Trajectory prediction of moving objects in flying state is an important technology involved in a large number of human-computer interaction tasks. Because uneven objects are easily affected by air resistance, Magnus effect and other factors in flight, it is difficult to carry out force analysis and realize exact trajectory prediction based on physical modeling. To solve these problems, we proposed a trajectory prediction method based on deep learning neural network. Multi-dimensional feature data of uneven flying object trajectory was collected by optical motion capture device, then an informer model was built to predict long-term sequence data. The model can effectively reflect the changing trend of the motion state of the object during flying and obtain the variable length pose data by model output. At the same time, we propose a valid method to optimize the structure of model, which can effectively lightweight the attention module in the encoder and decoder, change model generation mode to rolling output and improve the computing speed and accuracy of our model. Finally, we verified the improved model performance on the NMP public dataset, the experimental results show that our improved model has a better prediction effect than the original model and other methods such as Long Short-Term Memory network and Sequence to Sequence neural network. It gets an improvement in error accumulation which reduce 0.017 in MAE, 0.005 in MSE and 0.173 in MAPE compared to original model and has a good generalization performance in prediction task of other objects.