Prediction of spring flows using nonlinear autoregressive exogenous (NARX) neural network models

Area of Science:

• Hydrology
• Environmental Science
• Artificial Intelligence

Background:

• Climate change in the Mediterranean causes frequent droughts, depleting groundwater resources.
• Effective groundwater management and sustainable development require accurate spring discharge prediction.

Purpose of the Study:

• To apply non-linear AutoRegressive with eXogenous inputs (NARX) neural networks for predicting spring discharge.
• To develop and evaluate discharge prediction models for 9 springs in the Umbria region, Italy.
• To assess the impact of precipitation as an exogenous input on prediction accuracy.

Main Methods:

• Utilized non-linear AutoRegressive with eXogenous inputs (NARX) neural networks.
• Developed models for 9 monitored springs in the Umbria region, incorporating precipitation data.
• Evaluated model performance for both short-term (1-month lag) and long-term (12-month lag) predictions.

Main Results:

• Achieved good prediction performances for all springs, with R² values ranging from 0.9005 to 0.9842.
• Demonstrated reliable forecasting for lag times up to 12 months.
• Assessed the sensitivity of forecasts to temporal resolution changes (weekly to monthly).

Conclusions:

• NARX networks are effective for spring discharge prediction in Mediterranean karst aquifer areas.
• The methodology provides a valuable tool for groundwater resource management.
• The study highlights the potential for applying NARX models in similar hydrogeological contexts.