Long-run forecasting surface and groundwater dynamics from intermittent observation data: An evaluation for 50 years.

Area of Science:

• Hydrology and Water Resource Management
• Artificial Intelligence in Environmental Science
• Geospatial Data Analysis

Background:

• Accurate prediction of water dynamics is crucial for effective water resource management.
• Existing forecasting methods often struggle with long-term predictions and data gaps.
• Need for a robust system to forecast river levels, river discharges, and groundwater levels.

Purpose of the Study:

• To propose and evaluate a novel approach for long-term daily water dynamics forecasting (7-30 days).
• To enhance prediction accuracy and consistency using advanced neural network techniques.
• To address challenges of missing measurements and gauge installations in long-term operational forecasting.

Main Methods:

• Utilized a state-of-the-art bidirectional long short-term memory (BiLSTM) neural network.
• Developed an adaptive scheme for regular network adjustment and re-training to handle changing inputs.
• Leveraged a 50-year in-situ database from Normandy, France, covering rivers, a karst aquifer, and meteorological data.

Main Results:

• Achieved high accuracy and consistent predictions for river levels, discharges, and groundwater levels.
• Demonstrated prediction errors of approximately 3% for 7-day-ahead and 6% for 30-day-ahead forecasts.
• Successfully filled data gaps, detected anomalies, and revealed the influence of physical processes on prediction performance.

Conclusions:

• The BiLSTM-based data-driven model offers a unified and accurate approach for predicting multiple water dynamics.
• The system's performance is influenced by the physical characteristics of the water dynamics, such as groundwater's slow filtration.
• This novel approach significantly improves operational water resource management through reliable long-term forecasting.