Suggested Citation:
Lv, Zhilong and Jingyuan Zhao (2025)

Resource-Efficient Artificial Intelligence for Battery Capacity Estimation Using Convolutional FlashAttention Fusion Networks

Accurate battery capacity estimation is crucial for optimizing lifespan and monitoring health conditions. Deep learning has made notable strides in addressing long-standing issues in the artificial intelligence community. However, large AI models often face challenges such as high computational resource consumption, extended training times, and elevated deployment costs. To address these issues, we developed an efficient end-to-end hybrid fusion neural network model. This model combines FlashAttention-2 with local feature extraction through convolutional neural networks (CNNs), significantly reducing memory usage and computational demands while maintaining precise and efficient health estimation. For practical implementation, the model uses only basic parameters, such as voltage and charge, and employs partial charging data (from 80 % SOC to the upper limit voltage) as features, without requiring complex feature engineering. We evaluated the model using three datasets: 77 lithium iron phosphate (LFP) cells, 16 nickel cobalt aluminum (NCA) cells, and 50 nickel cobalt manganese (NCM) oxide cells. For LFP battery health estimation, the model achieved a root mean square error of 0.109 %, a coefficient of determination of 0.99, and a mean absolute percentage error of 0.096 %. Moreover, the proposed convolutional and flash-attention fusion networks deliver an average inference time of 57 milliseconds for health diagnosis across the full battery life cycle (approximately 1898 cycles per cell). The resource-efficient AI (REAI) model operates at an average of 1.36 billion floating point operations per second (FLOPs), with GPU power consumption of 17W and memory usage of 403 MB. This significantly outperforms the Transformer model with vanilla attention. Furthermore, the multi-fusion model proved to be a powerful tool for evaluating capacity in NCA and NCM cells using transfer learning. The results emphasize its ability to reduce computational complexity, energy consumption, and memory usage, while maintaining high accuracy and robust generalization capabilities.