Abstract
Industrial refrigeration systems are significant energy consumers, often managed by Real-Time Optimization (RTO) strategies that rely on accurate equipment models. However, these models are frequently unavailable or outdated. This paper presents ARTEO (Adaptive Real-Time Exploration and Optimization), a framework that utilizes Gaussian Processes and a "safety-first" exploration strategy to learn and optimize system efficiency during live operation without risking equipment damage or power violations.
Introduction
In large-scale industrial facilities, such as food freezing plants, refrigeration accounts for a substantial portion of total energy consumption. To minimize costs and environmental impact, engineers employ Real-Time Optimization (RTO) to balance cooling demand with power efficiency. A major hurdle for RTO is the requirement for precise mathematical models of components like compressors. In practice, these models are often unknown, change over time due to wear, or are too complex to characterize manually.
While active exploration could allow a system to learn its own efficiency curves, doing so in a production environment is risky. Exceeding power grid limits or failing to maintain critical temperatures can lead to equipment damage or product loss. This research addresses the challenge of "safe exploration": enabling a system to autonomously learn and improve its efficiency on the fly while strictly adhering to safety constraints.
The ARTEO Framework
The authors propose the ARTEO algorithm, which treats optimization as a "safe" variant of the multi-armed bandit problem. The framework operates as a supervisory layer above local controllers.
Key components of the framework include:
- Gaussian Processes (GPs): Instead of static formulas, ARTEO uses GPs to model compressor power consumption. GPs provide both a prediction and a quantified measure of uncertainty (standard deviation).
- Adaptive Utility Function: The optimization objective balances estimated cost with uncertainty. A "curiosity" hyperparameter ($z$) allows operators to tune the system's drive to explore unknown operating points.
- Confidence-Bound Safety: Using Upper Confidence Bounds (UCB), the algorithm ensures that new settings are only attempted if there is high statistical confidence that power consumption will remain below safety thresholds.
- Steady-State Detection: To ensure data quality, a statistical detector prevents the system from learning from transient data before the compressors have stabilized.
Experimental Setup
The framework was validated using a high-fidelity simulation of an industrial food freezing plant based on a real-world case study from Norway. The simulated system consists of two subsystems with 11 compressors. The study focused on a subsystem of five screw compressors of varying sizes (Large, Medium, and Small). Performance characteristics were derived from established literature and integrated into a MATLAB/Simulink environment. The evaluation utilized a realistic cooling load profile representing a fishing season, characterized by significant fluctuations over a three-day period.
Key Results and Insights
The ARTEO framework demonstrated several critical capabilities:
- Learning During Operation: Unlike traditional methods requiring disruptive "step tests," ARTEO treats every control decision as a learning opportunity, refining its models during normal production.
- Near-Optimal Efficiency: The framework achieved performance levels closely approximating a benchmark optimizer with perfect prior knowledge of the system.
- Constraint Adherence: The system successfully tracked fluctuating cooling demands while maintaining total power consumption below a safety limit of 1580 kW (80% of maximum capacity).
- Model Convergence: As the simulation progressed, the GP-predicted performance lines converged toward the actual, hidden characteristics of the machinery.
Conclusion
Despite its success, the authors identified a "jump" problem: sudden, massive spikes in demand might force the system into unexplored regions where GP confidence bounds are less reliable, potentially leading to safety violations. To mitigate this, future iterations could include rate-of-change or distance-based constraints to prevent excessive leaps into unknown operating states. Additionally, future research will explore how the system handles "drifting" characteristics, such as gradual efficiency loss due to mechanical degradation over months of operation.
Summary updated