MACHINE LEARNING-BASED APPROACHES FOR MODELING VARIABILITY IN INFLUENT AND EFFLUENT QUALITY AT WASTEWATER TREATMENT PLANTS

Abstract

A rigorous analysis was performed using machine learning techniques such as regression trees, generalized linear models (GLM), and random forests to identify the relationships between climatic variables and quality parameters of two wastewater treatment plants located in very different geographical locations namely Agra, Uttar Pradesh, India and Colorado, US. Regression trees were used to classify the quality parameters based on climatic variables (temperature and precipitation) and time lagged values of the considered quality parameter. Temperature was identified as a major factor in classifying the influent values of quality parameters by showing a nearly inverse relationship, indicating that seasonal variation is seeming to take place. A correlation was developed between influent BOD and influent NH3 values using regression trees and time series analysis. The effect of extreme weather events was observed for the Agra plant (located in a tropical region), showing that extreme heatwaves events may lead to non-compliance of effluent BOD values. However, some other extreme events like large amounts of rainfall will favor in reducing the effluent BOD values. Similar effects were not observed for the Colorado plant (located in a temperate region). Rather it was observed that comparing the performance of machine learning models, random forests showed a high level of accuracy for predicting effluent BOD with a range of 80-90 percent. Random forests also helped in predicting effluent ammonia with an accuracy of nearly 70 percent, without taking influent ammonia into consideration. The GLM model was also useful in predicting effluent BOD values but was not as efficient as random forests