A Classifier-Assisted and Bayesian-Optimized Filter Ensemble for Battery Model Parameter Calibration
Author ORCID Identifier
Semester
Summer
Date of Graduation
2026
Document Type
Thesis
Degree Type
MS
College
Statler College of Engineering and Mineral Resources
Department
Lane Department of Computer Science and Electrical Engineering
Committee Chair
Anurag K. Srivastava
Committee Member
Sarika Solanki
Committee Member
Muhammad Choudhry
Abstract
Grid-scale battery storage provides frequency regulation and peak shaving, and supports renewable integration, with a direct positive impact on grid operation and economics. Accurate calibration of lithium-ion battery equivalent-circuit model (ECM) parameters is a prerequisite for reliable terminal-voltage prediction, state-of -charge (SOC) and state-of- health (SOH) estimation, power-limit computation, and health-aware control in battery management systems (BMS). The internal parameters of an ECM, including ohmic resistance, polarization resistances, capacitances, and the associated RC time constants, drift over time with changing SOC, current, temperature, and aging. Fixed or time-varying sensor noise further degrades parameter calibration, and no single Kalman-filter-based estimator remains uniformly reliable across all noise levels and perturbation severities. This work develops a unified, robustness-oriented framework for offline parameter calibration of second-order RC (2RC) ECMs in three stages, with selected calibration analysis extended to third-order RC (3RC) ECMs. This study uses experimental HPPC- based ECM parameters with controlled synthetic noise-perturbation cases for robustness testing. First, a nonlinear state-augmented formulation is established, and the extended Kalman filter (EKF) and unscented Kalman filter (UKF) are derived and compared across perturbation and Gaussian-noise levels, establishing the algorithm-dependent strengths of each. Second, a classifier-assisted ensemble runs the EKF, UKF, and adaptive unscented Kalman filter (AUKF) in parallel, integrating their estimates through innovation-likelihood weighting, fuzzy rule-based prior weighting, and a dominance-aware parameter-fusion stage. A Random Forest classifier, trained on a nineteen dimensional feature vector extracted from voltage current profiles, identifies the operating condition and provides condition- aware priors to adjust the filter weights in the ensemble. Third, a machine-learning- assisted hyperparameter-optimization layer replaces heuristic tuning with surrogate-model- based Bayesian optimization to improve robustness under varying noise and perturbation conditions. Across six noise-perturbation conditions and multiple parameter bases, the ensemble reduced average MAPE by up to 82.7% under high-noise conditions and 73.3% under time-varying-noise conditions compared with EKF, while maintaining reliable voltage reconstruction and achieving 96.30% classifier test accuracy.
Recommended Citation
Sen, Varsha, "A Classifier-Assisted and Bayesian-Optimized Filter Ensemble for Battery Model Parameter Calibration" (2026). Graduate Theses, Dissertations, and Problem Reports (ETD). 13472.
https://researchrepository.wvu.edu/etd/13472