Welcome to NOAH
Water Quality Mixing and Water Level Prediction
Tucson, Arizona

Problem Definition

Tucson, Arizona, located within a mountain valley in the scenic Sonoran Desert, has over the last century severely overexploited its limited groundwater resources, causing water levels in the aquifer to drop several hundred feet. In order to reduce its over-dependence on groundwater, the City now utilizes Colorado River water, which is conveyed to and discharged into huge recharge basins located in near proximity to a large public supply wellfield, as shown in Figure 1. The recharged river water blends with the groundwater, which is extracted by large public supply wells for distribution to the community. Because the Colorado River water has a higher dissolved ion content than groundwater, recharging the groundwater system increases the dissolved ion content of the aquifer. Cognizant of public sentiment, Tucson Water is interested in reducing ion concentrations of the extracted blended aquifer water to the extent possible. In this project, artificial neural network (ANN) models were developed to accurately predict dissolved ion concentrations (i.e. specific conductivity) at the point of entry for the blended water 24-hours into the future as a function of pumping rates, basin inflows, and initial production well and basin water levels. In addition, ANN models were also developed to predict water levels in the aquifer one-week ahead, as measured by monitoring well WR-412a. Monitoring well water level measurements were usually made on a weekly basis, but there was a relatively short time period where daily measurements were collected in WR-412a. The data were used to generate weekly data by selecting water level measurements separated from 6 to 8 days. The ANN input variables included initial water levels in wells and basins, average basin inflows and production well pumping rates over the prediction week, and average basin flows for both one week prior and two weeks prior to the prediction event.



Table 1 presents the statistical performance of the ANN model develop for predicting conductance 24-hours ahead.

As shown, the ANN model achieved superior prediction performance. Figure 2 is the time-series depicting measured versus ANN predicted conductance values, respectively, for the 24-hour ahead prediction period. The ANN model accurately reproduced the higher and lower conductance values.

The statistical performance of the ANN model developed for predicting water levels in the observation well is summarized below in Table 2.

The ANN model achieved surprisingly high predictive accuracy, given the relatively few water level measurements available for development. Figure 3 below compares the ANN predicted water levels against the measured values for the entire data set.

As shown, the ANN model accurately reproduced the highly dynamic water level changes, closely tracking the higher and lower water level periods.


The ANN models can be used to optimize water quality conditions at the point of entry, as well as manage groundwater levels in the aquifer to minimize energy extraction costs. Because this dual objective management problem will be conflicting, in that supply wells located closest to the recharge basins will have the highest dissolved ion concentrations but the lowest lift requirements, this problem can be solved using multiobjective optimization, where the optimal balance between water quality and energy costs can be identified, as was done for the Toms River wellfield contamination issue.

A more detailed description of this study can be obtained by requesting the Preliminary Feasibility Study Report.