Safe Optimization of an Industrial Refrigeration Process Using an Adaptive and Explorative Framework

Buse Sibel Korkmaz, Marta Zagórowska, Mehmet Mercangöz

TL;DR

This paper introduces a framework called ARTEO (Adaptive Real-Time Exploration and Optimization) to improve the energy efficiency of industrial refrigeration systems. By using Gaussian Processes to model unknown machine characteristics and a "safety-first" exploration strategy, the system learns to optimize itself during normal operation without risking equipment damage or power overages.

What Problem Does This Paper Address?

In large-scale industrial settings, like food freezing plants, refrigeration is a massive energy sink. Engineers typically use Real-Time Optimization (RTO) to minimize power consumption while meeting cooling demands. However, RTO usually requires an accurate mathematical model of the equipment (like compressors).

The reality is that these models are often unknown, outdated due to wear and tear, or too complex to map manually. While "exploring" different settings could help a system learn its own efficiency curves, doing so blindly in a factory is dangerous—you might violate safety constraints, such as exceeding the power grid's capacity or failing to keep food at the required temperature. The authors tackle the challenge of safe exploration: how can a system learn to be more efficient on the fly without ever crossing a "red line"?

Methods

The authors implement the ARTEO algorithm, which treats the optimization problem as a "safe" version of the multi-armed bandit problem. Key components include:

  • Gaussian Processes (GPs): Instead of assuming a fixed formula for compressor performance, the system uses GPs to model power consumption. GPs provide not just a prediction, but a measure of uncertainty (standard deviation).
  • Adaptive Utility Function: The optimization goal is a weighted sum of the estimated cost and the uncertainty. A hyperparameter ($z$) acts as a "curiosity dial." Increasing $z$ encourages the system to explore unknown operating points to reduce future uncertainty.
  • Confidence-Bound Safety: To ensure safety, the algorithm uses Upper Confidence Bounds (UCB). It only tries new settings if it is statistically confident (within a high probability) that the resulting power consumption will remain below the safety threshold.
  • Steady-State Detection: Since industrial processes have lag, the system includes a statistical detector to ensure it only learns from data once the compressors have stabilized after a change.

Datasets

The study uses a high-fidelity simulation of an industrial food freezing plant based on a real-world case study from Norway.

  • System Configuration: Two subsystems with 11 compressors (3 reciprocating, 8 screw). The paper focuses on a subsystem of 5 screw compressors of varying sizes.
  • Data Sources: Performance characteristics for the compressors were sourced from established literature (Manske et al., 2000) and embedded into a MATLAB/Simulink environment.
  • Demand Profiles: The evaluation used a realistic cooling load profile representing a fishing season, featuring significant fluctuations over a three-day period (Saturday to Monday).

Key Insights and Ideas

  • Learning While Doing: Unlike traditional methods that require a separate "training phase" or disruptive "step tests," ARTEO learns during normal operation. It treats every control decision as a data point.
  • The Exploration/Exploitation Trade-off: The paper demonstrates that by adjusting the $z$ parameter, operators can choose between immediate energy savings (exploitation) and faster learning (exploration).
  • The "Safe Seed" Necessity: The algorithm requires a small "safe seed set"—initial operating points known to be safe—to start the learning process. You can't start from zero knowledge if you want guaranteed safety.
  • Uncertainty as a Resource: By quantifying what it doesn't know, the system can strategically pick operating points that provide the most information, eventually "collapsing" the uncertainty blue-zones on its performance graphs.

Key Results

  • Near-Optimal Performance: The ARTEO framework closely approximated the performance of a benchmark optimizer that had perfect prior knowledge of the system's characteristics.
  • Successful Constraint Management: In the primary test case, the system successfully tracked fluctuating cooling demands while keeping total power consumption below the 1580 kW safety limit (80% of maximum capacity).
  • Convergence: As shown in the paper's figures, the GP-predicted performance lines converged almost perfectly to the actual (hidden) performance curves by the end of the simulation.

Limitations and Future Work

  • The "Jump" Problem: The authors identified a critical limitation: if the system receives a sudden, massive spike in demand that forces it into a completely unexplored region, the safety constraints might be violated. This happens because the GP's confidence bounds are only reliable near previously explored data.
  • Rate-of-Change Constraints: To fix the "jump" problem, the authors suggest adding distance-based constraints (similar to move-suppression in traditional MPC) to prevent the system from leaping too far into the unknown in a single step.
  • Dynamic Environments: While the system handles changing loads, future work could investigate how it handles "drifting" characteristics, such as a compressor becoming less efficient over months due to mechanical degradation.

Summary updated

Abstract

Real-time optimization can play a key role for improving the performance of industrial processes, but it becomes challenging in cases where the process characteristics are uncertain, particularly when the uncertain characteristics impact safety-related constraints. In this study, we present an adaptive and explorative real-time optimization framework that can effectively learn the characteristics of an industrial refrigeration plant through externally driven changes in cooling load targets and through self-exploration. We utilize Gaussian processes (GP) to facilitate learning unknown compressor characteristics of the plant and we leverage the uncertainty quantification of the GP to drive exploration using a weighted sum term in the objective function for real-time optimization. Furthermore, the uncertainty information is used to probabilistically enforce the maximum total power consumption constraint for the compressors with high confidence at all times. Our simulated experiments demonstrate that the proposed approach safely enhances the energy efficiency of the refrigeration process, closely approximating the performance of a best-case solution that has complete information about the plant performance characteristics. We also demonstrate the impact of varying the exploration term in the solution and how the uncertainty of plant behaviour is reduced even in the absence of cooling load target changes.
BibTeX
@inproceedings{korkmaz2023safe,
  author = {Buse Sibel Korkmaz and Marta Zagórowska and Mehmet Mercangöz},
  title = {{Safe Optimization of an Industrial Refrigeration Process Using an Adaptive and Explorative Framework}},
  booktitle = {IFAC-PapersOnLine},
  year = {2023},
  doi = {10.1016/j.ifacol.2023.10.1811},
}

Loading related content...

Figure