A novel convolutional informer network for deterministic and probabilistic state-of-charge estimation of lithium-ion batteries

Accurate and real-time state-of-charge (SOC) estimation for lithium-ion batteries (LiB) is crucial for battery management system. However, the nonlinearity and complex dynamic properties of LiB pose a great challenge to the estimation of SOC. Some previous methods have undertaken high-precision SOC point estimation; however, the reliability of the estimated results has not been evaluated. These methods are too conservative and unreliable to describe the SOC sequence with random fluctuation characteristics by a certain number. In this study, a novel method for deterministic and probabilistic SOC estimation is proposed, called the Laplace distribution-based convolutional Informer network. The convolutional neural network (CNN) was used to extract spatial characteristics from the original input and enhance the ability of the model to capture sequential location information. An Informer network integrating the attention mechanism was used to learn the mapping relationship between these high-dimensional characteristics extracted by CNN and the SOC. This design makes the model suitable for fully extracting feature information from these data with complex temporal degradation properties. Considering the universality of measurement error and the importance of uncertainty estimation, the Laplace-based loss function was derived and used to train the proposed convolutional Informer network, reducing the influence of outliers on the estimated results and making the proposed model achieve uncertainty quantization. The effectiveness of the proposed model was evaluated on two public battery datasets. The results demonstrated that the proposed model produced accurate SOC point estimates and reliable interval estimates for different batteries under various operating conditions. •A novel convolutional informer network is designed for SOC estimation.•A loss function based on Laplace distribution is used to train the proposed model.•Deterministic and probabilistic SOC estimates can be obtained simultaneously.•Two case studies are conducted to verify the effectiveness of the proposed model.