Data driven state of charge estimation for lithium ion batteries: Evaluating the influence of averaged input features using machine learning

For electric carbatteries to operate safely and dependably, a highly accurate State of Charge (SOC) is essential. While machine learning (ML) techniques have demonstrated superior performance over traditional methods, their effectiveness heavily depends on appropriate hyperparameter selection and input feature engineering. This study introduces a novel, systematic framework for SOC prediction by combining Bayesian optimization for automated hyperparameter tuning and voltage and current, alongside standard parameters, as model inputs. A key methodological advancement is the use of temporally averaged voltage and current measurements as model inputs, a simple preprocessing step that significantly enhances signal stability. Multiple ML algorithms like Neural Networks, Support Vector Machines, and Ensemble methods were evaluated under both raw and averaged input conditions across single and dual charge-cycle datasets. Results demonstrate that this feature-averaging approach, when combined with automated tuning, drives a dramatic improvement in accuracy, with models consistently achieving near-perfect performance (R² ≈ 1.00), with Ensemble methods exhibiting the highest robustness and lowest error rates. This comparative analysis establishes the clear superiority of the proposed preprocessing and tuning approach. The application of Bayesian optimization further enhanced model generalization and convergence stability. This work underscores the significance of input preprocessing and automated hyperparameter tuning in developing reliable, data-driven SOC estimation frameworks adaptable for real-time battery management systems.