In the complicated flight operation, the main factors that affect the flight phases are different. Based on the large amount data of Quick Access Recorder (QAR) used by B737NG aircraft, the flight segment is divided into cruise, climb and descent stages. The weight coefficients of different prediction models are determined by entropy weight method, and the combined prediction model of the whole flight is established. Pearson correlation coefficients are used to analyze and filter the modeling data, and stationary wavelet Rigorous SURE is used to pre-process and filter the data. Regression model is introduced to modify BP neural network (BP neural network) for the complex decline of flight state and the unsatisfactory prediction effect at ground stage. Entropy method is used to determine the dynamic weight coefficient, that is, combined with the flight phase to predict the segment, based on flight parameters to establish a combined forecasting model of fuel flow (FF) for the whole flight range. Through simulation analysis, the accuracy of the prediction model is verified by choosing the common and representative situation in the flight. The error range is within 3.5%, which proves that the model is reasonable and has a wide range of application.