Adaptive multi-domain capacity estimation for battery energy storage system based on multi-scale random sequence feature fusion

Monitoring battery capacity degradation in lithium-ion battery energy storage systems (BESSs) is crucial for ensuring safe and reliable operations. However, conventional data-driven methods primarily focus on single-domain estimation and feature engineering from fixed charging/discharging stages, limiting their adaptability in real-world scenarios. Therefore, this paper proposes an adaptive multi-domain capacity estimation method for BESSs based on multi-scale random sequence feature fusion. Firstly, this paper proposes the adaptive multi-domain capacity estimation theory, which utilizes the Pearson correlation coefficient (PCC) for health feature screening and maximum mean discrepancy (MMD) for domain discrepancy identification and domain classification. Secondly, an optimal random sequence feature is proposed based on short-duration raw voltage and incremental capacity, considering the effects of both sampling interval and duration. Subsequently, a multi-scale convolutional neural network (MSCNN) is developed to fuse ageing information from the random sequence feature and enable accurate adaptive multi-domain capacity estimation. Finally, the validation is conducted using 130 batteries operating under various working conditions, and it shows the proposed method is more robust compared to the single-domain estimation. The overall RMSE and MAE are reduced to within 1.53 % and 1.18 %, with the overall R2 value up to 99 %. This demonstrates the superiority of the proposed method for real-world applications.