Towards fast multi-scale state estimation for retired battery reusing via Pareto-efficient

Abstract

With the exponential increase in the adoption of lithium-ion batteries, reusing and recycling have become critical for extending the lifespan of retired batteries and reducing environmental impact. Recent developments in deep learning provide efficient solutions for the screening and reuse of massive retired batteries, as they can estimate multiple battery states in a short observation period. However, the existing methods ignore the timescale differences between battery states, causing the model to collapse in optimization conflicts. In this paper, we revisit the impact of this conflict and propose a dual-path deep method for fast estimation of the state of both charge (SOC) and health (SOH) in a short observation time of the discharge phase. Specifically, the shared lower layers capture local time-varying features, while the two specialized paths integrate them into global features each focusing on a different timescales. Furthermore, to solve the ensuing optimization conflict, we seek for Pareto-efficient to achieve the optimal estimation of the two states. Exhaustive experiments and analysis on 89 realistic retired batteries and 16 public batteries with different chemistries and working conditions show that our framework can obtain reliable estimation. Using only an observation time of 400 s, the average root mean square error of SOC and SOH estimations is 1.01% and 1.72%, improved by 16% and 33% compared with state-of-the-art methods. Notably, our framework only has a parameter size of 0.0542 MB and can be deployed on most edge devices, which significantly promotes the application of data-driven models in the real world.