Determination of optimal coagulant dosing rate using an artificial neural network

Determining coagulant dosing rates currently depends on Jar-test results and the experience of the operators in many cases. The nature of these practices makes it difficult to cope quickly with the rapid fluctuation of raw water quality, mainly because it takes a relatively long time to obtain Jar-test results. For promptly predicting required coagulant doses in response to water quality changes, a number of researchers have attempted to use the multi-variable regression (MVR) approach. However, the prediction capability of the MVR approach has not been satisfactory. An artificial neural network (ANN) is an excellent estimator of the nonlinear relationship between the accumulated input and output numerical data. Using this characteristic of the ANN, this study has attempted to predict the optimal coagulant dosing rate accurately and quickly. To train the ANN and deduce the MVR equation, a set of 142 units of data chosen from the 2-year operation of a water treatment plant was used. Another set of 72 units of data, not used in training, was also used to check the prediction capability of the trained ANN and MVR equation. Root-mean-square normalized error (RMSE) was used as a quantitative indicator of prediction capability. For the training data set and the raw data set, the RMSEs of the MVR equation were, respectively, 0.0143 and 0.0193 while those of the ANN were 0.0058 and 0.0092, respectively. These results indicate that the ANN reduced the prediction error for the training data by about 59%, and for the raw data by about 52%. Thus, our study demonstrates that the prediction capability of the ANN for raw data is enhanced by twice that of MVR. As the advancement of on-line monitoring techniques enables the ANN to update the weights periodically, its prediction capability can be also continuously enhanced.