Aplicación de técnicas de aprendizaje automático para el desarrollo de soft-sensors en el ámbito del tratamiento de aguas residuales

Abstract

[EN] In this work, data from the monitoring of a wastewater treatment plant (WWTP) are exploited through machine learning techniques to design a data-based sensor. Data-based sensors or soft-sensors make use of measures available online for the estimation of other difficult-to-measure parameters, either because they entail high cost, high time or can only be obtained sporadically. In this case, the objective of the sensor to be designed is to obtain the nitrogen as nitrate concentration in the anoxic reactor of a biological process for carbon and nitrogen removal in urban wastewater. Since many WWTPs do not have a large amount of online instrumentation, one of the objectives of this work is to compare the loss of effectiveness of the model when the number of variables is reduced and especially selecting those that are easy to measure in terms of cost of investment and maintenance. Data from the characterization of the influent as well as from the processes that take place in the WWTP have been used to evaluate the convenience of using a linear or non-linear model. Subsequently, we have studied the variability of the model error based on the partition of the data set in training and test fractions, to establish an appropriate validation method. Finally, once the convenience of using a non-linear model was observed, a regression model based on Boosted Trees, that is, sets or ensembles of trees constructed using the boosting technique, has been adjusted.