Using physics and Machine Learning to optimize energy usage
With “EnTune BuildingAI” we optimize energy usage for large buildings, such as offices – to save up to 30% in energy. Most buildings do not have the intelligence to adjust to their environment. Hence they use unnecessary energy, which is costly. EnTune BuildingAI uses data and physical models to optimally predict energy usage, tune settings and with that save money and energy (see our previous blog).
In this blog, we’ll dive a bit deeper into how it works from a technical perspective.
Flexibility and adjustability with a digital twin
To accurately predict and manage a building’s energy usage, it is key to understand the thermal characteristics of the building. This physical-model-based insight, combined with data such as weather forecasts, energy prices, usage patterns, heating/cooling system dynamics, and personal preferences, allows us to create a digital twin of the building. We then apply a technical simulation model to this digital twin which results in an optimal energy plan at the lowest energy costs.
For the technical model, we defined two key requirements:
- It should be possible to adjust to the building’s environment; think of people in the building, the weather outside
- It should be quick and easy to install, and work for any type of building, new and old
So, how to optimize energy usage and costs, while taking the requirements above into account?
Gray box modeling: a blend of black and white
Typically, there are two ways to tackle the challenge: with Machine Learning (black box) or with physics modeling (white box).
Black box models
Machine Learning (ML) models are often used for complex problems where a lot of (historical) data is available. For this, you use data to model reality. ML models can become very accurate, but need to be trained and fed with data in order to reach their full potential.
White box models
Physical models are models based on physical laws, which are often used for less complex, more straightforward problems. For this, you model reality with physics rules. Physical models require little data and no training, which makes them transparent and easily interpretable.
What do we need from our model? On the one hand, we need the model to be quick, and work for any type of building – also buildings that have little to no historical data available. On the other hand, we’re facing high complexity, due to all environmental influences, such as the weather and energy prices. For these reasons, we’ve decided to use both types of models and blend them into a so called ‘gray box model’.
How does it work?
We want to use our gray boxing to model how temperature behaves. There are four components that have an influence on the temperature inside a building:
- Heat exchange between walls and air or between buildings and the outside world. The heat exchange is determined by: the surface area of the room/building (building parameter), the difference (delta) in temperature (thermostat data), and the heat transfer coefficient (a parameter for tuning).
- Solar input is the component that is shaped by weather forecast and tuned solar fraction, which defines how much solar energy can enter the room (given the room has windows). Incorporating blinds introduces the need for ML algorithms.
- Human input and devices is the component that impacts temperature changes based on people or devices in a room. It is derived from occupancy profiles, which can be static or dynamic. For static occupancy profiles, such as those of a traditional office building with fixed working hours, a simple white-box model will do. But for dynamic occupancy profiles, such as those of a university campus, we will need a ML model and internal heat factors that determine how much energy is added to the room by people and/or devices.
- Installations for cooling/heating all have their own energy consumption factor. This component refers to the efficiency with which these systems use energy to achieve the desired indoor temperature. For heating and cooling, the model aims to optimize the Coefficient of Performance (COP): the ratio of useful heating/cooling provided to the energy consumed.
With the given components, our gray box model can simulate many scenarios and form the optimal energy plan, with the lowest energy costs while safeguarding preferred comfort levels. Let’s highlight some possible scenarios to get a feeling of how it works:
- Instead of a set temperature, the model works with a dynamic comfort range. This way, the model has the flexibility to lower and raise temperature setpoints, to form the most efficient energy plan.
- When the weather forecasts show large amounts of solar energy from 12:00 onwards, the model can decide to stop heating the building at 10:00 or 11:00 in order to save energy.
- When energy is cheap early in the morning, and weather forecasts and expected building occupancy show an expected rise in temperature during the day, the system can decide to cool down the building a bit extra to prevent high cooling costs when energy is more expensive. Of course, comfort is again safeguarded in this cooling process.
Conclusion
Our gray box model combines the benefits of working with a physical model (easy & quick to set up, no historical data required), with those of a Machine Learning model (dealing with many interdependencies, high complexity, requiring lots of data).
Based on the four main components, it simulates the feasible heating/cooling combinations upfront to define how to come to lowest possible energy consumption. Each hour, it will repeat this process and adjust temperature setpoints accordingly, safeguarding comfort bounds.