Forecasting effluent and performance of wastewater treatment plant using different machine learning techniques

Abstract

Expectation of wastewater quality in wastewater treatment plants (WWTPs) is significant and can decrease the sampling number, cost, decision time, and energy. This paper presents two methods for predicting and forecasting the removal efficiency of total suspended solids (TSS), chemical oxygen demand (COD), biological oxygen demand (BOD5), ammonia, and sulphide at El-Berka wastewater treatment plant, Egypt. The first method uses different prediction models, includes the use of traditional feed-forward (TF), deep feed-forward backpropagation (DFB), and deep cascade-forward backpropagation (DCB) networks. The TF was generated in three layers: input, hidden, and an output layer. The DFB network comprised of six layers: an input, four hidden, and an output layer. The DCB network was created with six layers (with skip connections): an input, four hidden, and an output layer. The other method is a forecasting model by using deep learning time series forecasting (DLTSF) with a long short-term memory (LSTM) network. The developed models were trained, validated, and tested on a real-life dataset over eight years. The results indicated that the presented models could effectively predict and forecast the future series values of the removal efficiency of the El-Berka WWTP. The DCB network achieved the highest accuracy as compared to those exhibited by the TF and DFB networks. The RMSE and R-squared for training with the DCB model are 1.95 and 0.90, respectively. The RMSE of DLTSF was 0.85 for forecasting of BOD5. Thus, the DCB and DLTSF models are recommended for evaluating and predicting the performance of WWTP.