Welcome to NOAH
Water Treatment of a Surface Water Source for Potable Water
New Jersey

Problem Definition

A preliminary water treatment modeling assessment was conducted for a large water system, operated by New Jersey American Water, which extracts its supply from a surface water reservoir. For the water treatment modeling, the data consisted of 2.5 years of daily data. The 57 input variables used in the water treatment problem can be classified into four general groups: physical water data, chemical water data, water treatment data, and weather data. Physical water data included variables such as total daily flow and average water temperature. Chemical water data included average daily measured turbidity, pH, and average daily chlorine levels. Water treatment data were quantity or doses of chemicals added for treatment, such as sodium hydroxide and hydryfluosilicic acid. The output variables included average daily turbidity and average daily number of readings above 0.1 NTU. Because distinct raw water quality conditions are expected for different seasons of the year, in addition to developing and testing annual ANN models (i.e. trained with data for all seasons), four seasonal ANN models were developed and tested for each season.



Table 1 compares the seasonal ANN models against the best ANN models developed for the entire data set (annual) for the average turbidity and daily reading prediction variables. For both cases, the seasonal models always outperformed the annual models in terms of mean absolute error (MAE). It is important to note that the two ANN summer models significantly outperformed their corresponding annual models. This suggests that the summer season has some unique conditions and/or behavior that are distinctly different from the other times of the year. It appears that when this season is individually modeled, the ANN models are more able to capture this distinct behavior. In contrast, autumn and winter do not appear to have as distinct behavior, particularly with regarding to daily readings.

Figures 1 and 2 depict time series for the spring and summer seasons, respectively, comparing measured predictions made by the seasonal ANN models against the corresponding annual model for average daily turbidity. In general, the annual model has a tendency to smooth out the predictions, whereas the seasonal models do a more effective job, in general, of capturing higher and lower values.



A perturbation analysis was conducted to better understand input-output relationships, where the input values for select variables were incrementally changed, and the corresponding average turbidity levels computed. Higher water deliveries and higher settled turbidity readings both resulted in higher average turbidity levels for finished water. Both of these tendencies make sense, as both variables are measures of the quantity of mass in the raw water. If all other variables (e.g. chemical dosages) are held constant, turbidity removal is not increased, and thus average turbidity levels of the finished water will be elevated. The surprising inverse relationship between raw water temperature and average turbidity was identified with the above analysis was verified with time-series plotting. At higher temperatures, raw water has on average lower turbidity. By extension, treated finished water would be expected to have lower turbidity during times when the raw water has a higher temperature, as in summer months. Lastly, chlorine effluent appears to be weakly positively correlated with average turbidity. This may have some physical basis, as the presence of organic material, such as algae, will increase raw water turbidity, which will increase the quantity of chlorine added.

Despite relatively few years of data, the ANN models achieved relatively high predictive accuracy. The ANN technology can quantify important cause and effect relationships, identify seasonal variations, and optimize water treatment strategies, all of which can reduce treatment costs and improve finished water quality.