District heating has a long history, all the way back to ancient Roman times times, with a modern resurgence in the EU and Canada. The definite upside of district heating is centralized power generation, which simplifies efforts to use the cleanest and the most efficient heat generation solutions available. The oldest source leveraged is geothermal energy. Some lucky folks can just dig a hole and, presto, one has plenty of warm water. The latest addition to power sources is the excess heat from data centers, already used in Finland to heat 500 homes.

To transfer the energy from the source to consumers, pressurized water is put into underground pipes. To consume the heat, one attaches an exchanger to the main pipe to transfer the energy into the internal circulation of the building. There are also other heat transfer systems, for example, air conditioning and factory steam networks, that operate in a similar fashion. The idea is very simple indeed.

However, minimising the energy spent and the cost of energy is less trivial. There are just a couple of controls, essentially input temperature and network flow, that affect thousands of customers in the network with propagation delays of multiple hours. While the heated water travels in the pipes, it always dissipates some energy to the environment, the rate of dissipation depending on outside temperature and insulation.

The traditional solution to the control problem has been to add large safety margins and tie the network input temperature to the outside temperature in a simple linear fashion. What if a neural network could learn the exact effects of controls for each customer? What if it were possible to find in real time the minimum energy use to keep all customers warm enough?

Controlling a system, such as a district heating network, is a natural task for our Curious Engine. The figure below demonstrates the concept, leveraging a simple linear model and automated optimisation for a period of 80 hours. The top plot shows the feed temperature to the network. The bottom plot shows the temperature at one of the buildings in the network. The original feed temperature and temperature at the building heat exchanger are shown in solid black. The dashed lines in the bottom diagram shows the predictions of the model – black for the original control, blue for optimised controls. The yellow line shows the outside temperature, translated to fit the graph, just to show that the original control is derived fairly directly from the outside temperature. The solid blue line in the top plot is the optimised feed temperature. The difference between the original and the automatically optimised controls is multiple degrees, which translates to substantial energy savings.

This leaves the blue solid line in the bottom picture. That’s what you’d get if the difference of predictions is applied to the original result, which is another way to estimate the results of suggested optimised controls. So, the system ‘imagines’ that the suggested sinusoidal wave control will keep the temperature flat at desired level.

The real temperatures are likely to deviate a bit. As this model represents everything as sums of sinusoidal basis functions, it has strict limitations on expressive power. We’ll share detailed results comparison between this linear model and deep neural model and the suitable Curious Engine optimisation magic at a bit later date.

Check back or follow us in LinkedIn, so you won’t miss the exciting part II!

Write a Comment

Your email address will not be published. Required fields are marked *