Lithium-ion batteries with potential safety hazards require accurate forecasts of their remaining useful life to be maintained and replaced. An online forecasting method is proposed, in which the training dataset before a pre-set operating time for the start of forecasting is automatically extended in order to train support vector regression model and then the key hyperparameters of the model is optimized by tree-structured parzen estimator algorithm to update the model used to forecast battery capacity for the next cycle. Verified by NASA lithium-ion batteries datasets, this method provides higher forecasting accuracy and generalization ability, which is suitable for the scenario of online accurate short-term forecasting based on a limited amount of historical data.