The invention discloses a fixed-wing UAV (unmanned aerial vehicle) cluster control method based on deep reinforcement learning. The method comprises the following steps: S1) offline training phase: establishing a stochastic UAV dynamical model, and after Q function evaluation based on a dueling double deep Q-network, performing action selection, wherein the dueling double deep Q-network is a D3QNnetwork; and S2) online execution phase: constructing the dueling double deep Q-network, and loading a trained network model, the network model and an action selection strategy operating on an onboardcomputer of a wing plane, a lead plane roll action being given by an operator, and autopilots of the lead plane and the wing plane realizing closed-loop control according to rolling actions respectively and repeating the steps until the flight mission is completed. The method has the advantages of high real-time performance and adaptability, and being capable of transferring the strategy obtainedby training in simulation to the real environment and the like.