Welcome to NOAH
Groundwater Level Prediction
Eden Prairie, Minnesota

Problem Definition

Eden Prairie, Minnesota operates a wellfield that induces high drawdowns, which results in high energy lift requirements. Artificial neural network (ANN) models were developed to accurately predict highly dynamic head values (i.e. water levels) in 11 high capacity production wells located in a complex multi-layered groundwater under variable pumping and weather conditions. The 11 wells are screened in a semi-confined system, consisting primarily of dolomite, and the Cambrian Jordan Sandstone, as shown in Figure 1.

Because each public supply well can pump more than 1,000 gallons per minute, and is subject to interference effects, drawdowns are extreme and can change suddenly. The prediction periods selected were 24-hour periods consisting of hourly time steps or stress periods. The ANN model was 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 trained ANN models accurately predicted head changes in all 11 public supply wells, with absolute mean prediction errors of approximately 1.1 meters. In addition, the ANN’s accurately predict head changes over the 24 consecutive hourly stress periods without using measured water levels for ANN re-initialization. That is, measured head values were used only for initialization of the ANN at the start of the 24-hour validation period. For the remaining 23-stress period predictions, the head values predicted by the ANN for the previous stress period were used to re-initialize the network. The figures compare measured head data (Measured) for two representative wells against the ANN predictions with both real-world re-initialization (ANN-Rei) and “continuous” or self-initializing (ANN-Con). As shown, there is minimal degradation with continuous ANN processing; the mean absolute error increased for all wells from 1.06 to 1.2 meters.


A sensitivity analysis conducted with the ANN also demonstrated that weather variables had minimal influence on the system, with hourly head changes almost entirely the result of pumping conditions.

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 minimize total costs while satisfying required demand constraints for the system.

A more detailed overview of this work can be found in the following journal article:
Coppola, E., M. Poulton, E. Charles, J. Dustman, and F. Szidarovszky. (2003). Application of Artificial Neural Networks to Complex Groundwater Management Problems. Journal of Natural Resources Research, 12, no. 4: 303-320.