Neural Networks for Seasonal Groundwater Level Forecasting

Reliable and sustainable water resource management largely depends on the accurate prediction of groundwater levels, a challenge that has become increasingly critical due to climate change, and the growing exploitation of aquifers due to human activities . Reliable forecasts are essential for effective decision - making in water resource management, particularly in regions facing seasonal fluctuations or long - term declines in groundwater availability. The hidden nature of underground reservoirs makes it impossible to gain detailed knowledge of structural variability and water paths. Apart from large homogeneous basins, the physical properties of aquifers are thus poorly understood, making them difficult to model using physical or physically based models. Also, difficulties encountered in delimiting their feeding basin make it a challenge to fully understand them and accurately simulate and forecast their evolution. Thus, the lack of physical knowledge of underground hydrosystems encourages the use of systemic modeling methods, based on statistical approaches that do not require such knowledge. These include models based on artificial neural networks, particularly multilayer perceptrons (MLPs), that have proven to be a promising approach for improving the accuracy of hydrogeological forecasts. These models offer the advantage of capturing complex, nonlinear relationships between meteorological variables and groundwater level fluctuations, often outperforming traditional statistical methods. Although IA methods need a sufficient volume of data, they do not require extensive information about the hydrodynamics of the aquifer, making them easier to implement in poorly known contexts. In this study, we develop a seasonal groundwater level forecasting model for 30, 60, and 90 days ahead, for a piezometer located in the basement of the Sélune River watershed, in the Manche region of France. To achieve this, we use a multilayer perceptron (MLP) neural network while integrating regularization techniques to prevent overfitting and enhance generalization capabilities. The model is trained on a historical dataset covering a 15 - year period with a decadal time step, incorporating several input variables, including temperature, actual evapotranspiration, precipitation, and the discharge of the nearby river. Cross validation and a separated test set are used to ensure the robustness of the model. Furthermore, the impact of future meteorological data on forecast quality is analyzed to assess its contribution to improving model performance. The results show a variation in forecasting accuracy depending on the lead time, with a general tendency toward reduced accuracy over longer periods. However, the integration of available future meteorological data significantly enhances model performance, highlighting its importance in hydro geo logical forecasting and resource planning, opening new paths of inquiry for future developments. Estimating future trends in meteorological variables, even if uncertain, therefore appears to be a significant way of improving seasonal water level forecasts.