Recently we released new artificial intelligence (AI) technology that changes the way energy is managed in buildings. How does it compare to generative AI technologies like ChatGPT, DALL-E and Midjourney?

There are similarities in approach, but let’s wind back a bit and start with a different question: what problem are we trying to solve here? What opportunities are we targeting? Why on earth would Buildings Alive invest in creating a state-of-the-art AI system when amazing tools are already freely available and evolving at an alarming speed?

Why we need a smarter approach

The problem—and the opportunity—is that buildings consume more than 55 per cent of the world’s electricity and, with electrification, that proportion is rising. Buildings are also the gateway to distributed energy resources (DER) like rooftop photovoltaics (PV), electric vehicles (EV), stationary batteries and mechanical equipment. We need buildings to provide a flexible, efficient, and grid-interactive resource and we need them to do it consistently, at scale, to drive more investment in clean energy generation and decarbonize the global economy.

Energy efficiency has been a consideration for building owners and operators since the oil shocks of the 1970’s and the techniques for achieving savings have not changed markedly since that time. Engineers with expertise in energy efficiency are a prized commodity, typically developing their skills over many years of conducting energy audits and experimenting with tuning and controls. There are few textbooks, fewer graduate courses, and even fewer datasets explaining the ‘how to’ of energy efficiency in buildings. It takes years to develop expertise because unlike vehicles, for example, every building is somewhat unique and changing all the time. Knowledge must be developed and applied in context.

Recent advances in fault detection and diagnostics (FDD) have changed things to a degree. Data is more readily available, and analytics can identify opportunities for energy savings that are just too time-consuming for humans to find. However, most FDD technologies are still only as good as the rules that engineers and technicians develop for them – they reliably and unintelligently follow their engineers’ instructions. They don’t learn, and they certainly don’t get smarter.

How our AI works, differently

This gets us to why we created a sophisticated AI tool for analyzing and improving the energy performance of buildings. Generative language models that run AI-powered chatbots and image generators like ChatGPT and DALL-E need to be trained on massive amounts of data. They essentially vacuum up everything on the web they can access. That data is then processed through an artificial neural network that spots patterns and makes predictions. When asked to do something, an AI chatbot or image generator will return what the patterns in its training data suggest should be a helpful response. Over time, these responses improve through reinforcement learning (RL) from training provided by users in the form of feedback (thumbs up, thumbs down, etc.) and other algorithmic reward methods.

Ask an AI chatbot to tell you why a building’s electricity demand profile differs from a target profile at a specific point in time, and what should be done about it… and you will find (as we did) that the responses will be vague and not particularly helpful! Often it will be completely wrong, and that stands to reason. How would it know? How could it have learnt? To be clear, the bot could learn to be helpful, but it would need to be extensively trained just like an expert. And that is what we have set out to do.

Our approach has been to process fine-grained timeseries and event data from hundreds of buildings gathered over many years on the Buildings Alive platform using a combination of supervised and unsupervised learning methods, such as Support Vector Regression, K-Nearest Neighbors, Random Forest, Change Point Analysis and others.

Having generated models with deep understanding of the patterns of energy use for each building on the platform, and the relationships between them, the technology is then able to identify anomalies, their possible causes, and suggestions for action and improvement. Every anomaly-cause-suggestion set is reviewed by a Buildings Alive engineer and optionally by the message recipients themselves (can have multiple reviews for each set) to ensure the AI is given the opportunity to learn rapidly from experts and given high-quality results to aim for. This dual application of reinforcement learning on human feedback (RLHF) helps it to build expertise quickly and function effectively with only very limited information to work off. For example, it can provide expert insights with only the basic meter data used to produce utility bills.

Importantly, the AI recognizes similarities in buildings’ operational performance patterns and characteristics. This allows observations and learnings at one building to be transferred to another to accelerate overall learning and savings. Already we have found that the rolling-average positive confirmation score (thumbs up) for each anomaly-cause-suggestion set has risen from just over 50% at release three weeks ago to 70% now – a remarkable rate of learning, even for an expert!

What does it mean for operators of buildings?

The goal in all of this is to draw attention to performance anomalies that can save building operators’ time, cut costs, and cut carbon emissions. Our goal is for every building, every day, to have a world-class energy expert hunting for savings and drawing attention only to those opportunities that are worth acting on.

In the screenshot below, for example, the building is clearly achieving significant savings compared to the pre-commencement standard (“base range”) and coming close on this particular day to its target performance – the model’s assessment of its full potential under the conditions. Accordingly, the AI has drawn attention to (1) the early start, a series of power spike and cycling issues (2-4), and also the effect of a recently implemented tuning initiative (5). (Please note, the interface is interactive, so much of the features and information can’t be fully demonstrated in this article.)

The AI is capable of offering a wide array of suggestions which depend on a lot of factors, many of which are not visible to the user. For example, a “spike” or “dip” may occur due to heating or cooling control issues. The platform knows what temperature and humidity conditions the building was under, hence what the “possible cause” is likely to be and it bases its suggestion on that. Likewise, it can identify a possible cause based on the size or ‘signature’ of the anomaly, or when it occurred – it might be a pump or a fan for example.

All of this, obviously, is only the beginning. We have an array of advances coming through our development pipeline that will more directly address our goal of helping all buildings to be carbon efficient and grid-interactive.

At Buildings Alive we know enough to be deeply concerned about where AI is taking humanity. The precautionary principle must always apply and some recent AI developments are raising the alarm. Likewise, we are deeply concerned about the impact of climate change. Our approach, therefore, is to try to build technology that can only do good. We believe the technology breakthrough described here is an example of that.

If you would like to know more or work with us, please reach out.

 

 

* Banner image generated by https://deepai.org/machine-learning-model/cyberpunk-portrait-generator from text prompt: “A cheerful image of an office building in the form of a brain on a happy sunny day with trees and nature”.