Integration of different models in the design of chemical processes: Application to the design of a power plant

José A. Caballero, Miguel A. Navarro, Rubén Ruiz‐Femenia, Ignacio E. GrossmannApplied Energy 2014 · DOI

TL;DR

This paper introduces a novel modeling framework that integrates diverse mathematical models, each with its unique numerical characteristics, for the synthesis and design of complex chemical processes. By leveraging Generalized Disjunctive Programming (GDP), the framework allows for the robust and efficient utilization of "best available" models, even those that are implicit or "black box," leading to more reliable process designs.

What Problem Does This Paper Address?

Designing complex chemical processes often requires using various mathematical models, each with its own strengths and numerical behaviors. While sophisticated tools exist for simulation, optimization, and synthesis, integrating these diverse models—ranging from explicit equation-based models to implicit "black box" simulators—presents a significant challenge. Traditional approaches often treat these models as rigid input-output structures, limiting their flexibility and efficiency in design and synthesis problems. This paper addresses the need for a unified framework that can effectively capture and utilize the specific numerical characteristics of each model, enabling a more robust and reliable synthesis of complex chemical processes.

Methods

The core of this work lies in the application of Generalized Disjunctive Programming (GDP) as a unifying framework. GDP allows for the direct formulation of problems involving discrete decisions and continuous variables, naturally accommodating the "if-then-else" logic inherent in selecting and integrating different models. The proposed framework involves:

  • Modular Approach: Each component of the chemical process is treated as a module, allowing for its specific numerical characteristics to be handled independently.
  • Model Encapsulation: Modules are "encapsulated" within disjunctions in the GDP formulation, isolating their specific procedures and enabling tailored numerical treatments.
  • Hybrid Simulation-Optimization: The framework combines explicit equation-based models with implicit models (like process simulators) and surrogate models (e.g., kriging) to leverage the "best available" model for each component.
  • Logic-Based Solvers: The GDP problems are solved using logic-based algorithms, avoiding the need for reformulation into Mixed-Integer Nonlinear Programming (MINLP) problems, which can sometimes lead to numerical issues.
  • Structural Analysis: An analysis of the input-output structure of each model is performed to select appropriate design variables and minimize recycle structures.
  • Derivative Calculation Strategies: Various methods for calculating derivatives are employed, including analytical, automatic differentiation, and finite differences, adapted to the characteristics of each module.

Datasets

Not discussed in the paper. The paper focuses on a case study of a utility system power plant, for which models are developed and simulated.

Key Insights and Ideas

  • GDP as a Unifying Framework: The paper convincingly demonstrates that GDP is a powerful and intuitive framework for integrating diverse models with varying numerical behaviors. It preserves the logical structure of the problem, making it easier to formulate and understand.
  • "Best Available Model" Utilization: The framework enables the use of the most appropriate model for each process unit, whether it's a rigorous simulator, a simplified correlation, or an experimental model, without requiring extensive reformulation.
  • Handling Model Heterogeneity: The ability to handle models with different numerical characteristics (e.g., noisy, computationally expensive) through tailored approaches like surrogate modeling (kriging) is crucial for complex system design.
  • Robustness through Modularization: Encapsulating modules within disjunctions allows for specific numerical treatments and derivative calculations, enhancing the overall robustness of the optimization process.

Key Results

The paper presents a case study of a utility system power plant, demonstrating the framework's capabilities through three instances with varying demands.

  • Efficient Synthesis: The framework successfully synthesized complex utility systems, showcasing the ability to integrate diverse models.
  • Computational Performance: For a fixed configuration, the optimization problems were solved efficiently. For instance, in the first example, the most complex non-linear problem was solved in around 60 seconds of CPU time, with NLP sub-problems typically taking 5-20 seconds. MILPs were solved in under 1 second.
  • Robustness: Numerical tests indicated that the optimal solutions were not overly sensitive to initial values, suggesting reasonable robustness.
  • Economic Optimization: The framework allows for the minimization of Total Annualized Cost (TAC), considering capital and utility costs. For example, the TAC for the first instance was $31.42 M$/year.

Limitations and Future Work

  • Convexity Assumption: Like most optimization algorithms, the MINLP and GDP algorithms used require convexity to guarantee convergence to a global optimum. Proving convexity for implicit models is difficult, and thus, a global optimum is not guaranteed.
  • Non-Convex Problems: For non-convex problems, the search is terminated heuristically after two consecutive iterations without improvement. The authors suggest performing a fixed number of iterations to potentially improve solution quality at the cost of increased computation time.
  • Future Work: The authors plan to concentrate on developing a new design and synthesis tool that can be applied to a wider range of processes.

Summary updated

BibTeX
@inproceedings{caballero2014integration,
  author = {José A. Caballero and Miguel A. Navarro and Rubén Ruiz‐Femenia and Ignacio E. Grossmann},
  title = {{Integration of different models in the design of chemical processes: Application to the design of a power plant}},
  booktitle = {Applied Energy},
  year = {2014},
  doi = {10.1016/j.apenergy.2014.03.018},
}
Figure