LSTM-based autoencoder models for real-time quality control of wastewater treatment sensor data

Abstract

Abstract The operation of smart wastewater treatment plants (WWTPs) is increasingly paramount in improving effluent quality, facilitating resource recovery and reducing carbon emissions. To achieve these objectives, sensors, monitoring systems, and artificial intelligence (AI)-based models are increasingly being developed and utilised for decision support and advanced control. Key to the adoption of advanced data-driven control of WWTPs is real-time data validation and reconciliation (DVR), especially for sensor data. This research demonstrates and evaluates real-time AI-based data quality control methods, i.e. long short-term memory (LSTM) autoencoder (AE) models, to reconcile faulty sensor signals in WWTPs as compared to autoregressive integrated moving average (ARIMA) models. The DVR procedure is aimed at anomalies resulting from data acquisition issues and sensor faults. Anomaly detection precedes the reconciliation procedure using models that capture short-time dynamics (SD) and (relatively) long-time dynamics (LD). Real data from an operational WWTP are used to test the DVR procedure. To address the reconciliation of prolonged anomalies, the SD is aggregated with an LD model by exponential weighting. For reconciling single-point anomalies, both ARIMA and LSTM AEs showed high accuracy, while the accuracy of reconciliation regresses quickly with increasing forecasting horizons for prolonged anomalous events.