Machine Learning for Retrofits

State-of-the-art

The renovation and retrofit of old buildings and state-of-the-art machine learning isn’t the most common marriage. Much like it would give a Canadian pause to see a polar bear and penguin living together harmoniously. However, the march of sustainable development continues, and my role as an associate researcher with the University of Victoria’s Energy in Cities group has given me experience in understanding the relationship between old and new. Machine learning can be used to identify potentially valuable renovation targets on a neighbourhood, provincial, or even federal level, and help define strategies about which upgrades will have the largest impact.

I’m excited by the broadness of the lab’s goals. Why model the performance of one building when there’s technology and methods available to test 1000s? The following describes some of the steps in this process, while recognizing the contributions and hard work of my colleagues in the lab and peers in the field. Hopefully readers will excuse my brevity in describing various aspects of the research. This article aims to be more communicative than rigorous academic document, lending understanding to these concepts when they inevitably arise in personal and professional settings.

Communities

So many buildings, so little time. Which of these buildings in East Vancouver should be renovated first?

Stripping down to the bare essentials, we need to imagine we are looking down at a typical community trying to decide which buildings are the best candidates to upgrade with our limited resources. This makes attacking the lowest performing buildings first a good strategy since it allows for proportionally greater impact.

There are many machine learning approaches to calculating an estimate of either a given building’s energy consumption or its suitability for renovation. In this vein we’ve seen projects apply computer vision to Google Streetview or satellite imagery. These methods have relatively high barriers of entry because of their need for extensive pre-processing of the data before training and high computational cost at time of execution. Building time series data is more readily available, and therefore training recurrent neural networks, which specializes in classifying and predicting sequences, has some attractive advantages.

RRNs might rank as one of the more abstract things to understanding in machine learning. Assuming some knowledge on the reader’s part, RNN layers have some similarity to a standard linear layer (the neuron) but with the addition of a state memory. This state memory is utilized when a sample sequence is fed into the layer one timestep at a time. As the timesteps accumulate within the layer’s state memory, it’s constantly being compared to the output of the training data for the sample and updating the layer’s weights to capture how the whole sequence is behaving to arrive at the known output during training. In this way, latent knowledge from within the same time series or a set of time series can be uncovered.

For the application described in Targeting Buildings for Energy Retrofit Using Recurrent Neural Networks with Multivariate Time Series (2019), the authors used RNNs to analyze a building’s electricity use along with a selection of meta data about the building. They used this information to, first, classify a building’s heating system and, secondly, predict a building envelope’s thermal performance. Glossing over many details, an RNN model essentially maps a multivariate time series containing features like interior temperatures and energy consumption, to categorical variables of known heating systems and envelope properties contained in the training data. For deploying such a technology, one feeds in an unknown building’s time series into the model and the network will work out an estimate of its heating system or thermal properties. As someone who works with all these technologies, this can be a more efficient approach.

Candidate Buildings

When walking down the street in Vancouver I often note the state and age of the buildings around me, and recognize their potential for energy retrofits.

I’m not sure how common an activity it is for people in the architecture and engineering field to be walking around a neighbourhood and reflexively note the age and state of the buildings around them. Certainly it’s something I will admit to doing. However, when faced with no many streets to walk down, the above section described how we can efficiently find candidate buildings without needing to check every street, avenue, nook, and cranny. Now having found a candidate building to potentially renovate, my lab colleagues have also investigated methods of discovering the most cost-effective upgrades for said target building.

Our lab has investigated methods looking at a candidate building and predicting which upgrades would be the most impactful and cost-effective. One approach written about in Using Multiple Linear Regression to Estimate Building Retrofit Energy Reductions (2018) blends together methods of simulation and optimization. The method uses multiple-linear regression techniques to then construct marginal cost curves. Linear regression defines a simple relationship between an independent variable and dependant variable. Multiple-linear regression extents this by relating multiple independent variables to one depend variable. In this case the independent variables are all sorts of building meta-data, and it’s balanced against energy consumption and cost. Marginal abatement curves are a way of measuring and visualizing tradeoffs to then predict the best renovation strategy. The details of how this works mechanically would need a lot of polishing and hype to make it interesting, but the results are significant in reinforcing known pathways to decarbonization which will be discussed in the next section.

