다중 인코더 트랜스포머를 활용하여 용량 재생 현상을 고려한 배터리 성능 상태 예측 방법

Abstract

This paper proposes a Multi-encoder Transformer based battery SOH (State of Health) prediction method considering the CRP (Capacity Regeneration Phenomenon). CRP is a nonlinear effect in which battery capacity temporarily increases after an inactive state. Legacy battery SOH prediction models do not account for CRP, resulting in less accurate SOH predictions. To address this issue, we propose to decompose the battery capacity signal into its high-frequency and low-frequency components using CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise). In the proposed method, the Multi-encoder Transformer is employed such that each Transformer encoder, deployed in parallel, predicts value of each IMF (Intrinsic Mode Function). The predicted IMFs are concatenated and then passed to the Transformer decoder for SOH prediction. Moreover, the dual attention mechanism is used to evaluate the relative importance of time series sensor data for SOH prediction. Experimental results on the NASA and CALCE Li-ion battery dataset show that the proposed method achieved better accuracy across several evaluation metrics and provided interpretable attention weights. This study is expected to contribute to more effective battery condition monitoring, improved predictive maintenance, and overall improvement in energy management efficiency.