A degradation trajectory prediction method applicable to various life stages of lithium-ion batteries under complex variable aging conditions

It is imperative to effectively utilize the limited measurement data available at any life stage to address the challenge of predicting the degradation trajectory of lithium-ion batteries under various operating conditions throughout their lifespan. This study presents a method for incorporating base points (points obtained that can be interpolated to predict the degradation trajectories) by applying a convolutional neural network, which learns the mapping relationships from charging data from any stage, even down to a single cycle. The objective is to accurately determine the specific locations of these base points appropriate for interpolation and extrapolation of degradation trajectories. This study utilizes three metaheuristic algorithms to optimize the base point locations and model hyperparameters using different input data to achieve optimal model performance. The proposed method is validated on two independent datasets: A public dataset containing 95 LFP cells and a private dataset comprising 14 NCA cells. The experimental conditions for the NCA test set fluctuate with aging, thereby simulating real-world scenarios. In addition, selection criteria for the input data and specific optimization algorithms are proposed to maximize the potential for achieving optimal accuracy. The results indicate that the minimum RMSE (and MAE) for LFP degradation trajectory predictions can attain 0.67 % (0.51 %) across various base point locations and hyperparameter combinations. These values can improve for the NCA test set to as low as 0.30 % (0.25 %). Furthermore, the optimal strategy in a relatively stable testing environment, such as a laboratory environment, involves the application of the differential evolution (DE) algorithm utilizing a limited number of base points. In addition, the present method is not sensitive to measurement quantity and the life stage at which the input data is located • A model-free method for degradation prediction in limited samples is proposed. • Method validation utilizing multiple datasets and mixed measurement conditions. • Predictions can be achieved with limited data from any life stage. • Multiple trend calibration improves method applicability to complex environments.