Development of artificial neural networks for the optimization of drinking water treatment

Effective filtration is essential to maintaining surface water treatment objectives. To maximize water quality in terms of effluent turbidity and filter run time, the Elgin Area WTP has adopted artificial neural networks (ANN) to predict settled water and filter effluent turbidity as well as optimum alum and carbon dioxide dosages. Settled water turbidity dosage-response surfaces were produced by applying a matrix of assumed alum and carbon dioxide dosages and comparing the results to historical averages. Trained ANNs were used to predict filter effluent turbidity for a variety of flow rates and compared to removal efficiency equations. ANN responses based on assumed values proved to be successful diagnostic tests for the ANN and ease the choice of operating conditions for optimal performance. Point-wise confidence intervals based on error and squared error values were produced using ANN. These ANN based confidence intervals performed well compared to statistic-based confidence intervals.