The invention relates to an aircraft neural network learning control method based on lumped composite estimation. Firstly, a longitudinal channel model of an aircraft is decoupled into a speed subsystem and a height subsystem; dynamic inverse control is used for the speed subsystem, and backstepping control is used for the height subsystem; a neural network is used for estimating the dynamic uncertainty of the system, and the coupling from an elastic mode is estimated by a nonlinear observer; a lumped forecast error of two estimations is constructed based on online data, and the forecast erroris applied to updating laws of the neural network and the nonlinear observer; height and velocity controllers are given based on two types of estimator information to achieve the tracking of height and velocity, and finally the control method is applied to the elastic model of the hypersonic aircraft.