Accurate state of charge estimation for lithium-ion batteries based on cross-domain transfer learning with limited data

Accurate estimation of the state of charge (SOC) is critical for safe and reliable operations of lithium-ion batteries, while most SOC estimation methods often entail extensive offline testing data, significantly augmenting the development complexity. To reduce offline data requirements, a cross-domain transfer learning approach that integrates machine learning-based modeling with filter is proposed for SOC estimation. A source domain battery model is pre-trained based on temporal convolutional neural network and gated recurrent unit to excavate the battery's electrical characteristics. Moreover, transfer learning is exploited to transfer the weights of the network for the source domain to the target domain with limited data. The square root-cubature Kalman filter is developed to estimate the battery SOC by calibrating the error of voltage prediction. Substantial experimental results demonstrate that the proposed method can accurately estimate SOC under various temperature conditions for both nickel‑cobalt‑manganese and lithium iron phosphate battery chemistries, with the root mean square error, maximum absolute error, and mean absolute error of 1.27 %, 1.44 %, and 1.27 %, respectively. Compared to conventional methods, the proposed approach reduces the offline testing data requirement by 60 %, highlighting its potential for practical SOC estimation with limited target-domain data. • A cross-domain model is established based on TCN-GRU and transfer learning. • Square root-cubature Kalman filter is developed to estimate state of charge. • The fidelity of the method is validated under different driving conditions. • The proposed method can effectively shorten the offline testing cycle.