Artificial neural networks vs. Gene Expression Programming for estimating outlet dissolved oxygen in micro-irrigation sand filters fed with effluents

Abstract

Sand media filters are specially used to avoid emitter clogging when water with large amount of organic pollutants like effluents are used in micro-irrigation systems. Estimation of water quality parameters such as dissolved oxygen at sand filter outlet (DOo) is of great interest for irrigation engineers. Artificial neural networks (ANN), Gene Expression Programming (GEP) and Multi Linear Regression (MLR) based models were trained for estimating DOo using data from 769 experimental filtration cycles. Instead of considering a single configuration of the training and test data sets, which is the usual procedure for those applications in agricultural studies, the performance of those models was assessed through k-fold testing, ensuring a complete performance evaluation. In general, the GEP model tended to provide the most accurate estimations, followed by ANN and, lastly, by MLR models. After the evaluation of the models, the GEP approach was used to provide a new equation to estimate DOo based on the complete data set. This procedure revealed that only inlet DO, pH, electrical conductivity and filter head loss were necessary to feed the models. Furthermore, the consideration of leave one out or, at least, k-fold assessment should be advisable to perform a suitable evaluation of the model performance. Otherwise, conclusions drawn might be only partially valid.