Prediction of biochemical oxygen demand (BOD5) in wastewater treatment plant with rough set and machine learning hybrid approach

During the operation of wastewater treatment in plants, it is important to quickly monitor and intervene the quality of water. Some bacterial-based biological methods are used for the wastewater which is coming to wastewater treatment plants. The pollution and oxygen balance of the incoming water affects the life cycle within the facility. In sudden rises in pollution parameters values at the entrance of the facility, bacteria that provide biodegradation begin to die in the cases where the operator does not intervene or improperly intervenes. Therefore, treatment cannot be carried out as the biological balance will be disrupted. When compared to other parameters, the measurement of Biological Oxygen Demand (BOD), which is an important variable in water quality management and planning, in the laboratory environment takes longer times (5 days). In this study, which was carried out in a wastewater treatment plant in Sakarya, it was aimed to transfer the data obtained from the parameters that can be measured with online measurement devices to the automation and control systems and to estimate the Biological Oxygen Demand (BOD5) by applying artificial intelligence methods to these data. In order to predict BOD5 a hybird model which includes Fuzzy Rough Set, SmoteR and machine learning Regression algorithms (Boosted Decision Tree Regression, Bayesian Linear Regression, Decision Forest Regression, Neural Networks, Linear Regression) has been developed. While applying Fuzzy Rough Set based feature selection and SmoteR methods in order to increase the performance of machine learning algorithms used for prediction, Fuzzy Rough Set instance selection algorithm has been applied to eliminate unnecesary instances. Machine learning regression algorithms were trained on the obtained data sets and their performances were compared.The comparative results show that Fuzzy Rough Set based feature selection, instance selection and especially SmoteR methods clearly increase the performance of machine learning algortihms. In the first stage of the study (without applying Fuzzy Rough Set based feature selection, Fuzzy Rough Set instance selection, SmoteR methods), Bayesian Linear Regression algorithm with a R2 value of 91.56% and Linear Regression algorithm with a MAPE value of 10.35% show the best performances. As a result of the study by applying Fuzzy Rough Set based feature selection, Fuzzy Rough Set instance selection, SmoteR methods, the Boosted Decision Tree Regression Algorithm showed the best performance among the machine learning algorithms, with a R2 value of 97.18% and a MAPE value of 6.07%.