The remaining useful life (RUL) is the key element of fault diagnosis, prediction and health management (PHM) during the equipment operation service period. The prediction result of RUL is the premise for equipment to adopt preventive maintenance, condition-based maintenance, fault maintenance and other maintenance strategies. Lithium battery is an important energy component of new energy vehicles, mobile phones, etc. Its RUL is related to the state of its equipment system. Many model-based methods have been used to predict the lithium batteries’ RUL, and some studies have begun to use lithium battery monitoring data to predict its remaining service life. With the continuous detection and monitoring capability of equipment throughout its life cycle gradually improved, a large number of monitoring and detection data promote the wide application of data-driven residual life prediction in the field of equipment. At present, the data-driven prediction method of the lithium batteries’ RUL mostly adopts a single time-series forecasting model. The robustness and generalization of the prediction method are insufficient. It needs to be further improved to improve the prediction accuracy and robustness. Preventive maintenance measures shall be taken immediately according to the prediction results to ensure the effective supply of energy at any time. In this paper, an integrated learning algorithm based on monitoring data is proposed to fit the degradation model of lithium batteries and predict their RUL. The ensemble learning method consists of 5 basic learners to achieve better prediction performance, including relevance vector machine (RVM), random forest (RF), elastic net (EN), autoregressive model (AR), and long short-term memory (LSTM) Network. The genetic algorithm (GA) is used in the ensemble learning method to find and determine the optimal weights of the basic learners, and obtain the final prediction result of lithium batteries. Then, the simulation is carried out on the CS2_35 lithium battery data set. The simulation results show that the method proposed in this paper has a smaller Root Mean Square Error (RMSE) than another 5 single methods. The RMSE is respectively 0.00744 for RVM, 0.01097 for RF, 0.01507 for EN, 0.03223 for AR, 0.01541 for LSTM, and 0.00483 for ensemble learning, and the RMSE of ensemble learning is reduced by 0.0274 at the highest and 0.00261 at the lowest, so the ensemble learning algorithm has better robustness and generalization effect.