20 November 2023
By Giulia Cirri
This blog is divided into two parts. If you’ve missed part one, don’t fret! You can find it here.
In the previous part of this blog, we looked at the state of the UK building stock, and just how important retrofitting homes is to reach net zero goals, but also for households’ health and wallets. When we think about AI in buildings, we most often think about smart buildings. I was curious to investigate other ways for AI to support improvements in energy efficiency, and in particular, to improve the physical building stock.
The use of AI for retrofitting seems to be in the early stages of research. While there is definitely interest in the academic community, most studies are recent, and still in their initial phases. Nevertheless, there are some interesting studies using AI to support decision-makers.
A promising area for the use of AI relates to removing uncertainty from the retrofitting process. Indeed, experts note how retrofitting is marked by a high level of uncertainty and risk due to the complexity of the task, the lack of information on buildings, and competing interests (e.g. heritage-related concerns). This makes tools to support decision-making very useful. Two main ideas for the use of AI caught my attention:
One thing AI is very good at doing is using large quantities of data to produce predictions: this makes AI very useful for climate adaptation in general, and this also applies to retrofitting. Usually, current tools to simulate the impact of retrofitting require intensive data input, high expertise, and long computation time. Moreover, retrofitting is often found to perform worse than expected, achieving lower energy efficiency than anticipated. One of the reasons for this underperformance is the large heterogeneity in the retrofit effects: a retrofit action might have different levels of effectiveness in different buildings. This heterogeneity is hard to capture through current simulation tools: while simulation tools allow full control of physical and environmental parameter settings, they also require a lot of data input, expertise, and computation time. To simplify the model, many tools use prototype buildings, speeding up the analysis and providing a quick screening of retrofit alternatives. Nevertheless, this leads to a loss in accuracy: the prototype building does not necessarily match the real, existing building that will be retrofitted. This mismatch often leads to the retrofit performing worse than expected. Moreover, most simulation-based tools do not adequately assess operational or behavioural changes or their impacts on energy efficiency, which further increases the mismatch between simulation and reality.
The authors of this 2022 study on commercial buildings propose a data-driven approach, based on AI, to solve this problem. They use data from the U.S. General Service Administration (GSA), a government agency providing office space, services, and goods to government agencies. Between 2010 and 2015, a series of retrofit actions were undertaken in buildings in the GSA portfolio. The study quantifies the retrofit effect of those actions and uses the information to provide decision support for future retrofits.
In short, the researchers quantify the retrofit effect of six groups of retrofit actions: for instance, improving the building’s roofs, facades, or windows, modifying the heating, cooling, and ventilation systems (HVAC), or installing a system of smart metres. In order to assess the precise impact of the retrofit actions, researchers also needed to identify what variables other than the retrofit actions could have had an impact on the building’s energy consumption – the confounding, or control, variables. For instance, these variables include building characteristics, such as size and age, energy use, and weather variables. Researchers then apply a machine learning method with causal forest to quantify the retrofit effect considering all the control variables. They also identify and rank the importance of the confounding variables.
The model can be reused to support decision-making: based on measured data on energy use and weather history, building characteristics, and other control variables. Portfolio owners and policymakers can train models that estimate energy savings for past retrofits, predict savings for future retrofits, and target buildings with high predicted savings. Moreover, since it is based on measured data, this approach has the potential to better reflect the savings actually achieved. It is also less computation and data-intensive, therefore making it easier to be applied to a whole portfolio of buildings.
As mentioned, in its current form, the retrofitting process is expensive in terms of time, effort, resources, and expertise, and it fails to deliver quick and objective solutions for massive datasets. Therefore, researchers in this second study aim to create a data-driven decision support system using AI to quickly support the choice of the best retrofitting strategy.
Researchers use AI (in particular, clustering) to process and group a massive database of building information in Lombardy, Italy. This first database, CENED, provides information on a building’s period, energy label – and therefore energy consumption – and the transmittance of walls, windows, and roofs, which provides an indication of how much heat is lost through each of these respectively. The researchers were able to link this data with another database, TABULA, containing the typical building materials used for each period and transmittance value. Through their analysis, the researchers can understand which materials correspond to higher energy efficiency for buildings of the same period – and therefore which retrofitting practices work better for buildings with particular characteristics.
In short, this study uses AI to process a massive amount of open data and energy retrofit decision-making systems. The model identifies a given building’s construction technology and materials based on its energy parameters and records in national databases, and determines which materials bring the most optimum retrofit strategy to obtain a particular (higher) energy-efficient label. According to the authors, the study provides a precise, reliable and quick assessment tool that can serve as a reference point for energy retrofit strategic decision-making and planning on a building or urban scale.
While the field of AI for retrofitting is still emerging and requires further research and development, AI has the potential to support the building stocks’ net-zero objectives by supporting retrofitting practices.
However, the success of AI in this field, as in all the fields, really hinges on one thing: having enough data to work with. Experts highlight how building databases often lack the essential information needed to assess the viability of specific energy conservation measures. Moreover, these databases might vary differently across, and even within, countries: data availability might depend on local legislation, or even on the typology of buildings. In order for AI to be widely implemented in the area, it is fundamental that data is consistently available. The European Union is making important progress in harmonising databases across its member states – for instance the TABULA database used in the second study was created as part of an EU project.
As we head into winter, the issue of energy-efficient and well-insulated homes will once again be on our minds daily. While the will to implement retrofits is entirely human, AI might come to give us a nudge in the coming years.