Deep learning-based screening approach for priority pollutants: a case study on retired power battery recycling

Abstract

With the rapid increase in the production of retired power batteries, the potential environmental risks during recycling must urgently be identified and assessed. This study presented a novel screening framework for pollutant prioritization utilizing deep learning algorithms coupled with hierarchical clustering analysis. An integrated model for pollutant screening called McA was constructed based on five deep learning methods with performance-based weighting. Compared to traditional machine learning models, both the accuracy and reliability of the McA model were significantly improved (R² = 0.9999, MSE = 0.300, and MAE = 0.220 for the test set). By applying this approach to the retired power battery recycling, 13 pollutants were identified and divided into four priority levels: level I (highest priority), including 1 pollutant; level II (high priority), including 6 pollutants; level III (medium priority) including 1 pollutant; level IV (low priority) including 5 pollutants. Finally, SHapley Additive exPlanations (SHAP) visualization was performed to reveal the differences in risk priority by identifying the primary influencing factors, including acute toxicity, irritation and corrosivity, and endocrine disruption. The results of the study provide constructive schemes and insights for screening priority pollutants in the recycling process of retired power battery, suggesting the high potential to develop and implement deep learning methods in pollutant prioritization and risk management.