To achieve precise prediction of pitch angles， this paper proposes a TCN-iTransformer model， named MFP-TCN-iTransformer， that integrates multi-flight phase （MFP） encoding. This method constructs a joint prediction architecture： The iTransformer module extracts global temporal features from quick access recorder （QAR） data and captures cross-variable dependencies， while the temporal convolutional network （TCN） module models multi-scale temporal dependencies of the pitch angle through dilated convolution. Additionally， MFP encoding is introduced， dividing the flight process into five phases to distinguish the data characteristics of different stages. Finally， a feature fusion mechanism is designed to combine discrete phase information with continuous QAR data， enhancing the model’s adaptability to phase characteristics. Experiments based on 264 352 pieces of QAR data show that the proposed model achieves an average improvement of 19.16% and 22.05% in mean absolute error （MAE） and root mean square error （RMSE）， respectively， compared to other benchmark models. Systematic ablation studies subsequently verify the effectiveness of each core component and confirm that refined flight phase encoding brings stable performance improvements. The results indicate that the model can achieve high-precision pitch angle prediction， which has practical value for enhancing flight safety.