Author Identifier (ORCID)
Hoda Khoshvaght: https://orcid.org/0000-0001-8766-419X
Amir Razmjou: https://orcid.org/0000-0002-3554-5129
Mehdi Khiadani: https://orcid.org/0000-0003-1703-9342
Abstract
Unlike previous studies that rely on high-frequency (15-min or hourly) datasets, this study is among the first to use low-frequency (weekly) data to evaluate the performance of linear and nonlinear machine learning (ML) algorithms for predicting biochemical oxygen demand (BOD) and ammonium nitrogen (NH4+-N) in the primary and secondary treatment effluents from the Subiaco Water Resource Recovery Facility (WRRF) in Western Australia. Various feature selection methods, including filters, wrappers, and embedded methods, were employed to identify the most effective approach that achieves the highest model performance while enhancing computational efficiency. The results demonstrate that a reduced set of key features can achieve comparable predictive accuracy with lower computational complexity. For BOD prediction in primary effluent, a multilayer perceptron (MLP) achieved a root mean square error (RMSE) of 23.50 mg per liter (mg/L) using features selected based on mutual information. In the secondary effluent, SVR (rbf) and random forest feature selection yielded the best predictions, achieving an RMSE of 3.26 mg/L. Similarly, for NH4+-N, multiple linear regression with backward elimination achieved an RMSE of 2.71 mg/L in the primary effluent. In comparison, a random forest with five key predictors achieved an RMSE of 1.51 mg/L in the secondary effluent, indicating high accuracy in NH4+-N prediction. These findings demonstrate that data-driven models can predict BOD and NH4+-N using low-frequency monitoring data, supporting supervisory-level operational decision-making in wastewater treatment plants and near-real-time wastewater quality assessment. Furthermore, generalization analysis indicates that linear models perform more consistently across multiple targets and evaluation metrics.
Keywords
Effluent management, machine learning prediction models, primary treatment effluent, supervised learning algorithms, wastewater treatment
Document Type
Journal Article
Date of Publication
5-15-2026
Volume
172
Publication Title
Engineering Applications of Artificial Intelligence
Publisher
Elsevier
School
Mineral Recovery Research Centre / School of Engineering
Funders
Australian Government Research Training Program Scholarship / Water Corporation of Western Australia
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Comments
Khoshvaght, H., Permala, R. R., Razmjou, A., & Khiadani, M. (2026). Machine learning approaches for predicting biochemical oxygen demand and ammonium nitrogen: A decade-long weekly field study at a full-scale water resource recovery facility. Engineering Applications of Artificial Intelligence, 172, 114349. https://doi.org/10.1016/j.engappai.2026.114349