Few-shot state-of-health prediction for lithium-ion batteries with LSTM network

This work proposes a novel solution for few-shot state-of-health prediction in lithium-ion batteries, addressing the critical challenge of data scarcity in new battery configurations. The proposed solution applies an LSTM network optimized through the model-agnostic meta-learning algorithm, enabling rapid adaptation to new batteries by fine-tuning the LSTM network with only 10% state-of-health data of ground truth. Particle swarm optimization algorithm is also used for automatic hyperparameter optimization, leading to improved training convergence and therefore enhanced prediction accuracy. In our work, historical SOH values and operational features such as current, voltage, and temperature are jointly analyzed through a novel attention module that adaptively integrates features from two LSTM blocks, producing more informative representations by capturing both temporal degradation patterns and context-dependent operational dynamics. By only requiring 10% SOH data from new batteies, our solution demonstrates a significant reduction in prediction errors over existing baselines on the NASA PCoE dataset and the Oxford dataset, enhancing model generalization by addressing the issue of data scarcity for new batteries from distinct configurations and operational conditions. Furthermore, our lightweight structure only consists of around 0.04M trainable parameters, and single-step inference only costs 6.1 ms in an Apple M3 CPU. All these characteristics make our solution highly suitable for real-world battery management systems, where effective real-time state-of-health predictions in resource-constrained environments are essential.