Using model predictive control to optimize variable trajectories and system control

Abstract

A method and system of predictive model control of a controlled system with one or more physical components using a model predictive control (MPC) model, determining an iterative, finite horizon optimization of a system model of the controlled system, in order to generate a manipulated value trajectory as part of a control process. At time t sampling a current state of the controlled system a cost function minimizing manipulated variables trajectories is computed with the MPC model for a relatively short time horizon in the future, wherein the MPC uses a quadratic programming (QP) algorithm to find the optimal solution, and wherein the QP algorithm is solved using an Active Sets solver (AS) class algorithm with simple constraints based on gradient projection and using Newton step projection. A move of the manipulated value trajectory is implemented and the control process is moved forward by continuing to shift the prediction horizon forward.

Description

TECHNICAL FIELD

The present technology relates generally to automated process control, and more particularly to model predictive control.

BACKGROUND

Many advanced control techniques are formulated as optimization problems, typically as mathematical programming. One very successful advanced control technique is Model based Predictive Control (MPC). The MPC method is very well known in the process industry and it has been proven by many practical applications. There are MPC formulations for both linear and nonlinear systems. Nonlinear MPC requires solution of nonlinear mathematical programs in real-time which can be a challenging problem, for example due to a limitation of the computing resources, the complexity of the problem to solve, or the time available to solve it. Therefore, most of the practical applications are based on a linearity assumption or approximation. The linear MPC solutions are usually formulated as Quadratic Programming.

In the process industry, it is possible to use relatively powerful standard computers. This is not possible in many embedded applications where the embedded platform has relatively limited computational and storage resources. Therefore, more “modern” control techniques based on real-time optimization may require more computational resources than are available for such applications. Furthermore the sampling periods of embedded digital control systems can be much faster. For example, in an industrial process the sampling periods are typically at least in order of seconds while in embedded applications they often start at about 10 milliseconds. It is therefore a challenge to implement the advanced control methods that require real-time optimization for such small sampling periods.

BRIEF DESCRIPTION OF THE DRAWINGS

FIGS. 1A and 1B illustrate a process and software 10A and 10B incorporating an MPC control technology according to the present technology.

FIG. 2 illustrates a controller 110 incorporating an MPC control technology according to the present technology.

FIG. 3 illustrates an example computing system 200 for use in a controller according to the present technology.

FIG. 4 illustrates an example application 400 of the MPC control technology described herein applied to control of automotive components.

FIGS. 5A-5D and FIGS. 6A-6F illustrate an example application of the present technology to control of a diesel engine.

DETAILED DESCRIPTION

The following description should be read with reference to the drawings, in which like elements in different drawings are numbered in like fashion. The drawings, which are not necessarily to scale, depict selected embodiments and are not intended to limit the scope of the invention. Although examples of systems and methods are illustrated in the various views, those skilled in the art will recognize that many of the examples provided have suitable alternatives that can be utilized. Moreover, while the various views are described specifically with respect to several illustrative control systems, it should be understood that the controllers and methods described herein could be applied to the control of other types of systems, if desired.

The technology described herein utilizes an MPC-based control strategy that uses an iterative, finite horizon optimization of a system model. At time t the current system state is sampled and a cost function minimizing manipulated variables trajectories is computed, for example using a numerical minimization algorithm, for a time horizon in the future: [t, t+T]. Specifically, an online or on-the-fly calculation is used to explore state trajectories that emanate from the current state and find a cost minimizing manipulated variable trajectories until time t+T. Such a strategy may be determined through a solution of quadratic program (QP). A step of the manipulated variables trajectories is implemented, then the system state is sampled again and the calculations are repeated starting from the now current state, yielding a new control and new predicted state path. The prediction horizon keeps being shifted forward and for this reason MPC is also called receding horizon control.

According to one example embodiment, the inventive subject matter hereof provides an Active Sets solver (AS) class algorithm for solving the QP problem with simple constraints only based on gradient projection and using Newton step projection. In an AS class algorithm, simple (or box) constraints for optimization variables are defined as the lower and upper bounds, e.g. “LB<=X<=UB” where X is the optimization variable, and LB and UB are the lower and upper bounds, respectively. The gradient projection based method enables adding or removing of multiple constraints from the working set in each iteration of the algorithm, to enable quicker identification of the set of active constraints in the optimum solution. In addition, the Newton step projection speeds up the convergence of the algorithm. According to one embodiment, the method and corresponding system is used to solve QPs with simple constraints arising from MPC control.

