Abstract:

Accurate estimation of the state of charge of Li-ion batteries is required to guarantee the safe operation of battery systems. SOC estimation methods based on recurrent neural networks, like Gated Recurrent Unit, have recently received much attention because they can achieve accurate SOC estimation without using pre-defined battery models. However, due to their high computational complexity, these estimation methods are difficult to apply in engineering. To address the issues of high computational complexity caused by the large number of hidden state iterations required for SOC estimation in traditional GRU neural networks, a recursive update method with hidden state temporal succession is proposed, and it is possible to obtain the accurate SOC estimate at the current moment with only one network calculation of the sampled data at the current moment by improving the output structure of GRU networks. When compared to the traditional GRU method reported in the literature, this recursive GRU method can reduce the computational effort by more than 99% while maintaining SOC estimation accuracy, which has a better application prospect. Furthermore, in some application scenarios where there is a lack of battery training data, the method can combine migration learning to quickly complete network training. The method has been validated using laboratory test datasets and public datasets, and it is capable of performing accurate SOC estimation for different temperature environments, aging states, and Li-ion battery models, with a maximum estimation error of less than 3%.

Key words: lithium-ion battery, gated recurrent unit network, transfer learning, state of charge