An innovative multitask learning - Long short-term memory neural network for the online anti-aging state of charge estimation of lithium-ion batteries adaptive to varying temperature and current conditions

As the new industrial revolution accelerates, new energy storage systems are becoming increasingly vital to the industrial chain. The overall performance of the battery management system can be improved by using a long short-term memory neural network online multi-task learning model based on physical model constraints to estimate the state of charge of the lithium-ion battery accurately. This paper uses a model reference adaptive system to perform parameter identification on the equivalent circuit model to avoid inaccurate SOC-OCV voltage curves caused by battery aging. At the same time, drawing on the ideas of PINN neural networks, the equivalent circuit model is used as the physical information constraint of the long short-term memory neural network, which improves the model's generalization ability. Finally, the Mahalanobis distance is used for offset determination, and an online learning method is used to improve model robustness. A multi-condition aging experiment on a lithium-ion battery showed that the proposed model improved the accuracy of equivalent circuit model parameter estimation by up to 67.38 % and state of charge estimation by an average of 59.73 %. This work introduces a new model design to drive innovation in battery management systems.