Evaluation of LSTM‐Based Short‐Term Prediction of MBR Operational Parameters: Applicability and Limitations for Aeration Control Applications

ABSTRACT

This study evaluated the predictive capability of long short‐term memory (LSTM) neural network models for forecasting key operational variables in a membrane bioreactor (MBR) using a comparatively modest training dataset, a limited number of sensors, and a basic model architecture, with practical implementation in mind. Time‐series data for dissolved oxygen (DO), pH, transmembrane pressure (TMP), mixed liquor suspended solids (MLSS), and airflow rate were collected from a laboratory‐scale MBR over several months of continuous operation. Separate training and testing periods were used to develop and evaluate the models. The LSTM models achieved good predictive accuracy for DO, pH, and TMP over a 6‐h prediction horizon and retained useful predictive capability at 12 h. However, the 12‐h predictions showed a systematic delay in responding to process changes, indicating reduced reliability at longer horizons. Correlation analysis between airflow rate and the other operational variables further showed that the LSTM models did not fully capture the dynamic responses of the MBR to changes in aeration. This limitation likely reflects the complex and nonlinear behavior of activated sludge processes. Overall, the results indicate that a basic LSTM architecture trained on a comparatively modest dataset can support operational forecasting over horizons of 6 and 12 h. However, improved representation of aeration‐related dynamics is required before the model can be used reliably for advanced operational decision support.