Using AI for home energy assessments

I've been researching how AI could be used to electrify homes faster and more efficiently. We asked ChatGPT to help create an electrification plan for our house.

The order of recommendations ChatGPT gave were spot on but other info is less reliable

I have been playing around with ChatGPT to see if I can get it to create an energy assessment of our home without having to try and program something from scratch.

It's easier said than done. Those who know more about the inner workings of ChatGPT will know that as a large language model, it is very good at sounding convincing, but this doesn't mean what it says is even close to facts.

So, I've been working on a loose plan to address this. To avoid ChatGPT's urges to go off the rails, I began trying to guide it through an analysis of my home energy step by step, hand-holding it along the way through a systematic process.

After quite a bit of back and forth and trial and error, which I'll get into in just a second, I managed to get ChatGPT to create a rough guide on how we should electrify our home based on the probable savings and the cost of the upgrade.

In the context of our conversation, ChatGPT rightfully identified that replacing the gas hot water heater with a heat pump should be the highest priority. But let's explore how this list of priorities was created.

I uploaded photos of my house, a full year of consumption history, gas bills, and all the pricing

In the conversation with ChatGPT I gradually gave it the context of our home and what our energy consumption looks like. To do this I uploaded a dozen images around our home and told it to examine each image, identify the device in the image and give me an approximate of how much energy the device consumes. Identifying objects in photos is actually a strong point of the AI, so this went pretty well.

For example, ChatGPT guessed a MacBook would use around 60 watt-hours. The actual rated consumption for the MacBook is 69 watt-hours, a 15% mistake but not bad given it's a best guess.

Once we identified all of the devices in the house worthwhile, I uploaded a complete year of my energy consumption data, along with the price of the energy and my bills into the conversation.

We then used all of this information to create a profile of how all these devices were consuming energy data. For example, if our energy consumption from 6:00pm to 7:00pm on July 10th was 1.8 kWh, I wanted ChatGPT to guess which devices were most likely the culprits for that consumption.

This required ChatGPT to ask me a few more questions to get the context of our home's consumption, such as how often we run the reverse-cycle air conditioner and clothes dryer, what kind of light bulbs we use, and so on.

Calculating the rough savings of each upgrade relative to its impact on our bill

From here it was a matter of doing a few calculations of how much each proposed upgrade would save us. This was based on a rough esitmate of the impact each upgrade could have on our energy bills. Say if we replace the gas hot water heater for a heat pump how much gas consumption would that save, and then how much money does that convert into.

Once we had a table of the predicted savings, it was just a matter of prioritising those savings by highest to lowest.

The final step was asking ChatGPT to give a rough estimate about how much each of these upgrades might cost in Melbourne. In the end, it got close enough on most of the upgrades, but these are self-generated values and would need more work in the future. A couple of upgrade suggestions, like insulating the walls and ceiling are pretty off.

There were a few obvious mistakes that ChatGPT made, but overall not a bad first effort

Rome wasn't built in a day, and neither is trying the route to make ChatGPT to do a full accurate energy assessment for a home. Naturally, without strict guidelines and data being fed in, it will start to make up numbers and suggestions. Even when it is fed data it will sometimes still ignore those numbers.

One noticeable missed opportunity is that ChatGPT never realised if we removed our a stove, we would be fully off the gas and, therefore, would no longer need to pay the daily gas supply charge. Then, of course, sometimes ChatGPT gave into its basic instincts and would occasionally create some numbers from thin air.

But this is just one small step down the road of creating a tool that can make fast simple home assessments that anyone can use at anytime. Its certainly possible to see a future around the corner where an AI is trained on lots of home energy assessments and trained to behave.

A future where every house can get a free assessment within a few minutes or, at the very least, a tool for assessors to use to enhance their assessments.