Comment by Craig Mellis, Senior Advisor- Decarbonisation of Complex Sites, and Samuel Young, Practice Manager – AI, at Energy Systems Catapult

What do we mean by ‘Artificial Intelligence’ (AI)? It seems like every software solution nowadays boasts that it uses AI, but, as with other buzzwords like ‘sustainability’ or ‘digital twin’, what people mean by it and how people interpret it varies hugely.

This can be a problem when it comes to procuring and using AI tools. If what users think of AI as being differs from how it actually works, then real world decisions can be made based on flawed assumptions and misplaced confidence.

In the course of our work with non-domestic energy data and supporting Net Zero innovators, we have spent time working with AI experts to understand how the market is evolving, what AI products are available and how they deliver. The question we want to address is:

How can you tell whether a particular AI solution is appropriate for your situation?

Generative vs predictive AI

Generative AI (like ChatGPT and other ‘large language models’) essentially works by answering the question “what are the most likely words to come next?”. It is focused on words and what is most commonly written, rather than underlying data and facts, so the answers it gives are usually plausible but not always grounded in fact. When you hear software “uses AI” and the software involves documents or a chatbot, this is probably what it means.

Predictive AI analyses more structured data, often in a spreadsheet or database, and looks for patterns in that data. When you give it new data, it compares that data to the patterns it found in previous data and suggests things that might also be true of this new data. You may sometimes see this referred to as machine learning (ML). This is a bit more grounded in data and facts, but it is still only extracting patterns from past data, so if the data is incomplete or the future is quite different from the past, then it won’t necessarily give the right answers.

Applying AI to heat decarbonisation plans

Let’s look at an example to illustrate this. Consider the case where we want to develop heat decarbonisation plans for sites or buildings based on previous heat decarbonisation plans for similar sites or buildings.

A generative AI approach would feed the existing pdf-based heat decarbonisation plans into a ChatGPT-type tool which might then generate pre-populated decarbonisation plans for new sites. However, because the approach is “what text is most likely”, these new plans will tend to be based on the most common applications and recommendations, with a thin veneer of site customisation, rather than reliably understanding the needs of each site and customising plans accordingly.

A predictive AI approach would take a wide range of structured information about a site (e.g. location, building areas, energy usage) and use similar information from other sites to find correlations and make recommendations. It is more likely to recommend approaches that have been pursued at similar sites and remove options that are clearly unsuitable for a site from examination of the data.

Of course, neither can compete with the nuance and detail available from a site-based walk around and audit (“Phew!” I hear some of you engineers say) but the second is likely to provide a better suggestion of the options which could be then investigated whereas the first may well provide spurious results that are less useful.

It is also important to recognise that neither approach is ever going to be more accurate than its input data. An AI can only go on the information it has been provided with, and if that information is insufficient or incorrect it will make incorrect recommendations – usually without alerting the user to the fact key information is missing/wrong! This is where engineers and people who really know the site really shine – they can much more reliably spot when something isn’t accurate or realistic for a site.

In some cases, hybrid approaches combining different types of AI may be more effective. For example, generative AI may be able to extract structured data from text for a predictive AI approach to use (e.g. “Does this decarbonisation plan involve solar PV?”) or turn the output of a predictive AI model into more user-friendly text.

When developing building decarbonisation plans:

• Generative AI is less likely to provide reliable insight about specific sites
• Predictive AI with access to structured data about lots of buildings may be able to group buildings to help identify and prioritise investigation and action
• Expert knowledge will still be required to tailor a plan to each site

Does AI actually help us?

Well-chosen AI tools can be a powerful force in helping us pick up the pace in getting our wide and varied non-domestic estate to Net Zero. However, reliance on AI tools that are inappropriate for the task also has the potential to undermine good decision making and slow us down.

To ensure we use AI appropriately it is important to:

1. Understand what a specific instance of AI is, and is not, capable of (you may sometimes need support from people with greater AI knowledge for this).
2. Evaluate whether the data provided to AI contains enough of the important information for the AI to make reasonable recommendations.
3. Define a small number of detailed examples that you can use to test how reliable an AI’s outputs are.
4. Ensure human expert knowledge is applied after or alongside AI recommendations, rather than following them blindly.