Electrochemical Theory Guided Modeling of the Conditional Generative Adversarial Network for Battery

In many energy storage applications, the inevitable calendar aging of batteries results in both short service life and decreased battery performance. In this article, a generative adversarial network-based (GAN-based) model is proposed for both point and probabilistic forecasts of battery calendar aging, i.e., the capacity forecast GAN (CFGAN), which will be the first work that applies GAN to calendar aging forecast. GAN’s ability to learn arbitrarily complex distributions has enabled CFGAN to approximate all the possible (arbitrarily shaped) joint distributions. By taking electrochemical knowledge as the guidelines for designing CFGAN’s crucial part, i.e., the conditioner, CFGAN has maintained a satisfying consistency between knowledge and data, making it both knowledge-driven and data-driven, i.e., knowledge+data-driven, which has improved its theoretical strength and forecast performance significantly. Illustrative results on practical calendar aging case studies demonstrated the superiority of CFGAN in forecasting and generalizing to unwitnessed conditions, implying that the CFGAN built in deep structure has grasped the complex multimodality of the condition-varying calendar aging process.