Autodomos: Predicting Dwelling Heat Load Based On Internal Heat Pump Data 

Andries van Wijhe, Olav Vijlbrief, Nederlandse Organisatie voor toegepast-natuurwetenschappelijk onderzoek (TNO), Netherlands

A digital twin of an individual dwelling, including a heating system, can be used for predicting heat load. This is useful for determining when a heat pump should run based on forecasted ambient conditions (e.g., temperature, solar gain, wind), internal heat load, and electricity availability. In work conducted by TNO, the Autodomos algorithm has been developed, which generates a digital twin based on operational data from the heat pump. An Autodomos-generated model is able to predict the dwellings’ heat losses/gains and is able to calculate the required heating to be provided by the heat pump. This model can be used for self-configuring heat pump installations or to unlock the flexibility potential of a heat pump.

Introduction

Domestic heat pump installations face two important challenges.

1. The availability and price of electricity on the electrical grid.
2. The time and knowledge required by installers for installing and commissioning heat pumps, including after-sales service.

Both of these challenges can only be addressed if more information is known about the dwelling and its heating system. It needs to be known how much is the heat loss. How much will the dwelling heat up due to solar gain, and when? What happens if the heating is turned off/down for one or more hours, and what is the effect on efficiency when it is restarted at a later stage?

These questions can be answered with help from a digital twin of the installation, including the dwelling, heat emitter system, and the occupant. However, the time and knowledge required to make such a digital twin makes it unfeasible for mass deployment. In this work, TNO has developed an algorithm that can create a digital twin from the internal measurements of a heat pump without human interaction. The algorithm requires the following data points:

1. Supply water temperature; 2. Return water temperature; 3. Room temperature obtained from the thermostat; 4. Water flowrate or produced heat; 5. Weather conditions (can be obtained from weather station)

The digital twin uses a physics-informed neural network model of the dwelling with a heating system. Instead of putting all the data in a machine learning algorithm, it starts with physical relations of heat transfer. The unknown closure relations are then fitted using data-driven techniques. The result is that less data is required for training, and more realistic predictions can be made outside the bounds of the training data. Also, the found parameters have physical meaning.

Heat Balance of a Dwelling

A typical dwelling has several heat flows, which can be sources or sinks. In a dynamic system, an imbalance leads to a change in its state parameters. The rate of change of the temperature of a mass is defined by the amount of imbalance and the thermal inertia of the mass.

For a dwelling, five heat flows are identified:

1. Transmission heat loss due to conduction via the façade of the dwelling. This is a function of the insulation value and the temperature difference between the indoor and outdoor environments.
2. Ventilation heat loss: The air exchange from indoor to outdoor and vice versa. This is a function of the amount of the temperature difference between indoor and outdoor, and the ventilation/air leakage rate. The latter can be a function of wind speed, wind direction, and occupant behaviour.
3. Internal gain: The amount of heat the occupants and appliances release in the dwelling. This is a function of mainly the user behaviour
4. Solar gain: This is the heat gained by the dwelling via the sun. This is a function of solar strength, sun azimuth and elevation, and occupant behaviour (shading).
5. Heating gain: This is the heat added to the dwelling by the heating system. This is the variable to be controlled to keep the indoor conditions within satisfactory boundaries.

Functional Description of the Autodomos Algorithm.

A digital twin of a dwelling created by the Autodomos algorithm is able to distinguish the five heat flows in a prediction.

The Autodomos algorithm works in three consecutive steps:

Step 1: Building energy signature. In this step, a three-day average of the heat load is fitted against the three-day average ambient temperature.

Step 2: Fitting of the dynamic system. For an accurate dynamic response on the short- and long timescale, the model should be redefined to a two-mass building model, which is shown visually as an RC-network in Figure 1. In this step, the building energy signature from step 1 is used as a starting point for fitting a dynamic system.

Step 3: After completing step 2, the model is quite effective in determining the dynamic behavior of the dwelling and heating system. This model can be used to predict the effect of turning off the heating system for several hours. However, the model has some residual error because it lacks a few aspects:

-proper definition of solar gain, ventilation, and internal gain

-influence of user behaviour on the different heat gains and losses.

