Accurate state-of-charge (SOC) estimation is essential for range prediction, power allocation, and operational safety in electric vehicles. Low-temperature operation introduces coupled effects of capacity degradation and polarization dynamics, which challenge conventional fixed-capacity models and single-timescale estimation methods. This paper proposes a hierarchical multi-timescale SOC estimation framework that combines data-driven capacity prediction, online parameter identification, and nonlinear state estimation. At the upper layer, a temperature-aware temporal self-attention long short-term memory network (TS-LSTM) estimates the available battery capacity. At the lower layer, forgetting-factor recursive least squares (FFRLS) and a gain-scheduled unscented Kalman filter (UKF) jointly track impedance variations and recursively estimate SOC. The framework is validated on public lithium-ion battery datasets at 4, 24, and 43 °C under pulse and sparse high-current conditions. Under the challenging 4 °C scenario with a 10% initial SOC bias, the method achieves an SOC root mean square error (RMSE) of 1.35%. Additional perturbation tests yield SOC RMSEs of 1.35–2.40% under −5% to +10% initial-SOC offsets, 1 mV voltage noise, a +2 °C temperature bias, and a +20% impedance-prior error. Furthermore, the computational overhead remains substantially lower than the data sampling interval, indicating promising potential for online estimation in battery management systems.