Work

Performance–Complexity Trade‐Offs in Battery Lifetime Prediction with Task‐Aware Transformers

Year: 2026

Type: article

Abstract: Accurate battery lifetime prediction is essential for improving reliability and safety in energy storage systems. However, balancing predictive accuracy, inference latency, and energy consumption remains challenging. We introduce FAST-BatPro, a Flash-Attention Sparse Transformer for Battery Prognosis. This task-aware architecture combines convolutional feature extraction with dual attention mechanisms to enable robust, efficient, and scalable prediction. The 1.869-million-parameter model is evaluated from both perspectives. Across four datasets comprising more than 240,000 cycles across chemistries, FAST-BatPro demonstrates consistent performance across fast-charging and discharging protocols, temperature variations, and chemistry-dependent degradation behaviors. With limited early-cycle data, it achieves high accuracy, with coefficient of determination values approaching or exceeding 0.90 in most test settings. It maintains an inference time of 0.103 s and requires 1.65 billion FLOPs per battery over the full lifecycle. Hidden-dimension scaling identifies a compact configuration that preserves comparable accuracy while reducing inference latency, FLOPs, and energy consumption by 12.6%, 68.1%, and 12.6%, respectively. Module-level pruning further shows that a feed-forward-network-pruned lightweight variant improves accuracy while reducing inference latency, FLOPs, and energy consumption by 11.65%, 54.64%, and 11.65%, respectively, indicating that FAST-BatPro can serve as a reference architecture for identifying task-specific redundancy and guiding efficient AI model design for battery diagnostics and predictive maintenance.

Source: Advanced Science

DOI: 10.1002/advs.202524179