Diseño de una red neuronal artificial para la predicción de la dosis optima de policloruro de aluminio en el tratamiento de agua potable de la EPMAPA-SD

Abstract

Technological development has been taking place in basic water treatment operations, for this reason a predictive model was developed to know the optimal amount of coagulant in the clarification analysis of EPMAPA-SD. At the beginning, a database was generated with the information provided by the quality control area of EPMAPA-SD for the purification of water from 2015 to 2020, 121 data on the amount of coagulant used in jug tests. For the design of the network, the input variables were classified as turbidity, color and Ph; as an output variable the optimal dose of coagulant; to avoid redundancy in the data they were normalized. The neural network has three neurons in the input layer, 350 in the hidden layer and one neuron in the output layer, the selection process of this architecture was carried out through prediction tests with algorithms such as Levenberg Marquad, Bayesian Regularization and Scaled Conjugate Gradient, among these the Bayesian algorithm presented a mean square error of 1.94e-03 with a correlation in the training and test of the network of 0.947 and 0.923 respectively. The validation of the data predicted by the neural network was carried out statistically by means of a paired test that contrasts the variance between the data, with a p value greater than 0.05, the null hypothesis is accepted, stating with 95% confidence that it does not there is a statistically significant difference between the mean of the actual and predicted data. The prediction model calculates the optimal coagulant dose according to the quality requirements for drinking water established in the INEN 1108 standard. It is recommended to incorporate the prediction model in an automatic control system that allows the coagulant to be dosed in real time.