For power system transient stability assessment, the characteristic differences of critical samples are not obvious in the early stage after fault clearance, and the prediction accuracy is low. Over time, the evaluation accuracy improves, while the timeliness of the evaluation is difficult to ensure. Aiming at the contradiction between the accuracy and timeliness of transient stability assessment, a time-adaptive transient stability assessment method of power system based on ensemble learning is proposed. First, the unbalanced data are sampled by the EasyEnsemble algorithm. Ensemble LSTM classifiers with different evaluation cycles are trained. Thus, the stability prediction results of samples in different evaluation cycles are output. Second, the evaluation moments are divided and multi-stage threshold classification rules are proposed. The threshold is adjusted adaptively to evaluate the reliability of the prediction results. Finally, the samples whose prediction results are evaluated as unreliable are handed over to the model of the next evaluation cycle for judgment until the reliability reaches the threshold. The simulation results in the IEEE39 bus system show that the proposed method has better performance than other time-adaptive methods; in the case of unbalanced samples, the method achieves a better correction effect. This work is supported by the National Natural Science Foundation of China (No. 51707040).