Portfolio Decarbonization

The Role of Machine Learning and Building Information Modeling

Imagine you are overseeing a large portfolio of buildings and need to reduce your carbon footprint. You may have several options to choose from when upgrading each of your buildings, all while balancing a strict budget and timeline. How do you decide how to allocate funds among the buildings? Which buildings should be prioritized? And how can you know whether you will ever reach your goals with the funds you have? This challenge is what we call ‘portfolio decarbonization’, a key aspect of capital planning.

Traditional methods for portfolio decarbonization often depend on labor-intensive analysis and lengthy simulations. These methods lead to static and obsolete reports and are becoming outdated; incorporating automation and machine learning (ML) can lead to much more efficient outcomes. Although machine learning has come a long way, there is significant promise for what can be done with future developments.

What are the key steps for portfolio decarbonization?

The strategy varies widely among companies, influenced by the nature of their properties and client base. The process is also significantly impacted by the location as different regions have different policies in place. Despite these variations, the general approach to portfolio decarbonization can typically be summarized into the following four key steps:

1️⃣ Defining the possible decarbonization strategies

There are different ways to decarbonize a single building. One way is to decrease its energy demand through enhancements in its structure, such as upgrading the envelope or roofing system. This could mean installing high-performance windows, improving exterior wall insulation, or modifying the window to wall ratio (WWR). Another approach is to meet the existing energy demand more efficiently, for example, by implementing energy-efficient ventilation systems. Additionally, integrating solar panels and opting for clean energy sources over traditional grid power are other effective options. Each option is considered an Energy Conservation Measure or an ECM.

2️⃣ Determining the impact of each strategy

Each ECM will have an impact on the performance of the building in terms of energy usage that will translate into a number of factors:

1. Green House Gas Emissions (GHG): Less energy means less GHG emissions. Some ECMs might not change the overall energy usage but change its source. For instance, using electricity is cleaner than gas. Using solar panels is cleaner than both. Depending on the type of the impact, we need to determine the changes in GHG emissions for each ECM.

   
   

2. Energy Costs: A direct consequence of altering energy usage is the shift in operational costs that come with it. It’s essential to translate changes in both energy consumption and energy sources into their corresponding financial implications.

   
   

3. Carbon tax: Many regions have taxation policies in place for the amount of GHG produced by building owners. The change in GHG therefore, translates into changes in carbon taxes paid to the government.

   
   

4. Policy penalties: There are new policies in place in many regions that place a cap on how much energy is used in a building and how much GHG is produced by it. If those caps are exceeded additional penalties will be applied to the owners. Here is a list of some of such standards for the US.

   
   

5. Other operational costs: Implementing a new system means that maintenance costs will also change, although usually for the better. Older less efficient systems require more frequent maintenance and repair services which would be likely not required as much with newer systems. Estimating such savings is usually very difficult and not included in the analysis which would result in more conservative estimates.

3️⃣ Determining the initial investment costs

Each ECM will also incur costs. Replacing windows is not cheap, and solar panels do not come free of charge! For each option, we need to determine how much it’s going to cost us to implement. This would be especially challenging if we are developing a renovation plan that spans multiple years as installation costs change over time. Having skilled cost estimators by your side is essential for a reliable analysis of this estimate. It should also be noted that some regions offer financial incentives for building owners who enhance their buildings’ energy efficiency to help offset the required initial capital. These incentives should be factored into the overall cost assessment of each ECM as well.

4️⃣ Developing the plan!

Once all the options, their impacts, and costs are determined, we can proceed to develop the decarbonization plan. In this step, we decide which ECMs should be implemented for each building and in what sequence. Essentially, this step prioritizes certain buildings over the others while simultaneously deciding the most efficient ECM for each building. Ideally, this should be done using advanced optimization algorithms that consider different variations of options and sequences. However, industry practices often still rely on the heuristic judgments of building science experts, and use manual processes.

What makes the current practices inefficient?

