Multi-time Scale Attention Network for WEEE reverse logistics return prediction

In recent years, the reverse logistics (RL) management of waste electrical and electronic equipment (WEEE) has aroused widespread interest due to its vital role in improving ecological and economic benefits. As accurate quantification and estimate of future WEEE are fundamental to adequate planning and efficient treatment, we propose the Multi-time Scale Attention Network (MULAN) model in this paper for WEEE RL return prediction. In MULAN, we take multiple windows of historical data to capture distinct characteristics presented at different time scales and use attention-based alignment to fuse information. We then propose smooth time -series embedding method based on neighborhood data aggregation and temporal feature embedding method from four kinds of time positions to tackle the challenges of dealing with high noise data and high sensitivity to temporal dependencies. We apply MULAN to a WEEE recycling enterprise and conduct extensive experiments to demonstrate MULAN has significantly improved prediction accuracy over other baseline prediction methods in most cases.

Automated summary

In this paper, a Multi-time Scale Attention Network (MULAN) model is proposed for WEEE RL return prediction. MULAN captures distinct characteristics presented at different time scales and fuses information using attention-based alignment. Smooth time-series embedding method and temporal feature embedding method are proposed to address the challenges of high noise data and high sensitivity to temporal dependencies. Extensive experiments show that MULAN significantly improves prediction accuracy compared to other baseline methods in most cases.