Referring now to FIGS. 1A, 1B and 2, there is illustrated a flow chart and system diagram of a first embodiment of a system for model predictive control according to the invention. As will be described in more detail below, the improved process 10A and corresponding controller 110 provide a model predictive control approach that can be implemented with less computational resources and/or with greater speed than other approaches. In one embodiment, as illustrated below, the system is used in an embedded application with limited computational resources. Accordingly, the process 10A and controller 110 provide for using an MPC model to determine an iterative, finite horizon optimization of a system model, in order to generate a manipulated variables trajectories.

• • a. At time t, sampling the current system state (20) and computing a cost function minimizing manipulated variables trajectories with the MPC model, using the method described in FIG. 1B, for a relatively short time horizon in the future: [t, t+T]. (25)
  • b. Implementing a step or move of the manipulated variables trajectories. (30)
  • c. Repeating the control calculation starting from the now current state to yield a new control and new predicted state path.
  • d. Continuing the control process by continuing to shift the prediction horizon forward.

Referring now to FIG. 1B, there is illustrated the process 10B used to determine a cost function minimizing manipulated variables trajectories and set of active constraints based on the current system state:

• • a. Check if maximum of iteration of the algorithm was reached (45). If yes, then solution was not found (46) and the process 10B returns the suboptimal manipulated variables trajectories as the result. Else continue with new iteration.
  • b. Solve the set of equalities of Karush-Kuhn-Tucker (KKT) system with current active constraints set to get the Newton step. (50)
  • c. Move along the Newton step up to the nearest boundary constraint and consider the Newton step as the rest of full step (55).
  • d. Check if Newton step length is zero (57). If yes, continue by checking the stopping conditions (63). Else the new set of active constraints can be found using the projection of Newton step onto subspace of constraints. (60)
  • e. Check if all Lagrange multipliers associated with active constraints are negative and the coordinates of gradient corresponding to non-active constraints are zeros (63). If yes, the optimum was found (64), else look for better solution by using projection of the negative gradient direction onto subspace of constraints (65).
  • f. Project negative gradient onto subspace of constraints to get new active set of constraints and improve the objective (65).

According to one alternate embodiment, a warm-start is used to reduce the number of iterations and improve convergence. According to another embodiment, the simple constraints are divided into hard constraints for the manipulated variables (such as actuators) with the other constraints (such as system state, system output) formulated as soft constraints. In another embodiment, the optimization problem takes the form of:

$\underset{x}{\min }\frac{1}{2}{x}^T\mathrm{Gx}+x^{T}f,s.t.x_{\min }\le x\le x_{\max }$

Referring now to FIG. 2, there is illustrated a simplified block diagram of a controlled system 100 under the control of a controller 110. Controller 110 includes a computer system 90, an output module 112, and an input module 114. In operation, controller 110 generates sequence of manipulated variables (MVs) 120 from module 112 and receives sequence of controlled variables and disturbance variables (input variables) 130 from input module 114. Control outputs and inputs 120 and 130 may take the form of analog or digital control signals sent to and received from controlled system 100.

Referring now to FIG. 3, there is illustrated a block diagram of a computer system 200 that executes programming for implementing the above-described algorithms as described in particular in FIGS. 1A and 1B. According to one embodiment, the MPC control algorithms are implemented as software modules in computer programs 225 on the system 200. A general computing device in the form of a computer 210 may include a processing unit 202, memory 204, removable storage 212, and non-removable storage 214. Memory 204 may include volatile memory 206 and non-volatile memory 208. Computer 210 may include—or have access to a computing environment that includes—a variety of computer-readable media, such as volatile memory 206 and non-volatile memory 208, removable storage 212 and non-removable storage 214. Computer storage includes random access memory (RAM), read only memory (ROM), erasable programmable read-only memory (EPROM) & electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD ROM), Digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other tangible and physical medium capable of storing computer-readable instructions. Computer 210 may include or have access to a computing environment that includes input 216, output 218, and a communication connection 220. The computer may operate in a networked environment using a communication connection to connect to one or more remote computers. The remote computer may include a personal computer (PC), server, router, network PC, a peer device or other common network node, or the like. The communication connection may include a Local Area Network (LAN), a Wide Area Network (WAN) or other networks. Computer-readable instructions stored on a tangible and physical computer-readable medium in a non-transitory form are executable by the processing unit 202 of the computer 210. A hard drive, CD-ROM, and RAM are some examples of articles including a computer-readable medium. According to one example embodiment, computer programs 225 are stored in memory 104, in storage devices 212 or 214, in other storage locations on the system or outside the system such as a network storage device, in any combination.

