With the large-scale application of electric vehicles, accurate prediction of battery remaining capacity has become a core requirement to ensure driving range, safety, and economic efficiency. Traditional battery capacity degradation prediction methods mainly rely on aging experiment calibration or large-scale data-driven learning, making them difficult to adapt to complex real-world operating conditions. To address this issue, this paper proposes a mechanism-data hybrid-driven method for battery capacity loss prediction. On one hand, the Arrhenius model is employed to quantitatively describe the long-term dominant effects of temperature and cumulative ampere-hour throughput on capacity degradation from a mechanistic perspective. On the other hand, an LSTM network is used to capture dynamic perturbations in capacity loss under complex real-world operating conditions. Finally, the predicted battery capacity loss and its confidence interval are obtained. The proposed method is validated using charging data collected from electric vehicles under natural driving conditions, demonstrating its effectiveness in predicting capacity degradation in practical scenarios. Experimental results show that under the condition of training using only the first 30% of historical capacity degradation data, the method achieved a mean absolute error (MAE) of 0.73% and a root mean square error (RMSE) of 0.96% for subsequent capacity loss prediction, with the maximum error controlled within 2.18%. Overall, the prediction performance is superior to that of a standalone Arrhenius model and a pure LSTM model. The results indicate that the proposed mechanism-data hybrid prediction method can achieve high-precision and stable capacity loss predictions under real-world, complex onboard conditions, providing an engineering-applicable solution for assessing the health status and managing the lifespan of power batteries.