A highly accurate long short-term memory (LSTM)-Gaussian process regression (GPR) multivariate hybrid prediction model architecture was proposed to address the challenges of strong nonlinearity, significant time-series dependency in battery degradation, and the lack of uncertainty quantification in conventional models. The method utilized a LSTM network to process multi-dimensional time-series data (voltage, current, temperature, etc.). The LSTM effectively captured complex time-series dependencies and automatically extracted deep degradation features. Subsequently, GPR was introduced, operating on the extracted 30-dimensional feature vectors. GPR’s innovation lay in its ability to simultaneously output the 95% confidence interval, achieving reliable uncertainty quantification. Validation on the National Aeronautics and Space Administration (NASA) battery dataset demonstrates superior performance: the minimum coefficient of determination (R²) reaches 0.997 5, with the maximum root mean square error (RMSE) and mean absolute error (MAE) being only 0.004 99 Ah and 0.004 63 Ah, respectively. The performance metrics significantly surpass those of the comparative benchmark models, namely the convolutional neural network (CNN) and support vector regression (SVR). Research confirms that the proposed LSTM-GPR model, by combining high-precision point prediction with reliable uncertainty quantification, provides a viable solution for online state-of-health (SOH) prediction within battery management systems (BMS).