In this use case, there is a hot water grid, literally a set of pipes through which water flows. Throughout the plant, there are sources that release heat to the grid and consumers that take heat. In this way, the waste heat in the factory can be reused.

How does it work?

Different levels can be worked on in order to move, step by step, to a better but also a more advanced balancing.

Level 0
Let each consumer consume hot water according to their own need. This is the least ideal scenario, in some cases so much hot water has already been consumed by certain processes that there is nothing left for important processes that suddenly need urgent heat.

Level 1
Per consumer you set up a simple regulation where certain
consumers have priority over others. This can be, for example, based on the available amount of heat or temperature. Each consumer here looks only at itself and at the availability of the source.

Level 2
Allow the consumers to consider each other as well. Each consumer knows how much energy the other consumers are currently consuming.

Level 3
Predictive balancing also involves looking into the future. For example, if you know that an important process that is going to need a lot of heat is going to start up in 10 minutes, you can stop some of the smaller consumers now.

Level 4
Prescriptive balancing does not just look at what scenario will occur in the future. Alternative scenarios are also analyzed to see what the
optimal one is.

Level 5
For future initiatives, we look at promising techniques. One of them is using Reinforcement learning. This technique is already being
developed at this installation in collaboration with Howest. A machine learning model itself searches for the most optimal strategy to start balancing consumers.

Where is the difference made?

Load balancing involves working at different levels to safely make progress and depending on the needs of the customer. Balancing these loads makes better use of the energy resource.

This principle deals with the line of thought that is broadly applicable regardless of the application. Where there is a need to balance
consumers, especially when concurrency is involved, this principle will be able to prove its worth.

Advantages Smart Hot Water Grid

  • Work at different levels according to the maturity of the installation
  • Balance your loads on a resource in an optimal way
  • Guarantees availability of energy to high priority consumers
  • Helps you save on energy

 

 

Reinforcement Learning with Howest

It may be impossible to have a human find the ideal strategy. Therefore, we let an Agent search for this strategy by having the Agent take actions in a simulation environment.
An action is to control, not control or partially control all participants.
This results in a certain consumption of the hot water network.

All the heat that the consumers need extra is provided by a heat exchanger on steam. Since steam is much more expensive than reusing waste heat, our goal is to teach the agent to use as little steam as
possible. Good actions, with little steam consumption, are rewarded and bad actions are punished. In this way, the Agent learns more and more about the ideal strategy to balance the heat grid in all possible scenarios.

After the Agent is well and truly trained in the simulation environment. Then it is set up in parallel in production. That is, we let the Agent choose actions without executing them. If this also goes well, then the Agent is given the ability to execute actions in production as well.

At all times, the Agent is monitored. Deviations in performance are often due to changes in the environment. At that moment, or preferably in advance, the Agent is retrained to take the new situation into account.

 

Back to cases overview

How can we help you?

Contact us