Referring now to FIG. 4, there is illustrated an application 400 of the MPC system of FIG. 3 to an internal combustion engine, and in particular to an engine that is designed to operate so as to reduce unwanted emissions. In this application, controller 110 is interfaced with one or more automotive systems or components 410. Systems or components 410 can include, for example but not by way of limitation, turbo-chargers, exhaust gas recirculation, valves, throttles, carburetors or fuel injectors, ignition systems, catalytic converters, braking controls or transmissions. According to other example embodiments, the MPC system of FIG. 3 may be applied to aerospace systems, such that systems or components 410 can be aerospace systems or components, as opposed to automotive components. In addition, the MPC system of FIG. 4 is suitable for any other embedded applications, for example to replace existing standard control methods such as networks of PID (proportional-integral-derivative) controllers, or for any other applications.

According to one example implementation of the present technology, it is applied to a simple typical diesel engine control problem, wherein the objective is to follow a prescribed set-point for fresh air flow (system output 1) while keeping the boost pressure (system output 2) within specified limits. The actuators are EGR (Exhaust Gas Recirculation), (system input 1) and waste gate valve (system input 2). Positions of both valves are constrained. This MPC control problem can be transformed to a QP problem with simple box constraints. The engine is linearized for a selected operating point. The resulting linear model is shown in FIGS. 5A, 5B, 5C, and 5D. In this example the resulting QP optimization problem for the real-time has 18 optimization variables. Therefore, there are 18 box constraints (total number of constraints is 2*18=36). To solve this problem the above-described process was used. Note that this QP problem needs to be solved at each sampling period.

The results are shown in FIGS. 6A, 6B, 6C, 6D, 6E, and 6F. There are two cases. In the first case, the optimal solution to the optimization problem is illustrated in solid black line. In the second case, the maximum number of iterations of the solver was limited to three. In the example, the warm start technique (the solver is started from the previous solution) was used, which reduced the number of iterations to find the optimal solution. The maximum number of iterations to find the optimal solution is nine. The effect on control performances is negligible when compared with the optimal solution.

According to one example embodiment but without limitation, the base MPC model used herein is a multivariable control algorithm that uses an internal dynamic model of the process, a history of past control moves, and an optimization cost function J over the receding prediction horizon to calculate the optimum control moves. In one example implementation, the process to be controlled can be described by a time-invariant nth-order multiple-input multiple-output (MIMO) ARX (Autoregressive Model with External Input) model:

$y\left(k\right)=-\sum _{i=1}^{n}A\left(i\right)y\left(k-i\right)+\sum _{i=0}^{n}B\left(i\right)u\left(k-i\right)+\sum _{i=1}^{n}C\left(i\right)v\left(k-i\right)+e\left(k\right)$ where u(k) is a vector of n_u inputs or manipulated variables (MVs), v(k) is a vector of n_v, disturbance variables (DVs), y(k) is a vector of n_y outputs or controlled variables (CVs), e(k) is a white noise sequence of measurement noise (an n_y vector) with n_y×n_y covariance matrix Σ, and A(i), B(i) and C(i) are coefficient matrices (of appropriate dimensions n_y×n_y, n_y×n_u and n_y×n_u). Note that the latest data that are available for the prediction of the output y(k) are the disturbance v(k−1) and the values of manipulated variable u(k). With Kalman filter enabled, also the state space model or OE (Output Error) model can be used as an alternative.

According to one example embodiment, example AS solver algorithms suitable for implementation of the inventive subject matter are described in more detail in J. Nocedal, S. J. Wright, Numerical Optimization, Springer, 2006 or R. Fletcher, Practical Methods of Optimization, Wiley, 2003.

As described above the MPC technology described herein is particularly but not exclusively suitable for QP problems arising from the MPC control of embedded applications. For example, one application of the proposed method and system is an MPC controller for solving the multivariable control problems for internal combustion engines. The method can be also used as a core QP solver in Sequential Quadratic Programming (SQP). The SQP is a practical method for solving nonlinear optimization problems, for example in a nonlinear MPC.

Thus, the methods and systems described above reduce the number of iterations of the QP algorithm required to calculate a useful and effective control path. Accordingly, the improved MPC control strategy can be deployed using less computational resources. The reduced computation time afforded by the improved approach is particularly useful for applications with limited resources and/or short sampling period.

The Abstract is provided to comply with 27 C.F.R. §1.72(b) is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims.

Claims

1. A method of model predictive control of a controlled system with one or more physical components, comprising: at time t, sampling a current state of the controlled system and computing a cost function minimizing manipulated variables trajectories with a model predictive control (MPC) for a time horizon in the future, wherein the MPC uses a quadratic programming (QP) algorithm to find an optimal solution, and wherein the QP algorithm is solved using an Active Sets solver (AS) class algorithm with constraints based on gradient projection and using Newton step projection, wherein the Newton step projection facilitates adding of multiple constraints to a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in the optimal solution;implementing a move of the manipulated variables trajectories including outputting a control signal for directly or indirectly controlling the controlled system; andshifting the prediction horizon forward in time and repeating the sampling, computing and implementing steps.