By using parallel shallow neural networks, this residual error can be linked to the missing aspects of the model:

-The effect of solar gain can be linked to the elevation and azimuth of the sun. The effect in this case is determined by how much the solar radiation contributes to the heat balance of the dwelling.

-The ventilation rate can be an effect of the time of day and day of week and is assumed to be determined by occupant behaviour.

-The internal gain can also be matched to time of day, day of week, and month of year to describe user patterns.

By fitting the residual error after step 2 to the three aspects above, the unknown gains (and losses) can be attributed to physical effects. In many AI/Machine learning applications, the outcome is a black box. Since Autodomos uses physics-informed neural networks, the outcome can be linked to physical phenomena, which also makes it explainable. Another benefit of physics-informed neural networks is that it requires less data to train since the main relations are already established before the learning procedure. This should also make it more robust in conditions that did not occur during the training period.

A typical prediction made by the model can be found in Figure 2.

Physical Explanation of the Machine Learning Fitted Parameters.

Since the model acquired by the physics-informed neural network is not a pure black box, the outcome has physical meaning. For validation purposes, the outcome can be sanity checked against physical parameters. In Figure 3, this is done for the internal heat load. The graph shows a typical internal heat load by occupants and appliances during the course of 24 hours. A clear morning and (larger) evening peak can be distinguished, which likely corresponds with this occupant’s daily habits.

A second explainer is shown in Figure 4. On the left, the outcome of the model: the sensitivity of the dwelling to solar irradiation. Blue indicates that the solar irradiation will not cause the dwelling to heat up; red indicates high sensitivity of the dwelling to solar irradiation. It is noted that the solar sensitivity is high if the sun’s azimuth is South-West between 210° and 250° and the sun’s elevation is above ~16°.

Figure 4 (right side) shows an aerial view of the dwelling. In this case, it is a terraced house. By marking the directions 210° and 250° as found from the analysis of solar gain influence on the heat balance, it actually matches the orientation of the house façade (and thus windows)

On the bottom graph, it shows what the angle of the sun should be above to add influence the heat balance of the dwelling. This coincides very well with the data found from the learned relation between the position of the sun and the influence of the sun on the heat balance.

The same conclusion can be drawn for the elevation. If the sun sets below an elevation of 16°, it is hidden behind the roofs of the adjacent housing block. In Figure 4-right, this distance is shown in the aerial photograph. In Figure 4-bottom, this is graphically represented, and in Figure 4-left, the effect is clearly visible in the result.

It is worth noting that the model is trained solely on data from the heat pump, in conjunction with open-access meteorological data. The required heat pump data comprise supply temperature, return temperature, flow rate, and thermostat temperature.

Another advantage of the physics-informed neural network is that the fundamental physical relations are fixed. Since merely the closure relations are fitted, the model will not likely predict unphysical behaviour outside the domain of the training data, which is the case for black box models.

 Conclusions

In this work, it was shown that a comprehensive dwelling model can be trained using mainly heat pump operational data without user/installer intervention. It is Important that the physics-informed neural network techniques give physical meaning to fitted results and require less data when compared to deep learning techniques. Also, these models are more likely to predict well outside the training data range.

The working of the algorithm has been shown; however, it is still in the proof-of-concept phase. It is yet to be determined what the minimum data required is, and it should also be tested on a larger variety of dwellings and heating systems. Currently, two projects on the development of the application are being set up. In one project, the dwelling model will be used for deriving ‘ideal’ settings for space heating without intervention, thus creating the basics for a self-commissioning heat pump.

In the second project, the dwelling model will be used in combination with a moving horizon controller. The dwelling/heating system/heat pump model will be used to create heat pump control scenarios for the coming period, hence the name horizon. These scenarios will be evaluated in terms of energy costs, comfort, and electricity availability. The scenario with the highest rating will be executed by the heat pump.

TNO is looking for partners for further development. As the national research institute of the Netherlands, TNO is not a product developer. If parties are interested in developing the IT infrastructure or using this technology in their products, please contact the author of this article.

Author contact information