초록 ▼

A fast and reliable technique for tuning multivariable model predictive controllers (MPCs) that accounts for performance and robustness is provided. Specifically, the technique automatically yields tuning weights for the MPC based on performance and robustness requirements. The tuning weights are pa

A fast and reliable technique for tuning multivariable model predictive controllers (MPCs) that accounts for performance and robustness is provided. Specifically, the technique automatically yields tuning weights for the MPC based on performance and robustness requirements. The tuning weights are parameters of closed-loop transfer functions which are directly linked to performance and robustness requirements. Automatically searching the tuning parameters in their proper ranges assures that the controller is optimal and robust. This technique will deliver the traditional requirements of stability, performance and robustness, while at the same time enabling users to design their closed-loop behavior in terms of the physical domain. The method permits the user to favor one measurement over another, or to use one actuator more than another.

대표청구항 ▼

What is claimed is: 1. A method of automarically tuning a multivariable model predictive controller (MPC) for a multivariable process that comprises the steps of: (a) identifying a process model for the multivariable process; (b) scaling inputs, which are manipulated variables, and outputs, which a

What is claimed is: 1. A method of automarically tuning a multivariable model predictive controller (MPC) for a multivariable process that comprises the steps of: (a) identifying a process model for the multivariable process; (b) scaling inputs, which are manipulated variables, and outputs, which are controlled variables, of the process model; (c) calculating closed-loop transfer functions for performance and robustness as functions of tuning parameters for the MPC wherein the process model is used to calculate the closed-loop transfer functions or gain functions; (d) identifying at least one performance requirement for the MPC and transforming the at least one performance requirement into constraints on frequency responses of the closed-loop transfer functions or gain functions; (e) identifying at least one robustness requirement for the MPC and transforming the at least one robustness requirement into constraints on frequency responses of the closed-loop transfer functions or gain functions; (f) calculating possible ranges of tuning parameters such that a solution is known to exist within the range; (g) determining optimal tuning parameters with respect to performance and robustness; and (h) automatically yielding tuning parameters for the MPC based on performance and robustness requirements. 2. The method of claim 1 wherein step (g) comprises generating tuning parameters such that performance is optimized and at least one robust stability condition is satisfied. 3. The method of claim 1 wherein optimal actuator positions for the MPC are determined by solving an object function that employs weights in the form of positive semi-definite matrices and wherein in step (g) values for the weights. 4. The method of claim 3 wherein the MPC is approximated as a linear unconstrained controller, which is characterized by a constant prefilter matrix K, and complex transfer matrix K(z), and closed-loop transfer functions employed are calculated to satisfy the three closed-loop requirements which are characterized as disturbance rejection requirements, set point tracking requirements, and a closed-loop system that are given as (C1),(C2), and (C3) as follows: S(z)=(I+G(z)K(z)) -1, (C1) T(z)=(I+G(z)K(z)) -1G(z)K(z)Kr, (C2) R(z)=-K(z)(I+G(z) K(z))-1, (C3) wherein S (z) is the sensitivity function linking disturbances d and outputs y, T(z) is the closed-loop transfer function linking the targets ysp and the outputs y, R(z) is the control sensitivity function linking the disturbances d and the actuators u, and wherein the closed-loop transfer functions S(z), T(z), and R(z) are functions of the tuning parameters Γ, Λ, and Φ. 5. The method of claim 4 wherein in step (g) of determining optimal tuning parameters the values of the weights are determined by employing the following gain functions: wherein ∥Aith row∥2 denotes the vector 2-norm (Euclidean norm) of the ith row vector of matrix A, and characterized in that the target peaks (the maximum elements) of the gain functions s(ω), t(ω), and r(ω) are defined as Ms, Mt, and Mr, respectively, ω is a parameter representing the frequency, S(z), T(z), and R(z) with z=ejω are from (C1), (C2), and (C3), n and m designate numbers of controlled variables y and manipulated variables u respectively. 