A Study on the Development of a DNN Model for Predicting the Inflow BOD in Sewage Treatment Plants

The aim of this study is to predict BOD (Biochemical Oxygen Demand) more accurately in sewage treatment plants, allowing for the suggestion and implementation of various control strategies. This would enable the proper regulation of organic load rates, ultimately enhancing the efficiency of the treatment process. For this purpose, the study was conducted based on 2,100 daily data measurements from three sewage treatment plants in City B collected between January 1, 2015, and September 30, 2020. To predict the complex behavior of the influent characteristics that determine the BOD of the influent, a DNN(Deep Neural Network) model composed of nonlinear functions was used. To validate these results, model verification was conducted using R2 and RMSE with MLR using stepwise selection and principal component analysis. Additionally, the water quality characteristics from the immediate past are closely correlated with the BOD of the influent on the current day because of the continuous inflow nature of the influent. Consequently, derived variables of the water quality characteristics from the previous day and two days prior were added and applied to the model. To verify the validity of these derived variables, the predictive power of the model with the derived variables was compared and evaluated against a model using a dataset without these derived variables. The results demonstrated that the DNN model outperformed the existing sewage treatment plant prediction methods and the MLR model. The application of a nonlinear model to sewage treatment plants showed the potential to improve the process efficiency. Additionally, it was verified that the performance of both the DNN and MLR models improved when the derived variables were added, compared to when they were not, confirming the validity of the derived variables.