Prediction of MBR operating parameter using LSTM neural network

Abstract

This study investigated the forecasting ability of the long short-term memory neural network model (LSTM model), which is a type of recurrent neural network (RNN), for the dynamic character of membrane bioreactor (MBR). MBR is an advanced wastewater treatment system that combines activated sludge process with a membrane separation system. In this study, dissolved oxygen (DO), pH, trans membrane pressure (TMP), mixed liquor suspended solids (MLSS), and air flow rate of a bench-scale MBR were measured to obtain the time series data, and the time interval for each time series was unified to 1 hour. The training period of 640 hours was adopted for the LSTM model, and the remaining 160 hours were used as the testing period. The trained LSTM model predicted DO, pH, TMP, and MLSS one step ahead (one hour ahead), and multiple steps forecasts up to 6 hours ahead were also tested. The LSTM model succeeded in predicting MLSS one hour ahead with high accuracy. On the other hand, for DO and pH, the values predicted one hour ahead by the LSTM model reproduced their temporal fluctuation patterns to some extent. However, all of them tended to show predicted values that were lower than the actual values. The predicted values from the LSTM model did not reproduce the pattern of TMP changes well. In addition, the LSTM model was investigated the effect of forecasting horizons and look back period.