This paper addresses the multiple control surfaces allocation problem for aircraft. To overcome inherent limitations of traditional control allocation methods, including significant actuator saturation and strong dependency on models, we propose a control allocation algorithm based on the autoencoder neural network. The encoder network maps desired virtual control commands (including force and moment components) to optimal actuator deflection commands, while the decoder network inversely reconstructs corresponding aerodynamic force and moment outputs to resolve model dependencies. The encoder added convolutional networks and residual networks to improve performance, and introduced control surface constraints into the activation function to reduce control surface saturation. Adding an energy loss term into the training optimization objective is used to reduce high-frequency fluctuations in the results.