Abstract: International audience
Introduction
Designing the Distributed Control System (DCS) for a power plant—essentially the "brain" that manages safety and efficiency—is a complex mathematical challenge. A DCS consists of a network of autonomous controllers that must be configured to handle specific operations while maintaining a clean, reliable physical hardware architecture. This design process involves two interlocking optimization problems: the functional level (assigning operations to Automation Units) and the physical level (connecting cables to electrical cabinets).
Because these levels are interdependent, they cannot be solved in isolation. However, attempting to solve them simultaneously using standard 0-1 Integer Programming leads to a "combinatorial explosion," where the number of possible configurations exceeds the capacity of even the most powerful commercial solvers.
Methodology
To address this computational bottleneck, this research proposes a "divide and conquer" strategy utilizing Generalized Benders Decomposition (GBD). The problem is partitioned into two distinct components:
- The Master Problem: Handles the functional assignment of operations to Automation Units (AUs), aiming to balance processing loads.
- The Subproblem: Manages the physical assignment of cables to cabinets, aiming to minimize cable separation.
In this iterative process, the Master Problem proposes a configuration, which the Subproblem then evaluates for physical feasibility. If the Subproblem identifies a conflict or sub-optimal layout, it generates a "no-good cut"—a mathematical constraint—that prevents the Master Problem from revisiting that specific configuration.
Additionally, the authors implemented Symmetry-Breaking (SB) constraints. Power plants often contain identical hardware components; without SB, solvers waste significant time evaluating redundant permutations of the same solution. By forcing the solver to ignore these identical layouts, the search space is pruned significantly.
Experimental Evaluation
The proposed GBD + SB method was tested against five generated instances representative of real-world DCS architectures. These instances ranged from Instance 1 (~2,200 variables) to Instance 5 (~9,500 variables). Performance was compared against Direct Resolution (DR) using the FICO® Xpress solver with a one-hour time limit.
Results and Discussion
The synergy between decomposition and symmetry-breaking proved essential for solving larger instances. As shown in the performance data, the GBD + SB approach consistently outperformed standard methods in both speed and solution quality.
Key highlights from the results include:
- Significant Speedup: For Instance 1, while the standard DR method timed out with a 34.2% error gap, the GBD + SB method reached an optimal (0% gap) solution in just 57 seconds.
- Improved Convergence: For Instance 3, GBD + SB achieved optimality in under four minutes (194s), whereas the standard solver maintained a 20% gap after a full hour.
- Scalability: On the most complex scenario (Instance 5), the standard method struggled with a 49.2% gap. The GBD + SB method narrowed this to 15.5% within the same timeframe, representing a substantial improvement for large-scale engineering applications.
Conclusion
While the largest instances still present a challenge for reaching absolute optimality within an hour, the combination of Generalized Benders Decomposition and Symmetry-Breaking constraints makes previously impractical DCS design problems manageable. This approach provides a robust framework for optimizing the complex infrastructure of modern power plants.
Summary updated