A hybrid transfer learning scheme for remaining useful life prediction and cycle life test optimization of different formulation Li-ion power batteries

Long-term cycle life test in battery development is crucial for formulations selection but time-consuming and high-cost. To shorten cycle test with estimated lifespan, a prediction-based test optimization method is proposed for Li-ion batteries with different formulations. A hybrid transfer-learning method optimally selects historical test data and trained prediction model of other formulations to help construct models of the target batteries. It can improve prediction accuracy despite short-term test data containing insufficient global degradation information. Firstly, a four-step transferability measurement method automatically selects the most transferable sample from a historical database of other formulations, although their degradation laws exist individual differences and inconsistency. Four-types of transferability evaluation criteria including curve shape, long-term degradation rate, lifespan concentration, and distance between curves, are sequentially integrated to fit capacity curves characteristics and long-term prediction. Then, a prediction model using Long Short-time Memory Network is quickly initialized by transferring a shared part of the previous model of other formulations instead of random initialization. The shared model parameters are optimally and selectively transferred according to test temperature and test data amount for improving modeling effectiveness. The rest-part of the model is trained by the selected transferable-sample to learn degradation trend similar to the target battery for accurate prediction. Finally, actual data from a battery company verify the performance of the proposed method in terms of prediction and cost-saving. It achieves 89.18% average accuracy and 0.7 to 5.5 months saving under the condition of different formulations and test-stop threshold.

Automated summary

The proposed method utilizes transfer-learning and prediction modeling to optimize test data selection and model training, improving long-term cycle life prediction accuracy for Li-ion batteries.