Corroborating some of the above conclusions in a completely different fashion, our colleagues at ETH Zürich (to who I am connected with through my supervisor Dr. Ralph Evins) used a more straightforward approach to uncover cost-effective retrofit solutions. We often review related work for ideas and guidance, and in A machine learning-based framework for cost-optimal building retrofit (2021) we can learn a lot about predicting energy use from RNNs and feature selection on a per building basis.

Feature selection is a necessary evil in machine learning, and this paper really makes a specialty of it. The topic is deeply intertwined with sensitivity analysis, and can greatly reduce the computational load of the calculation if only the most important data is fed into the neural network during training. When trying to approximate complex behaviour, there is a small risk with this approach of getting rid of data the model might have found useful. However, in practice, the decrease in accuracy is negligible compared to increase in computational efficiency.

This work is interesting because of the high-resolution of their forecasted behaviour after retrofit. Many approaches aggregate the energy savings predicted, Chirag Deb et al. work on the other hand returns a time series. The output will look very similar as the users’ behaviour of when they draw power and heat will be the same, however the amounts should differ by the degree of energy savings.

Conclusions

The Canadian Green Building Council says targeted building renovations and retrofits could contribute to a 51% drop in building greenhouse gas emissions by 2030.

When and where to use such techniques, and how to balance the tradeoffs of one algorithm against another, should be considered in the context of each project. However, the conclusions the two papers above have reached for individual buildings (along with many other supporting papers from the literature to really drive home the point) are consistent in suggesting two common strategies. Though the reported improvements differ in magnitude, the math is clear; a retrofitted building could expect to see a greater than 30% reduction in energy and heating costs with the implementation of 1) building envelope upgrades and 2) heat pumps retrofits.

As Canada turns toward electrification of everything and increased use of renewable energy sources, heat pumps have an important role in reducing the greenhouse gas emissions and of the existing building stock. BC has an excellent climate and renewable electricity supply for heat pump operation, though it’s accepted in Canada some supplementary heating will be required in the winter when temperatures drop outside the optimal heating range of heat pump systems. On the other hand, as climate change continues and summers get hotter, many will welcome the air conditioning benefits a heat pump provides.

Retrofitting a building with a heat pump system can be cumbersome, but building envelope upgrades on an existing building are truly extensive. The installation quality of the insulation and air barrier of decades gone by can be suspect because of the sometimes complex and difficult building conditions encountered on site, furthermore, these systems can degrade over time dragging down performance from the initial design intent. Specialized renovation firms today attempt to make it as smooth as possible an experience for building owners, but nothing overcomes the temporary pain of redoing a building’s exterior.

I don’t feel overwhelmed when considering how large the existing Canadian building stock is or the magnitude of details that realistically go into planning a building upgrade. I feel well armed with the tools required to narrow down the scope strategically to target high-value renovations. The Rebuild project undertaken by our lab intends to further shape, support, and expand these efforts in the coming years. Assuming a high degree of uptake, BC could expect to see a 51% decrease in building emissions by 2030. This is a substantial amount when buildings contribute about 40% of Canada’s overall greenhouse gas emissions.

References

Baasch, G.M. and Evins, R., (2019). Targeting Buildings for Energy Retrofit Using Recurrent Neural Networks with Multivariate Time Series. In Neural Inf. Process. Syst.

Bowley, W., Westermann, P., Evins, R. Using Multiple Linear Regression to Estimate Building Retrofit Energy Reductions., IBPSA-Canada’s biennial conference themed Building simulation to support building sustainability (eSim), 10–11 May 2018, Montreal, Canada.

Deb, C., Dai, Z. & Schlueter, A. (2021), A machine learning-based framework for cost-optimal building retrofit, Applied energy, vol. 294, pp. 116990.

About the Author

Blair lives in Vancouver, B.C. and is currently an Associate Researcher at the Energy in Cities group at the University of Victoria. Follow him on Instagram or connect with him on Linkedin.