Flight recommendation is to predict the next flight by capturing the passenger travel behavior history. Although existing methods have achieved convincing results in flight recommendation tasks, they ignore the temporal patterns of different periodicity and evolution of passenger preference. In this paper, we propose a novel Multi Hierarchical Time Structures aware Passenger Preference Evolution model for Personalized Flight Recommendation (MTER), which not only uses the time structure of the calendar system to model temporal patterns of passenger multiple behaviors but also attempts to use passenger attributes to alleviate the problems of cold-start and new passengers. Our model divides a passenger behaviors into sessions to form flight-session bipartite graph, and then generates session embeddings by aggregating the corresponding flight embeddings in them, which are further aggregated into the embeddings of different time units. In addition, MTER also captures the passenger attribute embeddings to evolve passenger preference to alleviate the problems of cold-start and new passengers. The latent passenger representations are generated by concatenating all temporal patterns, and attribute embeddings of passengers, which are used to predict the next flight and behavior through a polynomial decoder and a Softmax layer. To the best of our knowledge, this is the first attempt to recommend flight by using multi-hierarchical time structures aware passenger preference evolution model. We conducted experiments on real civil aviation dataset to show the effectiveness of our model.