A method for predicting lithium-ion battery capacity using an enhanced LightGBM model

Capacity grading testing is a crucial aspect of battery production in battery manufacturing, as it is closely linked to battery consistency. The capacity grading test is thus a vital part of the production process. The capacity of lithium-ion batteries is a crucial factor in their performance under varying current and voltage conditions. Currently, most battery capacity testing uses the experimental method, which consumes time, workforce, and material resources. This paper proposes an improved LightGBM model for battery capacity prediction using battery formation and partial capacity grading data, based on charge-discharge curves. This study utilizes a variable screening strategy, including Stepwise regression, Ridge regression, and LASSO regression, to select the characteristics that impact battery capacity from the charge/discharge curve. The battery capacity-affecting features are then fed into the improved LightGBM model. These features reflect the battery capacity characteristics from various perspectives. To address the hyperparameter optimization problem in LightGBM, this paper utilizes the PSO algorithm to optimize 11 crucial parameters. Determining the optimal parameter combination, eliminating the randomness of manual parameter tuning, and enhancing model stability. Battery capacity data from experiments were used to validate the predictive performance of the proposed method. Results demonstrate that this method achieves faster model training speeds and higher interpretability than deep learning approaches. Furthermore, tree-based ensemble learning yields superior battery capacity prediction accuracy, reliability, and robustness.