Water Quality Modelling and Parameter Assessment Using Machine Learning Algorithms: A Case Study of Ganga and Yamuna Rivers in Prayagraj, Uttar Pradesh, India

Abstract

Due to the rapid growth in population, industrialization and agricultural outputs, the stress on groundwater and surface water has increased exponentially. The water quality depends on the interaction of both the groundwater and surface water; therefore, management and monitoring of the surface water are the need of the hour. River water management is a major environmental challenge worldwide. Because of the nonlinear behaviour of various water quality parameters, estimating the water quality of a surface water at any point of its flow is a time-consuming task. River quality monitoring is a difficult, cumbersome, and costly process that can lead to many analytical errors. Therefore, the main objective of this work is to create a reliable model for assessing and forecasting changes in water quality in Prayagraj (earlier know as Allahabad), Uttar Pradesh, India at three separate places, including the Ganga River, Yamuna River, and confluence of both rivers (also known as Sangam) using artifical neural network (ANN) and genetic algorithm (GA) models. The developed model was used to statistically compare the results by analysing samples collected from the selected stations fortnightly. Based on the correlation matrix of the water quality for three stations, general prediction models for the selected parameters, namely DO, hardness, turbidity, and BOD, were developed. The prediction model was developed for DO, hardness, and turbidity for station 1 (Ganga River). The results showed that the correlation coefficient (R) for the ANN prediction model is 0.97, the average absolute relative error (AARE) is 0.002, and the model efficiency (ME) is 0.95 for the hardness prediction model. Similarly, the BOD-ANN prediction model performed well at station 2 (Yamuna River) and station 3 (Sangam), with R = 0.99, AARE = 0.006, root mean square error (RMSE) = 0.06, and ME = 0.99 at station 2. Overall, ANN outperforms all other modelling techniques for all four prediction models.