A new approach to estimating oxygen off-gas fraction and dynamic alpha factor in aeration systems using hybrid machine learning and mechanistic models

Abstract

The aeration process is the largest energy consumer in wastewater treatment plants (WWTP), and the optimization of the process based on computational models can offer significant savings for the plant. Recent theoretical developments have revealed that many of the parameters commonly assumed as constants in aeration modelling have a dynamic nature; however, there still lacks a universal way to model these factors accurately. This work proposed a new framework and optimal time scale for building hybrid machine learning-mechanistic oxygen transfer rate models. The data used for the modelling were from existing sensors in a full-scale WWTP supplemented by a two-month field sampling campaign for collecting 15 min of dynamic data, including off-gas air fraction. Two machine learning (ML) models were built: the first set was a novel hybrid ML-mechanistic model developed to estimate the off-gas aeration fraction from wastewater parameters and calculate dynamic alpha per the off-gas method. The second ML model directly estimated dynamic alpha from wastewater process and operation parameters. The performance of the ML models was compared with recalibrated published regression-based models. Results showed that the hybrid ML-mechanistic alpha model (NSE = 0.96 and RMSE = 0.03) were more accurate than the direct ML alpha (NSE = 0.67 and RMSE = 0.07) and recalibrated regression (NSE < 0 and RMSE = 0.18 to 0.28) models, and therefore could be efficiently applied to predict dynamic alpha factor of an activated sludge plant. In addition to typical surfactant indicators such as MLSS and COD, other parameters like temperature and aeration ammonia concentration were substantial in dynamic off-gas and alpha estimation.