Enhancing wastewater treatment efficiency through machine learning-driven effluent quality prediction: A plant-level analysis

The main objective of this study was to develop, validate, and comprehend machine learning (ML) models capable of predicting chemical oxygen demand concentration in the effluent (CODout) of a given wastewater treatment plant (WWTP). The parameter CODout was chosen as a target due to its crucial role in controlling and optimizing water treatment plants, emphasizing its vital importance in maintaining the efficiency and quality of the water purification process. A calibrated WWTP model available on WEST software (DHI), developed using the Umbilo WWTP (South Africa) as a reference, served as a source for extracting influent and effluent data. The dataset was organized daily and hourly to train the ML predictive models. ML techniques used in this study include Support Vector Machines (SVM), Long Short-Term Memory (LSTM), Multilayer Perceptron (MLP), and Random Forest (RF). Exploratory data analysis revealed multicollinearity between some of the model's input variables, requiring a selection of these parameters. In the case of CODout predictions using the daily dataset, the MLP model proved more effective than other ML models. When the hourly dataset was applied, the LSTM models performed better by incorporating historical data into the model structure. When real effluent measurements were used to predict CODout, the SVM model had superior results, even outperforming traditional mechanistic models. Variable importance analysis showed the significant influence of influent TSS to predict CODout. In conclusion, the study demonstrates the successful applicability of ML models to predict CODout in WWTPs and provides valuable insights for optimizing wastewater treatment. The emphasis on model interpretability and validation with real data enriches the results' reliability. [Display omitted] •ML algorithms predict COD concentrations in a wastewater treatment plant.•Exploratory data analysis has streamlined the data set for efficient ML.•ML models outperformed traditional methods in predicting COD concentration.•LSTM model excelled in COD prediction, especially for hourly dataset (R2 = 0.99).•LSTM models demonstrated ability to capture complex temporal dynamics.