Robust SOH Prediction for Lithium-Ion Batteries via ProbSparse Informer Architecture

초록

Accurate battery state of health (SOH) estimation is critical for ensuring safety and longevity of lithium-ion batteries used in electric and hybrid vehicles. In this work, we apply the Informer model, a Transformer variant with ProbSparse self-attention and attention distilling, to end-to-end SOH estimation using raw cycling data. Input features are first mapped to high-dimensional embeddings via value and positional encoding before being processed by the Informer encoder, which captures long-term dependencies through parallel distilling stacks. The decoder then integrates masked ProbSparse self-attention with encoder outputs to generate one-step SOH estimations. We evaluate performance on NASA and CALCE battery datasets using a leave-one-out strategy. Experimental results demonstrate that Informer consistently achieves high accuracy on unseen battery cells. Compared to LSTM and standard Transformer baselines, Informer outperforms both in estimation precision and generalization across datasets. These findings confirm that Informer provides a robust, end-to-end solution with superior performance for Battery Management Systems.

키워드