The invention discloses an improved self-adaptive generalized predictive control method for a data-driven time delay system, which comprises the following steps of: firstly, determining a controlled variable and a controlled variable, acquiring input and output data of a time-delay system to construct a historical database, processing abnormal data, and adopting a time-varying forgetting factor recursive least square method to establish an autoregressive integral moving average model with controlled quantity of the time delay system; designing a reference trajectory according to the expected value, solving a Diophantine equation, designing a GPC control law, and constructing an optimization performance index; and calculating and implementing the optimal control quantity, collecting input and output data, and updating a historical database. According to the invention, the time lag problem of the system can be effectively solved, the utilization rate of offline and online input and output measurement data is improved, the influence of input and output interference and abnormal data on control is effectively inhibited, and the stability and good robustness of the system are ensured.