Three energy modeling problems, solved again: what AI changed this year

Last spring I posted three short stories on LinkedIn about using Atlas, EP3’s AI assistant, on real EnergyPlus projects.

In June 2025 I wrote that energy modeling can’t just “move fast” with AI. Models drive expensive decisions, so results have to be verifiable. I posited that we still need deterministic energy simulation engines (EnergyPlus), and that we need to give modelers tools that allow them to stay in control of the process. I still believe that, but AI has gotten a lot better since then, and we can use it for larger and more complex tasks. We’ve also more deeply integrated AI into EP3, and provided more non-AI diagnostic and visual tools that are available to both the modeler and the AI.

Here are the same three cases, revisited.

Case 1: 2,000+ unmet heating hours

Then. Twenty years of EnergyPlus and I still got humbled. The model had over 2,000 unmet heating hours, on an HVAC configuration I don’t use often. I asked Atlas. In about a minute it had read the model, the simulated EnergyPlus file and the results, walked through its calculations and found the cause: the zones were served by VAV terminals with no reheat, and the central heating coil and setpoint managers couldn’t meet the zone loads on their own. It offered four possible fixes. I would have gotten there eventually. It got me there faster.

Now. EP3 has an Unmet Hours tab in the results viewer that works without AI. For each run it lists the zones with unmet hours, worst first, and the months they fall in. Expand a zone and you see its thermostat, setpoint schedules, zone equipment and sizing, each one a click away from the object in the model.

Three pieces make the difference on a case like mine:

  • Likely causes. EP3 checks each failing zone for the usual suspects: hard-sized equipment below the design load, setpoints outside a normal range, very high outdoor air, sizing factors below 1, equipment scheduled off, people in the zone while every system serving it is off, and no warm-up period. Each check names the objects involved. They’re labeled “worth checking,” because they are leads for an engineer to follow up.
  • Commonalities. EP3 lists what the failing zones share: the air loop, the thermostat, the schedules, the zone equipment type.- Profile charts and a tolerance preview. A typical day for the worst month, showing setpoints, zone temperature, occupancy, when the system was scheduled on, when it actually ran, and the heating it delivered. You can also preview the count at a different reporting tolerance.

Atlas reads the same diagnosis when you ask it a question, so the conversation starts from the evidence.

Why this matters:

Having visual tools makes it easier for human modelers to double-check AI suggestions, and often AI isn’t needed at all.

AI models do best when they have access to structured data and informative context. Providing AI with pre-structured information reduces AI cost and increases accuracy.

Case 2: A museum that couldn’t hold 45–55% RH

Then. The artwork needed 45–55% relative humidity. The air handler wasn’t dehumidifying at all, and humidification missed its setpoint often. Atlas queried the model, pulled the results files, worked through the problem step by step and found the dehumidification setpoint manager on the wrong node. Then it drew time-series graphs showing, hour by hour, where humidity control broke down. What I valued most was seeing its reasoning and its evidence side by side.

Now. Every setpoint manager in an EP3 export reports on the nodes it controls: node temperature, flow, humidity ratio for humidity managers, and the setpoint it is writing. In the time-series viewer, node values also appear on the HVAC diagram and follow your cursor along the chart, so a setpoint landing on a node where nothing responds to it is visible on the diagram itself.

When Atlas draws a chart for you, it builds it straight from the run data, in your unit system. The hourly values never pass through the conversation, which keeps the token bill small.

Case 3: Constructions from an architectural schedule

Then. I pasted a description of the envelope assemblies from an architectural schedule, and Atlas generated every material and construction correctly on the first pass. That was hours of transcription I didn’t have to do.

Now. You attach the schedule itself: PDF, Excel, CSV or text. Atlas reads the pages and sheets it needs. What it proposes arrives as a review card with a checkbox per item. Before you see the card, EP3 has already tested the change on a scratch copy of the model and exported it, so a proposal that would break the export goes back to Atlas first. Integrating it is one undo step. You can also aim a proposal at a new EEM, so the existing design case stays untouched until you decide.

What’s new that I didn’t have in the spring

Calibration against measured data. This is the one I’m most excited about for anyone managing existing buildings. The results viewer now has a Calibration tab: load a meter or utility bill export (CSV or Excel), and EP3 lines it up against a run and shows how well the model fits the way ASHRAE Guideline 14 measures it (NMBE and CV(RMSE)), along with the likely causes of any gap. It’s deterministic code with no AI involved, so the same data gives the same answer every time.

Atlas sees the same comparison when you work on calibration with it, so its suggestions start from the evidence. Fitting a model to bills will happily produce implausible answers. In a test on a small office model with synthetic measured data, a fit brought electricity bias from 11.8% to 0.2% and gas from −37.8% to 4.0%, partly by raising infiltration. Atlas flagged the infiltration increase as implausible: it adds gas use to a model that was already using too much. That’s the kind of judgement call I want help with.

Ready-made measures. Atlas can list the efficiency measures your model has objects for, ranked by their share of site energy from the latest run, and propose one as a new EEM.

Control over what the AI sees. Each model file has an Atlas access level: Off, Help, Analysis (reads the model and results, proposes nothing) or Full. New models start at Off. Atlas works with Anthropic, OpenAI or Google, or with a model on your own computer through LM Studio or Ollama, so nothing goes to an AI provider. Incognito mode swaps real names for placeholders before anything goes to a cloud model.

Bringing outside edits back. If you or your team have folders of EnergyPlus files edited in other tools, including by a general-purpose chatbot, EP3 can zip several variants into one model with EEMs, or merge an edited file back as its changes.

Why this matters for portfolios

If you manage dozens or hundreds of buildings, the bottleneck is the same diagnosis and the same data entry repeated across a portfolio, by different people, with different habits. Tools that make the evidence visible, keep a person approving every change, and let you choose which data goes to which AI provider are what make AI usable at that scale.

That’s the work I do through Infinite Atelier: helping property groups, campuses, hotel groups and funds figure out where AI fits in their building workflows, and where it doesn’t. If that’s a conversation you’re having internally, I’d be glad to compare notes.

In the spring I asked whether we underestimate how much of our time goes to diagnosis and transcription. Now I’m curious about the next step: if a tool could hand you the likely cause before you started digging, which part of your modeling week would you spend the saved time on?



Request EP3 pricing · Book a demo · More articles

EP3 for Mac

Tell us you'd use it on a Mac. It helps us decide what to build next, and we'll email you when it's ready.

Which computers would you run it on?
Mac is included because you asked about EP3 for Mac
Send me occasional news about EP3 software

Sending…

Thanks!

We've noted that you'd like EP3 on a Mac. Our reply is on its way to : check your inbox, and your spam folder if it isn't there in a few minutes.