Prediction of wastewater quality parameters using adaptive and machine learning models: A South African case study

The wastewater treatment process often faces challenges in monitoring water quality parameters (WQ), to overcome this there is a need for developing innovative modeling approaches. Hence, the present study is motivated by the potential application of adaptive and machine learning (ML) models as soft sensors to predict the WQ in one of the largest Municipal Wastewater Treatment Plants (MWWTP) in KwaZulu-Natal, South Africa. Seven different adaptive and ML algorithms were examined and compared, varying from adaptive strategies to ML architectures such as Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Time Difference (TD), Just in Time Learning (JIT), Moving Window (MW), and fusion of adaptive strategies (JITTD, and JITTDMW), Support Vector Regression (SVR), and Artificial Neural Network (ANN). Based on the results, BiLSTM consistently provided the most accurate estimation of effluent parameters, with an error rate ranging from 3.12 to 9.75 % for all variables. For Chemical Oxygen Demand (COD), ammonia, pH, and Total Suspended Solids (TSS), BiLSTM model yielded low errors (Mean Absolute Error (MAE) values of 1.54, 0.1, 0.22, and 1.14) with lower correlation coefficient values (<0.7) compared to the six other models proposed. As for conductivity, COD, TSS, pH, ammonia, LSTM, and JITTDMW, JITTD performed well with MAE values between 1 and 8 but had difficulty estimating soluble reactive phosphate (SRP). From a future perspective, these models could be applied to other MWWTPs facing similar challenges, potentially helping to improve their performance and effectiveness. Overall, this study identifies promising ways to optimize MWWTPs using ML-based predictive models. • ANN, JITTD, JIT, JITDMW, LSTM, BILSTM, and SVR models were employed to predict water quality parameters. • South African municipal WWTP data is used as a case study to verify the effectiveness of the models. • BiLSTM, produced sublime results with all parameters in the range of 3.12–9.75 % error. • Sublime results are obtained with BiLSTM with a strong correlation <0.90. • BiLSTM model predictions were very close to actual values of nutrient removals.