1D convolutional neural network combined with long-short term memory-based cell-to-cell inconsistency estimation considering recursive least squares and Euclidean distance

Abstract

As the scale of application utilizing batteries as a power source expands, operation at the cell-tray and pack level has become indispensable. Such expansion amplifies the importance of accurate and consistent monitoring of cell-to-cell inconsistencies during battery operation for efficient and safe use. Consequently, this study addresses the limitations associated with inconsistency evaluations that rely solely on voltage deviation, which may lead to inaccurate inconsistent information, contingent on operational conditions and assessment timelines. By integrating the internal dynamics of an RLS-based 24S1P cell-tray, the aim is to derive deviations in electrical equivalent circuit model (EECM) parameters between cells, facilitating real-time battery inconsistency evaluations. The extracted EECM parameters are quantified utilizing the Euclidean distance for an effective quantification and scoring of multi-variables for inconsistency assessment. Additionally, to alleviate the computational burden associated with deriving the inconsistency evaluation index, an approach has been adopted that learns 60 % of the data prior to the occurrence of inconsistency through model-based index training, thereby verifying the results of the inconsistency estimation. In this process, a 1D CNN-LSTM model is developed, capable of proficiently learning and estimating the patterns of data changes. Finally, by benchmarking performance against traditional imbalance evaluations based solely on voltage deviation, this study demonstrates the robustness and precision of the proposed methodology, underscoring its capability for accurate imbalance evaluation, irrespective of SOC segment or evaluation timing.