A prediction-based cycle life test optimization method for cross-formula batteries using instance transfer and variable-length-input deep learning model

Cycle life is a key performance indicator in the design and development of lithium-ion power batteries. In order to obtain an appropriate formula, developers need to conduct a large number of cycle life tests (CLTs). However, the high test cost and unbearable time overhead of CLT have seriously hindered the upgrade and development of lithium-ion power batteries. In this paper, a prediction-based CLT optimization method for cross-formula batteries is proposed, which can shorten the number of test cycles by predicting the remaining cycle life of batteries. Specifically, we design an AED-based instance transferability measurement method to select reference battery from the historical database according to curves distance and trend consistent. Then, a highly robust deep learning method named variable-length-input stacked denoising autoencoder (VLI-SDA) is proposed to achieve remaining useful life prediction. The VLI-SDA model adopts a variable-length input strategy to expand the receptive field, fully learn the degradation trend, and ensure an appropriate number of training samples. Combined with the inherent noise reduction capability of the SDA model, the VLI-SDA model can effectively solve the problem of cycle life prediction under high-temperature stress test and small sample conditions. The actual CLT data at three temperatures from a battery company verify the effectiveness of the proposed method. The test temperature, curve shape and other influencing factors are analyzed to help determine optimization strategies.