A novel long short-term memory artificial neural network (LSTM)-based soft-sensor to monitor and forecast wastewater treatment performance

Abstract

Commercial instrumentation for measurement of various wastewater treatment processes parameters is costly and time-consuming in wastewater treatment plants (WWTPs). Long short-term memory neural network (LSTM) based soft-sensors to monitor and forecast crucial performance parameters including chemical oxygen demand (COD), NH4+ and total nitrogen (TN), which was based on lower-cost sensors dataset (e.g., dissolved oxygen, oxidation reduction potential (ORP) and suspended solids), were developed for a two-staged anoxic-oxic (A/O) process for wastewater treatment. Pearson correlation analysis was conducted to identify the essential model input features before LSTM development. With optimization of look-back periods, the proposed LSTM-based soft-sensors outperformed multiple linear regression model (MLR)-based soft-sensors for prediction of influent COD, influent NH4+, effluent COD and effluent TN. It was supported by the lower mean absolute percentage error, lower root mean squared error and higher Pearson correlation for LSTM-based soft-sensors compared to those of MLR-based soft-sensors. The overall results indicated that LSTM-based soft-sensors can achieve automated high-resolution measurement and effectively forecast the crucial performance of biological wastewater treatment, potentially lowering the capital cost for sensor installation in WWTPs.