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.