6. The method of claim 5 wherein step (g) of determining the optimal tuning parameters comprises determining the value of the weights by comparing the maximum peaks of the combined gain functions s(ω), t(ω), and r(ω) in (C4), (C5), and (C6) to their targets Ms, Mt, and Mr, respectively, subject to a robust stability condition. 7. The method of claim 6 wherein step (g) of determining the optimal tuning parameters comprises determining the value of the weights by comparing the maximum peaks of the closed-loop transfer functions S(z), T(z), and R(z) in (C1), (C2), and (C3) to their targets Ms, Mt, and Mr, respectively, subject to a robust stability condition. 8. The method of claim 1 wherein the MPC employs an object function of the form: subject to: ΛΔu(k)<b-Cu(k-1) (C8) y(k)=G(q)u(k)+d (k) (C9) where ysp is comprised of the reference signals, usp is comprised of desired actuator positions, Hp and Hc are prediction horizon and control horizon, respectively, ∥x∥Γ2 denotes the weighted Euclidean norm of x defined as ∥x∥Γ2=xΓΓx; Δu(k)=(1-q-1)u(k)=u(k)-u(k-1), where q is the shift operator; weights Γ, Λ, and Φ are positive semi-definite matrices that include the following diagonal matrices: Γ=diag(γ1,γ2, . . . ,γn), (C10) Λ=diag(λ1,λ2, . . . ,λm), (C11) Φ=diag(φ1,φ2, . . . ,φm), (C12). 9. The method of claim 8 wherein step (g) of determining the optimal tuning parameters employs the weights Γ, Λ, and Φ in determining the optimal tuning parameters while satisfying the following three closed-loop requirements: (i) disturbance rejection requirement, (ii) set point tracking requirement, and (iii) the closed-loop system given by (C7), (C8), and (C9) is stable in the face of model uncertainty. 10. The method of claim 9 wherein the MPC is approximated as a linear unconstrained controller, which is characterized by a constant prefilter matrix Kr and complex transfer matrix K(z), and closed-loop transfer functions employed are calculated to satisfy the previously defined three closed-loop requirements as follows: S(z)=(I+G(z)K(z)) -1, (C1) T(z)=(I+G(z)K(z)) -1G(z)K(z)Kr, (C2) R(z)=-K(z)(I+G(z) K(z))-1, (C3) wherein S(z) is the sensitivity function linking disturbances d and outputs y, T(z) is the closed-loop transfer function linking the targets ysp and the outputs y, R(z) is the control sensitivity function linking the disturbances d and the actuators u, and wherein the closed-loop transfer functions S(z), T(z), and R(z) are functions of the tuning parameters Γ, Λ, and Φ. 11. The method of claim 10 wherein the closed-loop system is robustly stable for all plants Gp(z) if it is nominally stable and satisfies the following condition: where ∥R∥∞ denotes the H-infinity norm which is equal to the maximum singular value of the control sensitivity function R(z) in (C3), β and Δ(z) are from the plant with additive model uncertainty Gp(z) defined by where σ denotes the maximum singular value, z=ejω and the symbol ω represents the dynamical frequency, the symbol '∀ω' means for all the frequencies. 12. The method of claim 11 wherein a maximum singular value of R(z) is employed for satisfying a robust stability condition. 13. The method of claim 12 wherein the closed-loop system is represented in the form of gain functions that as follows: wherein ∥Aith row∥2 denotes the vector 2-norm (Euclidean norm) of the ith row vector of matrix A, and characterized in that the target peaks (the maximum elements) of the gain functions s(ω), t(ω), and r(ω) are defined as Ms, Mt, and Mr, respectively. 14. The method of claim 13 wherein step (g) of determining the optimal tuning parameters comprises determining the weights Γ, Λ, and Φ in (C7) by comparing the maximum peaks of the combined gain functions s(ω), t(ω), and r(ω) in (C4), (C5), and (C6) to their targets Ms, Mt, and Mr, respectively, subject to the robust stability condition (C13). 15. The method of claim 8 wherein step (g) of determining the optimal tuning parameters uses common weights ρ for Λ and a for Φ to determine the tuning weights Λ and Φ, respectively, in (C7), where common weights ρ and α are defined as Φ=ρ·diag(φ1t, φ2t, . . . ,φmt), (C15) Λ=α·diag(λ1t,λ 2t, . . . ,λmt), (C16) where φjt and λjt with j=1, . . . ,m are relative weights for each manipulated variable uj. 16. The method of claim 8 wherein step (g) of determining the optimal tuning parameters determine the tuning weight Φ first and the tuning weight Λ second.