Carbon source dosage intelligent determination using a multi-feature sensitive back propagation neural network model

Abstract

The carbon reduction concept drives the development of low-carbon and sustainable wastewater treatment plant (WWTP) operation technologies. In the denitrification stage of WWTPs in China, there are widespread problems of uneconomical dosage consumption and unstable total nitrogen (TN) concentration in effluent through manual experience to add external carbon sources. Deep learning methods can deal with these problems. However, the methods often require a large amount of data. This paper establishes a multi-feature sensitive back propagation neural network (BPNN) based on Shapley additive explanations (SHAP) and sensitivity analysis (MFS-BPNN-SSA) model to predict carbon source dosage in WWTPs and address short-term and limited data. The model also incorporates theoretical formulas to enhance prediction accuracy and feedback regulation to handle anomalous data. The prediction performance of the MFS-BPNN-SSA model surpasses traditional machine learning and deep learning models. R and R² reach 0.9999, 1.75% and 3.48% higher, respectively, compared to the best-performing traditional model. The model has been operating safely in the WWTP for over two years, achieving a 9% improvement in effluent TN concentration and a 14% reduction in carbon source dosage. This study provides a novel strategy for pollution reduction and carbon mitigation in WWTPs.