Hybrid model composed of machine learning and ASM3 predicts performance of industrial wastewater treatment

Conventional mechanistic models encounter difficulties in predicting treatment performance of industrial wastewater due to the substantial fluctuations in composition and biodegradability. This research constructed a hybrid model, integrating machine learning models with a mechanistic model Activated Sludge Model 3 (ASM3), to predict the treatment performance of two activated sludge reactors (ASRs) for the treatment of two different types of petrochemical wastewater. Prediction performance of ASM3 based on average wastewater biodegradability was found to be suboptimal, in terms of high Mean Absolute Percentage Error (MAPE) (31 %–47 %) and low correlation coefficient (0.47–0.50) for mixed liquor suspended solids (MLSS) and chemical oxygen demand (COD). To address this, multivariate linear regression (MLR) and decision forest (DF) models were employed to simulate the wastewater biodegradability, which was then integrated with ASM3 (referred to as hybrid models of MLR-ASM3 and DF-ASM3). The hybrid models were trained and subsequently validated using data from ASR-1 (R1), and ASR-2 (R2) which treated different compositions of petrochemical wastewater were utilized for additional validation of the developed hybrid models. Although predicted capacity for COD by MLR-ASM3 for R1 operation could be accepted (correlation coefficient of 0.78), it had a poor predictive capacity for the R2 operation (correlation coefficient of 0.48). Remarkably, DF-ASM3 exhibited superior prediction performance of both R1 and R2 as compared to MLR-ASM3 for MLSS, COD and nitrate with low MAPEs (all<25 %) and high correlation coefficients (all>0.7). This study underscored the effectiveness of hybrid modeling as a strong tool in minimising the labor required for biodegradability testing. Additionally, it demonstrated the improvement of the prediction capability of hybrid models for the treatment efficiency of petrochemical wastewater compared to mechanistic models.