Published in: Modeling, Identification and Control (2005)
Abstract
This paper investigates the application of Sequential Quadratic Programming (SQP) optimization algorithms to Nonlinear Model Predictive Control (NMPC). It evaluates the trade-offs between feasible vs. infeasible path methods, sequential vs. simultaneous methods, and reduced vs. full-space methods. The authors introduce a new optimization algorithm, rFOPT, which maintains feasibility with respect to inequality constraints throughout the optimization process. The effectiveness of these strategies is assessed through a Continuous Stirred Tank Reactor (CSTR) case study, which also examines the impact of various discretization methods on optimization performance.
Introduction
Nonlinear Model Predictive Control (NMPC) is a powerful control strategy capable of handling complex system dynamics and operational constraints directly. However, the real-time implementation of NMPC is often limited by the computational burden of solving a nonlinear programming (NLP) problem at every sampling instance. This paper explores the "under the hood" mechanics of these optimization problems, providing a comparative analysis of different architectural choices and introducing a novel algorithm designed for enhanced reliability.
The Three Pillars of NMPC Optimization
When designing an NMPC optimizer, engineers face three primary architectural decisions:
- Sequential vs. Simultaneous: Sequential methods (single-shooting) solve the physics model at every optimization step. In contrast, simultaneous methods (multiple-shooting) treat the model equations as equality constraints to be solved alongside the objective function.
- Full vs. Reduced Space: Full-space methods optimize over every variable in the system. Reduced-space methods use the model equations to eliminate dependent variables, focusing the optimizer only on the actual degrees of freedom (the control inputs).
- Feasible vs. Infeasible Paths: This determines whether the optimizer must stay within the "legal" bounds of the constraints at every intermediate iteration, or if it can temporarily violate constraints on its way to the final optimal solution.
Methodology and the rFOPT Algorithm
The authors focus on Sequential Quadratic Programming (SQP), the standard workhorse for nonlinear optimization. They compare three primary strategies:
- sOPT (Sequential): A feasible path method for equality constraints that solves the model at each iteration.
- rOPT (Reduced Hessian): A simultaneous method working in a reduced variable space. It only ensures the model equations are fully solved at the final optimal point.
- rFOPT (Reduced-space Feasible Optimization): The paper’s primary contribution. While standard rOPT may violate inequality constraints during the search, rFOPT uses a "tilting" mechanism to modify the search direction. This ensures that every intermediate iteration remains feasible with respect to nonlinear inequality constraints—a vital feature for the safety and reliability of real-world industrial controllers.
These algorithms were tested against various discretization schemes, including Explicit/Implicit Euler, Lobatto IIIC, and Runge-Kutta 4 (RK4), to determine how the choice of numerical integration interacts with the optimization logic.
Case Study: The Isothermal CSTR
The algorithms were evaluated using a simulated isothermal Continuous Stirred Tank Reactor (CSTR), a classic benchmark for nonlinear control due to its multiple steady states. The controller was tasked with navigating a 50% step change in feed concentration, requiring it to manage a highly nonlinear landscape while respecting physical bounds and nonlinear inequality constraints.
Key Insights and Results
1. The "Implicit" Trap
The study reveals that sequential methods (sOPT) are highly sensitive to the choice of discretization. When using implicit methods—which are necessary for "stiff" systems—sOPT becomes computationally expensive because it must solve a complex system of equations at every single iteration.
2. Robustness of Simultaneous Methods
Simultaneous methods (rOPT and rFOPT) proved largely indifferent to the discretization scheme. By treating the model as a set of constraints, they avoid the overhead of solving the physics model repeatedly, making them more versatile for complex system dynamics.
3. Computational Efficiency
In tests using standard RK4 discretization, the reduced-space simultaneous method (rOPT) was 2 to 5 times faster than traditional full-space SQP solvers (such as those found in the NAG toolbox). However, the authors noted that while rFOPT provides the safety of feasibility, it carries a performance "tax" due to the additional sub-problems required for the tilting mechanism.
4. The Importance of Derivatives
The paper strongly advocates for the use of analytic derivatives or Automatic Differentiation (AD). The results showed that relying on numerical finite differences can increase CPU time by an order of magnitude, acting as a significant bottleneck for real-time applications.
Conclusions and Future Work
The research demonstrates that simultaneous reduced-space methods offer significant advantages in speed and robustness for NMPC, particularly when paired with sophisticated discretization schemes. The introduction of rFOPT provides a path forward for applications where maintaining constraint feasibility during the optimization process is non-negotiable.
Future research directions suggested by the authors include scaling these methods to higher-dimensional systems, developing more robust infeasibility recovery strategies for when disturbances push the system outside of legal bounds, and transitioning the algorithms from Matlab to compiled environments (C/C++) for true real-time industrial testing.
Summary updated