Misker Shumeye

Mr. Misker Shumeye


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
MSMisker Shumeye
Misker Shumeye
EYEliyab Yoseph
Eliyab Yoseph
KDKejela Adane Dulecha
Kejela Adane Dulecha
Accurate state-of-charge (SOC) estimation is central to battery management, yet equivalent-circuit filters degrade when operating conditions drift from those at which they were parameterized, while purely data-driven estimators can fail outside their training envelope. This study compares five SOC estimators on a simulated lithium-ion cell with a richer ground-truth model than any estimator uses (two RC branches, hysteresis, temperature-dependent resistance and capacity, self-heating, and sensor bias and noise): Coulomb counting, an extended Kalman filter (EKF) with parameters identified at 25 °C, an EKF with Arrhenius-type temperature-scheduled parameters (EKF-TS), a feed-forward neural network (data-driven), and a hybrid that adds a neural-network voltage-residual correction to the EKF measurement equation. All methods were evaluated on 300 held-out runs (three drive-cycle-like profiles, five ambient temperatures, 20 seeds), including two temperatures absent from training. The hybrid reduced the RMSE of the 25 °C EKF from 7.24% to 1.37% SOC and outperformed the data-driven network overall (1.31 ± 0.08% versus 1.49 ± 0.17% across ten training seeds). At the training temperatures it was also better than EKF-TS (0.90% versus 1.10%), but at unseen temperatures the physics-structured EKF-TS generalized better (1.34% versus 1.94 ± 0.22%). A hybrid built on top of EKF-TS matched EKF-TS overall (1.17 ± 0.09% versus 1.20%). Ablation showed that the temperature input to the residual network is essential. The results are simulation-only and should be confirmed on experimental data.