Research supported by the Siebel Energy Institute has created a software platform that quickly identifies which structures would yield the most energy savings from the least expensive repairs.
Commercial buildings account for more than 72% of all energy consumption in the United States. Electricity is the largest operating expense in these buildings, with up to 60% of it powering heating and cooling systems. And yet, the majority of buildings waste significant amounts of energy due to minor system errors, like broken thermostats or improperly tuned HVAC systems.
Although these faults result in costly electricity use and the generation of unnecessary CO2 emissions, they often go unaddressed because they are difficult to diagnose and are rarely considered urgent.
But research supported by the Siebel Energy Institute has created a software platform that quickly identifies which structures would yield the most energy savings from the least expensive repairs.
The project, entitled, “Data Analytics to Assess Energy Efficiency Opportunities in Commercial Buildings,” leverages machine learning algorithms to recognize signature inefficiency patterns in buildings, zero in on the root causes of efficiency gaps, and assess quantifiable benefits of energy retrofits and emissions savings.
Led by Kameshwar Poolla, the Cadence Design Systems Distinguished Professor of Mechanical Engineering at the University of California, Berkeley, the research focuses on identifying buildings that would use markedly less energy with low-cost efficiency improvements, such as repair of HVAC infrastructure or recalibration of a building control system.
Poolla, with his research partner, Carlo Novara, a professor of Control and Computer Engineering at Politecnico Di Torino, developed and tested a large database of information about building electricity consumption that is cross-linked with meteorological data, construction details, and occupancy statistics. Their approach allows them to isolate and then compare equivalent systems.
Their findings show how modest operational adjustments could result in a 20% energy savings for a typical inefficient structure. The results also reinforce U.S. Environmental Protection Agency (EPA) studies, which indicate that a 30% energy savings can be realized through improvements to facilities and facility management.
Poolla spoke with the Siebel Energy Institute about the project.
Q: How would you characterize the potential societal impact of this work?
Buildings account for 33% or more of worldwide electricity consumption. It is well known that from design to construction to operation, buildings lose energy efficiencies. Even modest efficiency gains of say 5%, represent an enormous reduction in systemic greenhouse gas emissions.
Q: The platform could be valuable to organizations like government urban planners and commercial vendors of building energy management systems. What would it take for the software to be put into use?
We would have to validate our software tools at scale. Maybe hundreds of buildings. We would have to show that we can detect operational inefficiencies accurately, and that the error rates are small. This requires considerable field work, and much more funding than this Siebel seed grant program supports.
Q: It seems that identifying energy inefficient buildings is only half the battle. In your opinion, what would be an effective way to actually have the modifications enacted?
Since we are targeting operational energy efficiencies only, fixing the problem requires modifying the building management system. For example, the chiller staging schedule may be poorly chosen. In the case of faults, technicians need to repair thermostats. The burden of selecting and implementing modifications falls on the building owner. They have the resources (building maintenance staff) to do so, and once armed with energy savings estimates, we hope they are motivated to implement the necessary modifications.
Q: The study drew from publicly available data about energy use in U.S. buildings. What were some of the challenges you encountered when using the data?
The BPD database of the Department of Energy contains anonymized consumption data at 15-minute granularity. However, the BPD data base doesn’t allow you download data, it only permits access online. Further, the crucial metadata such as construction details, orientation, HVAC system sizing, etc. are not available.
Q: You needed to remove environmental effects, a process known as conditioning, to truly compare apples to apples in building systems. How did you deal with the conditioning variables?
We use modern machine learning methods to construct consumption models that incorporate dependence on conditioning variables. For example, a simple linear model describing the dependence on the conditioning variables of occupancy density and orientation might be
Consumption = α× occupant density + β × orientation angle + ….
We use Energy+ (a publicly available and commonly used building energy modeling tool developed by Lawrence Berkeley Lab) to construct appropriate model types, and verify these using comprehensive simulation. Of course, linear models are not good enough, and we have to use models that allow for more complex dependencies on conditioning variables.
Q: What were the types of variables you experimented with when testing the software on virtual building models?
Virtual building models allow us to compute counterfactuals. For example, how much energy would the building have consumed if it had a northwest exposure. We can condition and measure data and compute counterfactuals for square footage, occupant density, orientation, and shading.
Q: How does peer-to-peer comparison of buildings allow you to compute efficiency metrics?
We begin with a corpus of buildings and their consumption data. We then normalize the consumption data to compute per square foot counterfactual consumptions if all buildings have the same orientation and occupant density. This final metric allows us to compare buildings within their peer group. Building with high consumption metrics are likely to be operating inefficiently.
Q: How does machine learning enable the software to determine the root cause of operational energy inefficiency in a building?
This is work in progress. We will need data from buildings under various operational faults such as defective thermostats, or stuck dampers. This data is hard to come by. We may have to resort to Energy+ to synthetically engineer faults. Once we have a data corpus, we can use classification methods such as clustering to identify root causes.