The process discussed above, although practical, is vastly inefficient considering the automation capacity that is available to us with today’s technologies. The most important inefficiency comes from the very first step. There is a huge limit to the number of ECMs we can consider simply because it takes a lot of time and analytics to determine all the impacts and costs associated with each strategy. Considering an additional alternative for the roofing system means that we need to perform another energy simulation (or some other form of analysis) to determine its impact on the overall energy usage. We also need to calculate the required initial investments, its energy costs and carbon taxes. Now imagine going through this process 10 times for 50 buildings. What makes it even more challenging is that we might be interested in the combined effects of the ECMs as well.

What if we want to improve the walls and change the roofing system for the same building? Can we simply stack their impacts? Or is the combined impact different from simply the sum of the two?

Challenges such as this force the engineers to only consider a handful of ECMs for each building and use simplified assumptions to scale the assessments on the whole portfolio. Otherwise, it wouldn’t be practical for analysis and planning.

How Can Automation and Machine Learning Help?

Automation proves most effective in processes governed by specific rules. For instance, once we know the energy impact of an ECM, we should follow a set of rules to determine its overall costs. We first need to determine the amount of electricity / gas usage, then we need to access the grid information for the location of the building, and retrieve energy costs as well as GHG emission factors for that grid. We also need to retrieve the policy caps and penalties for the building’s region and then we can easily calculate its overall operational cost. This process is entirely automatable, and we have already developed such a tool here at Adaptis. But how do we determine the energy impact in the first place? This is where machine learning can help significantly.

Today, we have access to large databases of building energy performance. BPD is a good example. It is “the largest publicly-available source of measured energy performance data for buildings in the United States.” It includes recorded energy measurements for hundreds of buildings as well as their physical and operational characteristics. Databases such as this provide an opportunity for training ML models that can predict the impact of a potential ECM on a building. This is a major opportunity for shifting from time-consuming simulations and error-prone approximations to reliable and quick ML-based predictions. Research is already showing promising results in such approaches. For instance, this paper developed a random forest model for predicting the choice of building envelope wall material on energy performance; and this paper provides insights on which ML models and features to use when predicting building energy performance.

The second area where machine learning can be very useful is predicting the maintenance and implementation costs of different ECMs. Unfortunately, due to the sensitivity of such data, it is hardly available for public use. However, there are growing cost databases such as BCIS (Building Cost Information Service) or RSMeans that can be reliable starting points for cost estimation.

The Future: From Building Information Models to Graphs

Most of the existing solutions reduce buildings to a bunch of parameters such as overall size, usage type, WWRs, and wall R-values. This puts a major limit on the performance of ML models as the level of abstraction is relatively high. In other words, two buildings with the same set of parameters will not necessarily have the same energy performance. There are many other parameters that define a building.

Future ML models will not be trained on structured tables of some building parameters. They will be trained directly on the Building Information Models that provide full details of the design of a building. However, the current technology has a considerable gap to get to this point. We still do not know how to feed an entire BIM model as raw input for a neural network. One idea could be to represent BIM as a graph structure and train a graph neural network on top of it. An interesting recent article used this solution for quality assessment of building elements. With few modifications, we could use the same data structure for energy impact prediction.

Another advantage of using a graph structure is that it enables the integration with other forms of data. We can’t represent everything about a building in a BIM model. Other forms of information such as condition assessment reports should also be considered. A graph structure allows us to combine BIM models with relevant textual documents. Research has already shown the potential in such solutions too. For instance, in this paper, I showed how we can link maintenance text records to BIM objects when we leverage graph data structures. However, production-level-ready technology on this matter is yet to be investigated. The most obvious one is that BIM is not practiced as often as it should be in the industry.

Many buildings do not have any BIM models, and the ones that do are primarily for visualization purposes. This often leads to a lack of accuracy, and makes them less useful for analytical purposes, particularly in energy performance analysis. The energy performance of a building greatly depends on small details such as wall materials, window types, and mechanical systems. If the building walls for example are not accurately represented in the BIM model, we will not be able to conduct reliable energy performance analysis. Consequently, the investment that the owner made in developing the model can not return much value.

If you’re investing in a BIM model, it’s crucial to do it right: don’t settle for a basic model that’s only good for rendering images.

A comprehensive, accurately detailed BIM model is a valuable asset, unlocking a myriad of practical applications from reliable energy analysis to effective project management.

#Decarbonization