Abstract: In the existing model free adaptive control (MFAC) algorithms, the four model parameters λ, ρ, η, μ are kept constant during the control process, which leads to the problems of small influence of pseudo-partial derivatives on the control process and weak adaptive energy of the algorithm. Using the radial basis function (RBF) neural network, based on control input and pseudo partial derivatives, and taking the difference between the expected output and the real-time output as the training error, the four parameters can be adjusted in real time, which improves the existing compact format dynamic linearization MFAC method for discrete time nonlinear systems. Furthermore, a new BRF-MFAC algorithm was proposed, and its superiority of tracking performance was verified in the control of a nonlinear system. Compared with MFAC, the operation adjustment time of RBF-MFAC system for propylene concentration could reduce by 42.4% in the propylene separation column of a 0.3 Mt/a gas fractionation unit. The operational adjustment time of propylene product concentration and production in MIMO system could reduce by 78.0%.

Key words: process control, neural network, parameter estimation, gas fractionation unit, propylene separation column