Improving Stability Of Adversarial Li-Ion Cell Usage Data Generation Using Generative Latent Space Modelling

The quality and quantity of cell usage data (CUD) availability are crucial for reliable lithium-ion (Li-ion) battery modeling. Further, the model needs to encompass the non-linear and complex system dynamics, such as diverse aging mechanisms and dynamic operating characteristics. In general, the CUD acquisition from the electrochemical energy storage systems is a time-dependent, tedious, and lengthy, expensive process, which is often noise-corrupted with spurious outliers. Outliers’ robust, realistic synthetic CUD generation is essential for accelerating domain-specific technological developments. Time-series generative adversarial networks (TimeGAN) have been the state-of-the-art for latent space sequential data modeling by optimizing both the adversarial and supervised objectives while preserving the multivariate sequences’ temporal correlation dynamics [1]. The original TimeGAN formulation adopts the binary cross-entropy loss function, leading to vanishing gradient stability problems during the training process [2], [3]. Least-squares based formulation overcome such an issue without considering outliers influence [4]. In this treatise, some robust loss-functions for the TimeGAN architecture are explored for generating realistic Li-ion CUD. Extensive experimental validation on publicly avail-able datasets illustrates the amended TimeGAN framework’s improved stability w.r.t generator and discriminator scores.