Water quality prediction of the Yamuna River in India using hybrid neuro-fuzzy models

• Main Authors: Kisi, Ozgur, Parmar, Kulwinder Singh, Amin Mahdavi-Meymand, Amin Mahdavi-Meymand, Muhammad Adnan, Rana, Shahid, Shamsuddin, Mohammad Zounemat-Kermani, Mohammad Zounemat-Kermani
• Format: Article
• Language: English
• Published: MDPI 2023
• Subjects: TA Engineering (General). Civil engineering (General)
• Online Access: http://eprints.utm.my/107539/1/ShamsuddinShahid2023_WaterQualityPredictionoftheYamunaRiver.pdf; http://eprints.utm.my/107539/; http://dx.doi.org/10.3390/w15061095

The potential of four different neuro-fuzzy embedded meta-heuristic algorithms, particle swarm optimization, genetic algorithm, harmony search, and teaching–learning-based optimization algorithm, was investigated in this study in estimating the water quality of the Yamuna River in Delhi, India. A cross-validation approach was employed by splitting data into three equal parts, where the models were evaluated using each part. The main aim of this study was to find an accurate prediction model for estimating the water quality of the Yamuna River. It is worth noting that the hybrid neuro-fuzzy and LSSVM methods have not been previously compared for this issue. Monthly water quality parameters, total kjeldahl nitrogen, free ammonia, total coliform, water temperature, potential of hydrogen, and fecal coliform were considered as inputs to model chemical oxygen demand (COD). The performance of hybrid neuro-fuzzy models in predicting COD was compared with classical neuro-fuzzy and least square support vector machine (LSSVM) methods. The results showed higher accuracy in COD prediction when free ammonia, total kjeldahl nitrogen, and water temperature were used as inputs. Hybrid neuro-fuzzy models improved the root mean square error of the classical neuro-fuzzy model and LSSVM by 12% and 4%, respectively. The neuro-fuzzy models optimized with harmony search provided the best accuracy with the lowest root mean square error (13.659) and mean absolute error (11.272), while the particle swarm optimization and teaching–learning-based optimization showed the highest computational speed (21 and 24 min) compared to the other models.