model classes for operational groundwater level prediction tasks.

Main Methods:

• Evaluated MODFLOW, eigenmodel, transfer-function model, multi-layer perceptron, long short-term memory, and random forest models.
• Tested models on four groundwater level time series from the Wairau Aquifer, New Zealand.
• Employed Bayesian model averaging to assess ensemble forecasting performance and derive model weights.

Main Results:

• Data-driven models demonstrated competitiveness for predicting individual groundwater wells, even outside calibration ranges.
• No single model universally outperformed others for all prediction scenarios.
• Physics-based MODFLOW was favored by Bayesian model averaging for simultaneous prediction of all four wells, indicating its strength in integrated system representation.

Conclusions:

• Data-driven models are viable for localized, single-well groundwater level predictions.
• Physics-based models retain significant advantages for integrated aquifer management and multi-well forecasting.
• Ensemble forecasting using Bayesian model averaging effectively combines diverse model strengths, highlighting the continued importance of physics-based approaches.