Probabilistic neural network-based flexible estimation of lithium-ion battery capacity considering multidimensional charging habits

Abstract

As the adoption of electric vehicles (EVs) continues to rise due to increasing environmental concerns and policy support, the accurate estimation of battery capacity becomes crucial for efficient vehicle management and prolonging battery life. This study presents a framework for capacity estimation under real-world electric vehicle charging scenarios, reflecting a variety of driving habits. A novel CNN-LSTM neural network architecture featuring a probabilistic regression layer (CLPNN) is developed, utilizing short time interval data to provide accurate, reliable capacity and uncertainty estimates for diverse charging situations. The adaptability of the proposed method is evaluated by intercepting real-vehicle statistics to simulate charging events. We introduce two result fusion strategies for refining capacity estimations: the Trimmed Mean Strategy and the Uncertainty Screening Strategy. Their performances are meticulously compared to establish the most effective approach. Moreover, probabilistic layer transfer learning is employed to address the challenge of insufficient labeled data in the target domain, thereby improving the transfer and generalization performance of the model. The findings of this study underscore the potential of the proposed methods in accurately evaluating battery capacity and dealing with uncertainties in real-world EV operations.