2. A method according to claim 1, wherein the active set class algorithm uses constraints for optimization variables, wherein the constraints are defined as the lower and upper bounds.

3. A method according to claim 1, further wherein the gradient projection facilitates adding or removing of multiple constraints from a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in the optimal solution.

4. A controller used to control a controlled system with one or more physical components, comprising: a computer system including one or more computer programs operative on the computer system to: at time t, sample a current state of the controlled system and compute a cost function minimizing manipulated variables trajectories with a model predictive control (MPC) model for a time horizon in the future, wherein the MPC uses a quadratic programming (QP) algorithm to find an improved solution, and wherein the QP algorithm is solved using an Active Sets solver (AS) class algorithm with constraints based on gradient projection and using Newton step projection, wherein the Newton step projection facilitates adding of multiple constraints to a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in the improved solution; andimplement a move of the manipulated variables trajectories by outputting a control signal used directly or indirectly to control the controlled system.

5. A controller according to claim 4, wherein the active set class algorithm uses constraints for optimization variables wherein the constraints are defined as the lower and upper bounds.

6. A controller according to claim 4, further wherein the gradient projection facilitates adding or removing of multiple constraints from a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in the improved solution.

7. A controller used to control an automotive system with one or more physical components, wherein the automotive system comprises one or more automotive components requiring control, the controller comprising: a computer system including one or more computer programs operative on the computer system to: at time t, sampling a current state of the one or more automotive components and computing a cost function minimizing manipulated variables trajectories with a model predictive control (MPC) for a time horizon in the future, wherein the MPC uses a quadratic programming (QP) algorithm to find an improved solution, and wherein the QP algorithm is solved using an Active Sets solver (AS) class algorithm with constraints based on gradient projection and using Newton step projection, wherein the Newton step projection facilitates adding of multiple constraints to a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in the improved solution using a specified maximum number of iterations; andimplementing a move of the manipulated variables trajectories to facilitate control the one or more automotive components.

8. A controller according to claim 7, wherein the AS class algorithm uses constraints for optimization variables wherein the constraints are defined as the lower and upper bounds.

9. A controller according to claim 7, further wherein the gradient projection facilitates adding or removing of multiple constraints from a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in the improved solution.

10. A controller according to claim 7, further including moving the control horizon forward and repeating the determination of manipulated variable trajectories.

11. A controller for an automotive system, comprising: processing unit;an input for providing one or more controlled variables (CV) and/or disturbance variables (DV) to the processing unit from the automotive system;an output for providing one or more manipulated variables (MV) from the processing unit to the automotive system, wherein each of the one or more manipulated variables (MV) have a trajectory;the processing unit configured to sample one or more of the one or more controlled variables (CV) and/or disturbance variables (DV) at a time “t”, and to compute a cost function minimizing the trajectory of one or more of the manipulated variables (MV) using a model predictive control (MPC) over a time horizon, wherein the MPC uses a quadratic programming (QP) algorithm that has constraints and that uses Newton step projection, wherein the Newton step projection facilitates adding of multiple constraints to a working set in each iteration of the algorithm, to facilitate identification of a set of active constraints in an improved solution using a specified maximum number of iterations; andimplementing a move of the trajectory of one or more of the manipulated variables (MV), and outputting one or more of the manipulated variables (MV) to the automotive system.

12. A controller according to claim 11, wherein the quadratic programming (QP) algorithm is solved using an Active Sets solver (AS) class algorithm with constraints based on gradient projection and using Newton step projection.

13. A controller according to claim 11, wherein the processing unit shifts the time horizon forward, and then samples one or more of the one or more controlled variables (CV) and/or disturbance variables (DV) a second time.

14. A controller according to claim 11, wherein after the processing unit samples one or more of the one or more controlled variables (CV) and/or disturbance variables (DV) for the second time, the processing unit again compute a cost function minimizing the trajectory of one or more of the manipulated variables (MV) using a model predictive control (MPC) over a time horizon, and implements a move of the trajectory of one or more of the manipulated variables (MV), and further outputs one or more of the manipulated variables (MV) to the automotive system.

15. A controller according to claim 11, wherein the automotive system includes one or more actuators, and wherein the one or more manipulated variables (MV) are used to control one or more of the actuators.

16. A controller according to claim 15, wherein one or more of the actuators include an EGR valve.

17. A controller according to claim 15, wherein one or more of the actuators include an Waste Gate Valve.