Artificial Neural Networks for Coastal Aquifer Conductance Prediction | JoVE Visualize

Area of Science:

Hydrological modeling using machine learning
Coastal aquifer management in environmental engineering

Background:

Traditional models for predicting aquifer conductance require detailed hydrological parameters, which are not always available. This gap motivated the use of artificial neural networks (ANNs) to predict conductance from more accessible variables. Prior research has shown that physical-based models depend on hydraulic conductivity and fluid density, but these are difficult to measure in unconfined coastal aquifers. No prior work had resolved how easily measurable variables could replace these inputs. Climate and pumping changes affect brackish-freshwater interfaces, but modeling these effects is complex. This uncertainty drove the need for alternative predictive tools. Existing models may not capture time-variable conductance accurately. The lack of a simpler approach for aquifer management under changing conditions remains a challenge. This paper introduces a novel method using ANNs to predict conductance based on environmental and operational data.

Purpose Of The Study:

The study aimed to develop ANNs to predict specific conductance in a coastal aquifer using easily measurable variables. The researchers propose that ANNs can replace traditional models that require detailed hydrological inputs. The specific problem is the difficulty in modeling time-variable conductance due to brackish zone displacement. The motivation is to provide a simpler and accurate prediction tool for aquifer management. Traditional models may not adapt well to variable pumping and climate conditions. The researchers propose that ANNs can learn system behavior from accessible data. This approach could help minimize groundwater quality degradation. The study focuses on using ANNs to predict salinity at a monitoring well near a municipal well.

Main Methods:

The study used artificial neural networks to predict specific conductance in a coastal aquifer. Input variables included initial conductance, precipitation, temperature, and pumping extraction. The researchers trained ANNs using historical data from a monitoring well. The model predicted salinity over time periods ranging from 30 days to several years. Unlike physical models, ANNs do not require hydraulic conductivity or fluid density inputs. The study compared ANN predictions with measured and interpolated values. Linear regression was used as a benchmark for model accuracy. The researchers also conducted a sensitivity analysis to assess variable importance.

Main Results:

ANNs predicted specific conductance with an absolute mean error of 1.1% over 90-day periods. The average percent change in conductance was 39% over the same period. Predictions matched measured and interpolated values closely. The model outperformed linear regression in accuracy. The sensitivity analysis showed the importance of each input variable. Precipitation, temperature, and pumping extraction influenced conductance. Initial conductance also played a significant role in predictions. These results suggest that ANNs can accurately model complex aquifer dynamics.

Conclusions:

The study concludes that ANNs can accurately predict conductance in unconfined coastal aquifers. The authors suggest that ANNs are a powerful tool for managing groundwater quality. The results indicate that ANNs can replace traditional models requiring detailed hydrological inputs. The sensitivity analysis provided insights into variable importance. The researchers propose that ANNs can help identify optimal pumping policies. The model's accuracy suggests it can minimize aquifer degradation. The study supports using ANNs for adaptive aquifer management under variable conditions. The authors suggest that this approach can be applied to other coastal aquifers.