Safe reinforcement learning for demand-response scheduling of industrial refrigeration

This project develops smart control algorithms to optimise industrial refrigeration systems, reducing energy costs while keeping food safely cold. Refrigeration often accounts for nearly half of a facility’s electricity use and £1b of UK electricity consumption annually, creating opportunities for savings. The project centres around algorithms that learn to schedule cooling operations during periods of higher electricity availability, while maintaining safe food temperatures. Currently, we have developed initial algorithms and tested them on simplified processes. In the coming months, we plan to test these algorithms on CrossnoKaye‘s advanced industrial refrigeration simulator. This represents a crucial step toward real-world deployment of energy-saving refrigeration technology.

• Reduce industrial refrigeration energy costs through intelligent scheduling
• Unlock demand-side flexibility for the electricity grid
• Maintain food safety standards throughout operation
• Create algorithms that work safely in real industrial settings
• Establish international collaboration for broader impact

The team uses reinforcement learning algorithms that learn refrigeration scheduling schemes through trial and experience, similarly to how humans learn skills. Unlike traditional approaches requiring detailed system models, these algorithms adapt to uncertainty in electricity prices and food storage demands. Key innovations include building upon existing control systems to accelerate learning and incorporating safety constraints to prevent food spoilage during the learning process. The algorithms leverage the natural thermal storage capacity of cold storage facilities, treating them like distributed batteries that can shift energy demand while maintaining required temperatures.

Initial testing on simplified computer simulations using artificial historical data has successfully demonstrated the algorithms can learn effective refrigeration scheduling strategies. Expected results from the upcoming CrossnoKaye simulator testing include validation of performance under realistic industrial conditions and complex electricity pricing schemes, leading toward eventual deployment on actual refrigeration systems.

The immediate next step involves deploying and testing the algorithms on CrossnoKaye’s simulator, which comprise digital twins of real industrial refrigeration systems. This will validate performance under complex, real-world electricity pricing schemes and operational constraints. The project aims to demonstrate continuous learning capabilities, where the algorithms adapt to changing conditions over time.

Potential impact areas include:

• Energy sector: Providing significant demand-side flexibility to electricity grids, helping integrate renewable energy sources
• Food industry: Reducing operational costs while improving food safety through optimized temperature control
• Environmental benefits: Decreasing carbon emissions through more efficient energy use and better grid integration