Welcome to NOAH
Groundwater Level Prediction
Montville, New Jersey

Problem Definition

Artificial neural network (ANN) models were developed to accurately predict 30-days into the future highly variable head (i.e. water levels) at two monitoring wells located in a semi-confined buried valley glacial Towaco aquifer system in New Jersey under variable pumping and weather conditions. The two monitoring wells are located in near proximity to three public supply wells that pump from the semi-confined aquifer system, as shown in Figure 1. The ANN models were also used to conduct a sensitivity analysis to identify important relationships between the input predictor variables (e.g. pumping rates and precipitation) and the corresponding head values.



The ANN models achieved excellent prediction performance during validation, always correctly forecasting relative increases and decreases in head. The absolute mean error was significantly below the mean absolute change; for example, for the Indian Lane well, the mean absolute change from the initial to final measured head was 1.61 m, while the mean absolute error was just 0.27 m. Figures 2 and 3 below depict the validation results for the two monitoring wells, with the final measured versus the ANN predicted values shown in contrast to the initial measured water level for the respective prediction event. For both wells, the ANN models accurately reproduced higher and lower water table elevations, and in almost all cases, accurately predicted the relative movement of the potentiometric surface (i.e. water level) from the initial condition measured 30-days prior (i.e. rises or fall).

Sensitivity analyses conducted with the ANN also demonstrated that head in the semi-confined aquifer is least effected by precipitation over the 30-day prediction period, and more strongly correlated with temperature and pumping. The results conform to physical intuition, as the semi-confined nature of the system would dampen precipitation effects over short time periods. In contrast, temperature strongly correlates with pumping extractions, as well as areal recharge, and thus has a stronger correlation with water levels. Still, the weak correlation with precipitation does demonstrate hydraulic connection between the unconfined and semiconfined system, which can help justify stronger land management practices for protection of the drinking water supply.


The ANN could be used to help optimize pumping of the wellfield. Using the ANN derived state-transition equations as part of the constraint set, formal optimization methodology can be used to maximize total pumping without violating imposed water level constraints at the observation wells. The higher predictive accuracy achieved with the ANN that with the traditional physical-based model would result in identification of superior management solutions that can protect the scarce groundwater resources in the valley.

A more detailed overview of this work can be found in the following journal article:

Coppola, E., A. Rana, M. Poulton, F. Szidarovszky, and V. Uhl. (2005). A Neural Network Model for Predicting Water Table Elevations. Journal of Ground Water, 43, no 2: 231-24.