Data-Driven Prediction Methods for Lithium-Ion Battery State of Health Based on Elbow Rule

Lithium-ion batteries are extensively utilized in diverse sectors such as automotive applications. The imprecise estimation of the State Of Health (SOH) will significantly impact safe operations and cost reduction initiatives. Addressing challenges in extracting aging features and the complexity of modeling, this study proposed a feature parameter identification method leveraging Incremental Capacity Analysis (ICA) technology and the elbow rule. This approach employed data-driven methodologies to predict battery health status. Initially, voltage capacity and incremental capacity curves were obtained, and data-driven principles were employed along with data cleaning to mitigate noise. Subsequently, correlation and significance analysis methods were applied for preliminary health feature selection. To eliminate data redundancy, a novel principal component analysis strategy based on the elbow optimization rule was introduced. Next, two data-driven estimation models, namely BP (Back Propagation) neural network and Gaussian Process Regression (GPR), were established. These models were employed for SOH prediction. Comparative results demonstrate that feature parameters extracted using the elbow rule from both models closely approximate reality, validating the correctness and accuracy of the feature extraction method. Specifically, the GPR model exhibited higher prediction accuracy. Furthermore, the GPR model was applied to predict SOH for an additional set of three batteries. The findings reveal a relative error of approximately 4% when maintaining lithium-ion battery SOH at 80%, affirming the GPR model's high accuracy and robust adaptability for SOH prediction.