Disclosed is a multivariable predictive control method for an air separation device. The multivariate predictive control method comprises the steps that (a) an upper computer is additionally arranged on a DCS system local area network of the air separation device for creating an MVPC server of the air separation device; (b) variables are determined, specifically, according to the technological process optimizing target of the air separation device, the controlled variable (CV), the manipulating variable (MV) and the disturbance variable (DV) are determined; (c) the establishment principles of the size of an MVPC matrix are that key parameters are selected, the variables are controlled as less as possible, the valve opening degree, namely corresponding to the output of a PID controller under conventional control, is adjusted through the MV of an MVPC, when the MVPC works, the PID controller corresponding to the MV of the MVPC is in cascade control, the output of the PID controller tracks the MV of the MVPC, and when the MVPC does not work, the PID controller corresponding to the MV of the MVPC is in automatic control, and the MV of the MVPC tracks the output of the PID controller; (d) a model is built, specifically, an MVPC controller model is built in the air separation device, or independent MVPC controller models are built in different work sections in the air separation device; (e) the models are trained and adjusted, specifically, the gain K, the time constant Tau and the delay time Tdelay in the models are preliminarily set, and after the models are put into use, the optimized gain, time constant and delay time are provided by a system automatically; and (f) the models are put into use.