Performance prediction of an internal-circulation membrane bioreactor based on models comparison and data features analysis

Abstract

Complex hydrodynamic conditions and the dynamic process of multi-habitats have brought challenges for the real-time management of novel wastewater treatment techniques. Thus, based on a well-performing membrane bioreactor with internal circulation, this paper presented an intelligent data mining work of performance prediction and the analysis of data influence for efficient management. The experimental data with 440 samples and nine internal environment parameters were integrated, and the concentrations of COD, NH4+-N and total nitrogen (TN) in the effluent were used as multiple performance indicators. Then, a comprehensive testing was utilized for recognizing the most robust model from the deep neural network, regression tree, random forest, adaptive boosting and gradient boosting regression (GBR), which included the evaluations of model fitting and generalization ability. Furthermore, the feature relative importance (Imp) and principal component analysis (PCA) was applied to analyze the influence of data. The results indicated that the GBR model had a higher performance, in which, the R squared scores of 5-fold cross-validation were 0.847, 0.792 and 0.851 in predicting the concentration of COD, NH4+-N and TN, respectively. Besides, the quantification of Imp was accorded with the independent result of PCA, which indicated that the GBR model had well captured the dynamic information.