Loading...
Prediction of biological wastewater treatment performance using artificial neural networks
Author(s)
Sindane, Winile
Date Issued
2023
Type
Thesis
Publisher
Cape Peninsula University of Technology
Abstract
Process control and monitoring of wastewater treatment plants are essential, but have proven to be slow, expensive and dearth. It is currently achieved by examining effluent wastewater quality and adjusting the treatment process. This justifies the need to develop robust mathematical modelling tools known for high accuracy to predict the performance of wastewater treatment plants for future purposes.
A comparative study on the prediction of biological wastewater treatment performance of industrial wastewater, biodiesel- (BDWW), textile- (TTWW), polymer- (PWW), as well as pulp and paper wastewater (PPWW) using artificial neural networks (ANN), was carried out based on historical data from previous studies. Industrial wastewater was characterised by high levels of pollutants represented by the chemical oxygen demand (COD) since it is one of the important parameters used to evaluate the performance of wastewater treatment systems. Three ANN-based models, namely, nonlinear autoregressive neural network model with exogenous inputs (NARX), feedforward backpropagation (FFB) and cascade feedforward backpropagation (CFBP), were developed to predict the COD of the effluent using the Levenberg-Marquardt (LM) backpropagation algorithm. The ANN models were developed using a three-layered ANN architecture, including the input, hidden and output layer. The best ANN architecture from the three models was chosen after several steps of training, testing and validation using a trial-and-error method altering the number of neurons ranging from 2 to 11 in the hidden layer.
Based on all three model performances and prediction capabilities, the most appropriate ANN model was found to be the NARX for all four industrial wastewater and treatment methods with a mean square error (MSE) of 0.0239, 0.303, 0.0719, 0.343 and an overall model correlation coefficient (𝑅) for training, validation, and testing of 0.988, 0.838, 0.964, 0.809 for the BDWW, TTWW, PWW and PPWW, respectively. According to the MSE and 𝑅 values obtained, it was concluded that the NARX performed better and could accurately predict COD effluent concentration, which proves that ANN-NARX can be employed successfully to estimate COD effluent concentration from biological wastewater treatment systems. The CFBP model also showed better prediction results compared to the FFB model with overall 𝑅 values of 0.947 for BDWW, 0.736 for TTWW, 0.837 for PWW and 0.739 for PPWW. However, the model showed poor performance with an MSE values of 0.1024, 0.444, 0.297 and 0.457, respectively, which could result in poor generalisation when presented with new data sets.
Based on the results obtained from all the ANN methods, it can be concluded that ANNs are reliable modelling tools for successfully predicting biological wastewater treatment systems performance focused on the effluent COD. Proper selection of ANN input parameters resulted in good prediction and network performance of the ANNs. The quality and quantity of the historical data had a significant influence on the network performance, poor quality and fewer data resulted in poor prediction and ANN performance.
A comparative study on the prediction of biological wastewater treatment performance of industrial wastewater, biodiesel- (BDWW), textile- (TTWW), polymer- (PWW), as well as pulp and paper wastewater (PPWW) using artificial neural networks (ANN), was carried out based on historical data from previous studies. Industrial wastewater was characterised by high levels of pollutants represented by the chemical oxygen demand (COD) since it is one of the important parameters used to evaluate the performance of wastewater treatment systems. Three ANN-based models, namely, nonlinear autoregressive neural network model with exogenous inputs (NARX), feedforward backpropagation (FFB) and cascade feedforward backpropagation (CFBP), were developed to predict the COD of the effluent using the Levenberg-Marquardt (LM) backpropagation algorithm. The ANN models were developed using a three-layered ANN architecture, including the input, hidden and output layer. The best ANN architecture from the three models was chosen after several steps of training, testing and validation using a trial-and-error method altering the number of neurons ranging from 2 to 11 in the hidden layer.
Based on all three model performances and prediction capabilities, the most appropriate ANN model was found to be the NARX for all four industrial wastewater and treatment methods with a mean square error (MSE) of 0.0239, 0.303, 0.0719, 0.343 and an overall model correlation coefficient (𝑅) for training, validation, and testing of 0.988, 0.838, 0.964, 0.809 for the BDWW, TTWW, PWW and PPWW, respectively. According to the MSE and 𝑅 values obtained, it was concluded that the NARX performed better and could accurately predict COD effluent concentration, which proves that ANN-NARX can be employed successfully to estimate COD effluent concentration from biological wastewater treatment systems. The CFBP model also showed better prediction results compared to the FFB model with overall 𝑅 values of 0.947 for BDWW, 0.736 for TTWW, 0.837 for PWW and 0.739 for PPWW. However, the model showed poor performance with an MSE values of 0.1024, 0.444, 0.297 and 0.457, respectively, which could result in poor generalisation when presented with new data sets.
Based on the results obtained from all the ANN methods, it can be concluded that ANNs are reliable modelling tools for successfully predicting biological wastewater treatment systems performance focused on the effluent COD. Proper selection of ANN input parameters resulted in good prediction and network performance of the ANNs. The quality and quantity of the historical data had a significant influence on the network performance, poor quality and fewer data resulted in poor prediction and ANN performance.
Additional information
Thesis (MEng (Chemical Engineering))--Cape Peninsula University of Technology, 2023
File(s)![Thumbnail Image]()
Loading...
Name
Sindane_Winile_214223043.pdf
Size
2.09 MB
Format
Adobe PDF
Checksum
(MD5):8178dc72ae3d377221480e33cb39c749