Groundwater level response identification by hybrid wavelet-machine learning conjunction models using meteorological...

Area of Science:

• Hydrogeology and Water Resource Management
• Environmental Science and Engineering
• Computational Hydrology

Background:

• Groundwater level (GWL) prediction is vital for managing complex aquifer systems.
• Heterogeneous aquifers present challenges for accurate GWL modeling.
• Supervised machine learning (ML) offers potential for improved GWL forecasting.

Purpose of the Study:

• To develop and compare supervised ML models for delineating GWL changes.
• To assess the efficacy of wavelet transform (WT) in enhancing ML-based GWL prediction.
• To identify the optimal ML model and input variables for accurate GWL forecasting in the Zarand-Saveh aquifer.

Main Methods:

• Utilized a 15-year monthly dataset (2005-2020) for the Zarand-Saveh aquifer, Iran.
• Implemented and compared four standalone ML models: Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Group Method of Data Handling (GMDH), and Least Square Support Vector Machine (LSSVM).
• Applied Wavelet Transform (WT) in conjunction with ML models to improve predictions using precipitation, evapotranspiration, temperature, and GWL data for 1- and 3-month horizons.

Main Results:

• Hybrid wavelet-ML models significantly outperformed standalone ML methods.
• The Wavelet Transform-Least Square Support Vector Machine (WT-LSSVM) model demonstrated superior performance.
• The best predictions were achieved with WT-LSSVM using all influential variables, yielding high accuracy (e.g., R=0.99, NSE=0.99 for 1-month ahead).

Conclusions:

• Hybrid models integrating Wavelet Transform with ML substantially enhance groundwater level prediction accuracy.
• The WT-LSSVM model is highly effective for forecasting GWL in complex aquifer systems.
• Accurate GWL prediction is achievable using optimized ML approaches with comprehensive input data.