Groundwater level forecasting using ensemble coactive neuro-fuzzy inference system.

Area of Science:

• Hydrology and Water Resources
• Artificial Intelligence in Environmental Science
• Geoscience Modeling

Background:

• Accurate groundwater level (GWL) forecasting is crucial for water resource management.
• Existing models often struggle with long-term predictions and uncertainty quantification.
• The need for advanced AI techniques in hydrological forecasting is growing.

Purpose of the Study:

• To develop and evaluate a coactive neuro-fuzzy inference system (CANFIS) for multi-lead time GWL forecasting.
• To investigate the impact of various input variables and training data durations on forecasting accuracy.
• To enhance model performance using ensemble learning and assess prediction uncertainty.

Main Methods:

• Developed a CANFIS modeling framework for GWL forecasting in Texas and Florida.
• Utilized singular spectrum analysis (SSA) for input variable selection (GWL, precipitation, temperature, surface water level).
• Applied bagging ensemble learning and bootstrap sampling for uncertainty analysis.

Main Results:

• CANFIS achieved satisfactory daily GWL forecasts, with potential for improvement in monthly forecasts (2-3 months ahead).
• Temperature was identified as a significant variable for monthly forecasting.
• Ensemble CANFIS significantly enhanced monthly forecasting performance.
• Daily forecast uncertainty was satisfactory; monthly forecasts exhibited higher uncertainty, especially during fluctuating GWL periods.

Conclusions:

• The CANFIS model, especially with ensemble learning, shows promise for GWL forecasting, particularly for shorter lead times.
• Optimal input combinations and training data duration are critical for model performance.
• Uncertainty analysis highlights the challenges in long-term hydrological predictions and the need for careful interpretation of results.