A Scalable Recurrent Structure with Fast Transfer Learning for Lithium-Ion Battery State of Charge Estimation at Different Ambient Temperatures

State-of-charge (SOC) is a critical parameter of battery management systems to ensure safe, efficient, reliable and durable battery operations. However, SOC cannot be directly measured and is highly sensitive to different temperatures. Thus, the SOC estimation accuracy and uncertainty management are significant for robust control and energy dispatch. Gaussian process regression (GPR) is thus becoming appealing due to its non-parametric and interpretable probabilistic advantages. But the scalability of GPR is still challenging, suffering from cubic complexity. To solve these challenges, a SOC estimation method is proposed by an end-to-end scalable deep recurrent structure with fast transfer learning at different temperatures. First, convolutional and recurrent neural networks are employed to catch nonlinear temporal dependency within measurements. Second, a GPR layer is concatenated after neural networks, so that the estimation can be quantified with uncertainty while retaining nonlinear expressive ability. Then, a non-parametric fast transfer learning is designed to realize fast transfer between different temperatures. Next, structured sparse approximations and a semi-stochastic gradient procedure are established for scalable training. Finally, the accuracy and fast transfer of the proposed structure are verified through comparison. The structure demonstrates the state-of-the-art performance on estimation accuracy and efficiency with transfer learning faster than fine-tuning strategy by two orders of magnitude.