Ultra-Long-Term Early Prediction of Lithium-Ion Battery State of Health Based on Time-Series Decomposition and Learned Physics Constraints

Accurately predicting the State of Health (SOH) of batteries is of great significance. Although data-driven SOH prediction methods have made some progress, ultra-long-term early prediction still faces challenges such as insufficient short-term input information, error accumulation, and poor universality of traditional physical modeling. To address these issues, this chapter proposes the DFLPformer prediction framework. This framework employs Seasonal-Trend decomposition using Loess (STL) and Fourier transform to enhance the prediction robustness, and utilizes the Bidirectional Long Short-Term Memory network (BiLSTM) to learn physical constraints, thereby enabling the grasp of the long-term degradation trend of battery capacity. Comparative experiments show that, compared with the Informer model, in two different battery charge-discharge datasets, the Mean Squared Error (MSE) of the algorithm proposed in this paper is reduced by 52% and 49% respectively, and the Coefficient of Determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) is increased by 3% and 7% respectively, which verifies the effectiveness of DFLPformer.