Building owners and operators consult energy efficiency auditors to have their building’s energy efficiency evaluated and to obtain a list of recommended retrofits to consider for their building. Unfortunately, these retrofit measures can be difficult for building owners to identify on their own, resulting in the need for external talent.
While Google searching is something we can do when we don’t know an answer to a question we have, information about recommended retrofits to improve building energy efficiency is spread across academic papers, thick manuals, technical reports, and web articles. Many of these mentioned materials can be time-consuming to digest and difficult to judge whether an energy efficiency measure is established, experimental, or relevant for a given building type and climate region.
Experts spend time and resources becoming proficient in estimating the potential savings of energy efficiency projects so that these building managers do not need to.
But what if artificial intelligence (AI) could parse this information for building owners and operators directly and provide pragmatic, accurate advice on projects to implement, removing energy efficiency experts (and the associated cost) from the equation?
AIs are computer programs that output information or make decisions, typically at a scale and complexity than a human can do themselves. Typically, in the building energy efficiency community, AI is used for numerical problems, such as predicting the energy of a building based on a supervised learning model, trained on large observational data sets. This is how weather-normalized energy is calculated, for example.
Another type of AI involves processing text data like search engines. Much of the information about energy efficiency is contained in text documents and can be accessed using search engines like Google Search. We’ll touch on how AI is making these smarter in a little bit, but let’s first understand what’s happening in a search engine.
Search engines vary in their capabilities. Simple search engines have a set number of queries and responses. Simple search engines can be difficult to make general and flexible since the programmers must include all queries and responses. A complex search engine, like Google Search, can highlight points in an information database that are similar to a query.
Search engines like Google Search can aid users that are willing to organize the raw query results. An ideal search engine would sift through the data to determine relevant information and then summarize it to provide the user with the exact answer they need. This ideal search engine could summarize information for the average user.
Okay, now to the cool stuff.
OpenAI, a research organization dedicated to improving AI, presented a step toward the ideal search engine.
OpenAI’s GPT (Generative Pre-trained Transformer) is a language model that was trained to predict the next word after being given a sequence of words. For example, when GPT is fed the input “The quick, brown fox”, it completes the sentence with “jumped over the lazy dog’s back”. What GPT outputs is heavily influenced by recent input. The heavy influence of recent words means that GPT can be quickly adapted to new language data that it has not used for training. When GPT is fed relevant input, it can generate relevant information and act as a search engine.
For example, to prompt GPT about retrofits suitable for a car dealership, one can input into GPT “Common retrofits to a car dealership for improved energy efficiency include”, which could generate the following text as an example:
“Common retrofits to a car dealership for improved energy efficiency include on-site solar system installation, LED lighting retrofit, and optimization of heating and cooling economizers”. These, of course, are helpful measures that can reduce building utility expenditure and improve net operating income – some helpful advice!
To try GPT for yourself, follow this link: https://bellard.org/textsynth/
GPT is strong at outputting word sequences that sound reasonable, but due diligence is required to avoid errors in the language model’s output (this is a cautionary comment and we don’t mean to deter the model’s use!). The mistakes that an AI makes versus those of a human are different, and this difference can lead to better judgment together than if a human or an AI were to judge alone.
In Thinking Fast and Slow, Daniel Kahneman reviewed a study where cancer diagnoses were more accurate from human-aided AI than diagnoses from humans or AI alone. The increase in diagnoses accuracy was due to the human (radiologist) adjusting their safely pessimistic judgment with the statistics-driven judgment of the AI. Language models like GPT can make obvious mistakes, but the models can still generate relevant text in a readable format from many sources.
Text-oriented AI like GPT could be a tool for building owners and operators to confidently explore relevant energy efficiency measures for their buildings without the external help of an expert consultant.
OpenAI’s GPT is a step toward a language system that could be fed raw energy efficiency content from which insights could be generated. In an energy efficiency application, there would still be work to be done for GPT to streamline parsing of the enormous energy efficiency literature. If GPT pulled this off and was able to develop reliable models, non-energy experts could access tailored insights about energy efficiency with a simple voice command.
This could shake up the industry on how buildings evaluate and improve their efficiency, and do so at much lower cost to the building owner or operator.