Water quality prediction based on Naïve Bayes algorithm

• Published in: Water science and technology Vol. 85; no. 4; pp. 1027 - 1039
• Main Authors: Ilić, M., Srdjević, Z., Srdjević, B.
• Format: Journal Article
• Language: English
• Published: England IWA Publishing 01.02.2022
• Subjects: Accuracy; Algorithms; Ammonium; Ammonium compounds; Artificial intelligence; Bayes Theorem; Biochemical oxygen demand; Canals; Electrical conductivity; Electrical resistivity; Environmental Monitoring - methods; Machine Learning; Missing data; naïve bayes; netica; Nitrogen oxides; Orthophosphates; Oxygen; Oxygen content; Oxygen demand; Oxygen requirement; Photochemicals; Prediction models; prediction of water quality classes; Saturation; serbian water quality index (swqi); Solid suspensions; Suspended particulate matter; Suspended solids; Total oxygen demand; Usability; Water analysis; Water demand; Water management; Water Pollution; Water Quality; water quality parameters; Water sampling
• ISSN: 0273-1223; 1996-9732; 1996-9732
• DOI: 10.2166/wst.2022.006

In the fast-changing world with increased water demand, water pollution, environmental problems, and related data, information on water quality and suitability for any purpose should be prompt and reliable. Traditional approaches often fail in the attempt to predict water quality classes and new ones are needed to handle a large amount or missing data to predict water quality in real time. One of such approaches is machine-learning (ML) based prediction. This paper presents the results of the application of the Naïve Bayes, a widely used ML method, in creating the prediction model. The proposed model is based on nine water quality parameters: temperature, pH value, electrical conductivity, oxygen saturation, biological oxygen demand, suspended solids, nitrogen oxides, orthophosphates, and ammonium. It is created in Netica software and tested and verified using data covering the period 2013–2019 from five locations in Vojvodina Province, Serbia. Forty-eight samples were used to train the model. Once trained, the Naïve Bayes model correctly predicted the class of water sample in 64 out of 68 cases, including cases with missing data. This recommends it as a trustful tool in the transition from traditional to digital water management.