Prediction of parameters characterizing the state of a pollution removal biologic process

This work is devoted to the prediction, based on neural networks, of physicochemical parameters impossible to measure on-line. These parameters-the chemical oxygen demand (COD) and the ammonia NH"4-characterize the organic matter and nitrogen removal biological process carried out at the Saint Cyprien WWTP (France). Their knowledge make it possible to estimate the process quality and efficiency. First, the data are treated by K-Means clustering then by principal components analysis (PCA) in order to optimize the multi-level perceptron (MLP) learning phase. K-Means clustering makes it possible to highlight different operations within the Saint Cyprien treatment plant. The PCA is used to eliminate redundancies and synthesizes the information expressed by a data set. With respect to the neural network used, these techniques facilitate the pollution removal process understanding and the identification of existing relations between the predictive variables and the variables to be predicted.