Interpretable lithium-ion battery degradation trajectory prediction via multi-task critical point detection and symbolic construction

Accurate prediction of lithium-ion battery (LIB) degradation trajectories is vital for ensuring reliability, safety, and optimal maintenance in energy storage applications. This paper proposes a two-phase, interpretable framework that combines multi-task learning and genetic programming to predict battery degradation trajectories from early-cycle data. In the first phase, a deep neural network combining convolutional layers and a Transformer architecture is trained in a multi-task learning setting to jointly predict the cycles and corresponding capacities at four critical points: knee onset, knee point, end-of-life and a newly defined last alarm point. In the second phase, a symbolic regression model based on genetic programming constructs the full degradation trajectory from these sparse critical points, ensuring accuracy and interpretability. Extensive experiments on a private dataset and the public battery dataset demonstrate that the proposed critical point detection method achieves accurate results. Furthermore, the subsequent trajectory degradation construction procedure not only offers accurate trajectories, but also facilitates the examination of degradation patterns through its symbolic expression, providing valuable interpretation for battery prognostics and health management.