Accurate prediction of lithium battery lifespan is crucial for the proper functioning of electrical equipment. However, predicting battery lifespan accurately is challenging due to the nonlinearity of capacity degradation and the uncertainty of operating conditions.
Recently, Prof. CHEN Zhongwei and Ass. Prof. MAO Zhiyu from the Dalian Institute of Chemical Physicsof the Chinese Academy of Sciences, in collaboration with Prof. FENG Jiangtao from Xi'an Jiaotong University, designed a novel deep learning model, referred to as the Dual Stream-vision Transformer-efficient Self Attention mechanism (DS-ViT-ESA), which overcame the traditional reliance on extensive charging test data.
This model provided new insights for real-time battery life prediction and enabled end-to-end evaluation of lithium battery lifespan. Additionally, it is a key component of the team's first-generation Battery Digital Brain core algorithm, known as Power Battery Systems Research Department Digit (PBSRD Digit), which offers solutions for intelligent battery management.
Detailed illustration of DS-ViT-ESA model (Image by LIU Yunpeng)
The researchers developed the deep learning model using a small amount of charging cycle data. This model employed a vision transformer structure with a dual-stream framework and an efficient self-attention mechanism to capture and integrate hidden features across multiple time scales. As a result, it was able to accurately predict the battery's current cycle life (CCL) and remaining useful life (RUL). With just 15 charging cycle data points, the model achieved RUL and CLL prediction errors of only 5.40% and 4.64%, respectively. Moreover, the model maintained low prediction errors even when tested on charging strategies not included in the training dataset, demonstrating its zero-shot generalization capability.
This battery lifespan prediction model was also a crucial component of the first-generation Battery Digital Brain, called PBSRD Digit. By integrating this algorithm, the system's accuracy was significantly improved. Currently, the Battery Digital Brain system served as the core energy management system for large-scale commercial storage and electric vehicles, with the capability to be deployed on both cloud servers and client-side embedded devices.
"The battery lifespan prediction model effectively balances prediction accuracy with computational cost, thereby increasing the application value of the Battery Digital Brain for lifespan estimation," said Prof. CHEN. "We plan to further optimize the model using techniques such as model distillation and pruning, aiming to enhance robustness and resource utilization of the system," Prof. CHEN added.