Stable Decomposition for Dynamic Optimization

Purt Tanartkit, Lorenz T. BieglerIndustrial 1995 · DOI

TL;DR

This paper introduces a novel decomposition algorithm for solving dynamic optimization problems, particularly those involving differential-algebraic equations (DAEs) and unstable components. By combining a stable multiple shooting approach with a reduced Hessian Successive Quadratic Programming (SQP) method, the authors achieve efficient and robust solutions for complex chemical engineering problems that often challenge existing methods.

What Problem Does This Paper Address?

Dynamic optimization problems are crucial for tasks like process control, scheduling, and design in chemical engineering. These problems are often formulated as differential-algebraic equations (DAEs). Traditional methods for solving these DAE optimization problems fall into two main categories: sequential and simultaneous strategies. While effective for stable systems, these approaches can struggle with:

  • Unstable components: Many real-world chemical processes have inherent instabilities, leading to numerical difficulties and failure of standard algorithms.
  • Accumulating errors: Initial value problem (IVP) formulations, common in sequential methods, can suffer from error accumulation during integration, especially with unstable systems.
  • High dimensionality and nonlinearity: Simultaneous approaches, while better for handling constraints, can lead to very large and complex nonlinear programming (NLP) problems.
  • Mesh selection: Determining the optimal placement of finite elements (discretization points) is critical for accuracy and efficiency but is often treated separately or heuristically.

This paper aims to provide a stable, efficient, and robust method that can handle these challenges, particularly for problems with unstable dynamics and complex constraints.

Methods

The core of the proposed method is a stable decomposition algorithm for dynamic optimization problems. Key components include:

  • Multiple Shooting (MS) with Collocation: Instead of single shooting (which is prone to instability), the time horizon is divided into smaller subintervals (elements). Within each element, a collocation method is used to discretize the DAEs. This approach, implemented using the COLDAE routine, provides a stable framework for integrating the system equations.
  • Reduced Hessian Successive Quadratic Programming (SQP): To handle the large-scale NLP problem arising from discretization, a reduced Hessian SQP algorithm is employed. This method efficiently finds a search direction by partitioning variables into "decision" (control) and "dependent" (state) variables.
  • Bilevel Programming Framework: To address the critical issue of element placement, the authors propose a bilevel optimization framework. The outer problem optimizes the element lengths, while the inner problem solves the fixed-mesh NLP. This decoupling allows for a more systematic approach to mesh refinement and error control.
  • Multiplier-free Line Search: A specialized line search procedure is used that does not require the explicit calculation of Lagrange multipliers, simplifying the algorithm.
  • QPKWIK QP Solver: A tailored quadratic programming solver, QPKWIK, is used to efficiently handle the highly-constrained QP subproblems that arise in DAE optimization.

Datasets

The paper evaluates the proposed method on several benchmark problems from the literature, including:

  • PS polymerization problem
  • Van der Pol oscillator problem (with and without path constraints)
  • Parameter estimation problem
  • Zeolite distribution problem
  • A "car" problem (for demonstrating the bilevel framework)

These examples are used to demonstrate the algorithm's performance, stability, and efficiency compared to existing methods.

Key Insights and Ideas

  • Stability is paramount: The paper strongly emphasizes that for many chemical engineering problems, especially those with unstable dynamics, the choice of formulation (e.g., BVP vs. IVP) and solution method (e.g., MS vs. SS) is critical for stability and success.
  • Decomposition is key for scale: Breaking down the large NLP problem into smaller, manageable parts (via multiple shooting and SQP) is essential for tackling complex, large-scale dynamic optimization.
  • Element placement matters: The choice of finite element discretization significantly impacts the accuracy, efficiency, and even the ability to converge. A bilevel approach that optimizes element placement alongside control variables offers a promising path forward.
  • Infeasible path SQP excels for nonlinearity: For larger and more nonlinear problems, the infeasible path SQP approach generally outperforms the feasible path counterpart.
  • BVP formulation often preferred for stability: For problems that can be formulated as boundary value problems (BVPs), this often leads to better conditioning and more stable solutions compared to initial value problem (IVP) formulations, even if the analytical solutions are identical.

Key Results

The proposed method demonstrated robust performance across various test cases:

  • Handles unstable problems: The algorithm successfully solved problems that failed with sequential algorithms based on IVP solvers.
  • Efficient for large-scale problems: The reduced Hessian SQP and efficient QP solver (QPKWIK) contribute to good performance on large problems.
  • Improved accuracy and stability: The multiple shooting approach, combined with collocation, provided stable and accurate solutions, especially for problems with ill-posed characteristics.
  • Bilevel framework shows promise: The bilevel approach for element placement yielded better results and required less computation time than uniform meshing in the "car" problem.
  • Competitive performance: Computational statistics (CPU times, iterations) presented in tables show competitive or superior performance compared to existing methods for the tested examples. For instance, the infeasible path approach was significantly faster for the Van der Pol oscillator problem.

Limitations and Future Work

  • Gradient evaluation for bilevel optimization: The current implementation of the bilevel framework uses a finite difference scheme for computing gradients in the outer problem, which can be computationally expensive. Future work will focus on developing more efficient sensitivity algorithms.
  • Automatic element selection: While the bilevel approach is promising, a fully automatic procedure for selecting finite elements is still needed. Future work will explore adaptive element addition strategies.
  • Large-scale problems: The paper acknowledges that dealing with very large-scale problems remains an ongoing challenge, particularly in parameter estimation where stable side conditions are crucial but not always generalizable.
  • Outer problem performance: The success of the bilevel framework critically depends on the performance of the outer problem, especially its gradient evaluation. This is a key area for future research.

Summary updated

Abstract

ADVERTISEMENT RETURN TO ISSUEPREVArticleNEXTStable Decomposition for Dynamic OptimizationPurt Tanartkit and Lorenz T. BieglerCite this: Ind. Eng. Chem. Res. 1995, 34, 4, 1253–1266Publication Date (Print):April 1, 1995Publication History Published online1 May 2002Published inissue 1 April 1995https://pubs.acs.org/doi/10.1021/ie00043a029https://doi.org/10.1021/ie00043a029research-articleACS PublicationsRequest reuse permissionsArticle Views243Altmetric-Citations50LEARN ABOUT THESE METRICSArticle Views are the COUNTER-compliant sum of full text article downloads since November 2008 (both PDF and HTML) across all institutions and individuals. These metrics are regularly updated to reflect usage leading up to the last few days.Citations are the number of other articles citing this article, calculated by Crossref and updated daily. Find more information about Crossref citation counts.The Altmetric Attention Score is a quantitative measure of the attention that a research article has received online. Clicking on the donut icon will load a page at altmetric.com with additional details about the score and the social media presence for the given article. Find more information on the Altmetric Attention Score and how the score is calculated. Share Add toView InAdd Full Text with ReferenceAdd Description ExportRISCitationCitation and abstractCitation and referencesMore Options Share onFacebookTwitterWechatLinked InRedditEmail Other access optionsGet e-Alertsclose Get e-Alerts
BibTeX
@inproceedings{tanartkit1995stable,
  author = {Purt Tanartkit and Lorenz T. Biegler},
  title = {{Stable Decomposition for Dynamic Optimization}},
  booktitle = {Industrial},
  year = {1995},
  doi = {10.1021/ie00043a029},
}
Figure