Case study 1

A fired gas heater in live service at a 300,000+ bpd refinery

4 days¹

model deployment, from papers the plant had already prepared.

30 to 50%²

lower than vendors' total cost: deployment, installation and annual licenses.

0³

unsafe “keep running” calls across 2,000 simulated scenarios.

Request a demo

Want to see more?

The heater

Item	Spec
Type	Process heater
Service	Naphtha
Fuel	Fuel gas
Draft	Forced draft, air preheater
On fan loss	Natural draft, air doors
Burners	Gas, with pilots

The study

Question: can it forecast the heater's fuel gas pressure a week ahead?
Given: 27 months of process data, its P&IDs, operating procedures and cause and effect matrices.
Compared with textbook physics alone and an off-the-shelf forecasting tool alone, on its real history, in 922 overlapping one-week tests.
A second set of 614 overlapping tests, each spanning one week, compared an updated model with a model held fixed after its initial setup.

Can I defend these numbers?

The rules were locked before the first forecast. Each method learned only from the past and was scored on the next week.

We recomputed the 922-test results and found zero differences.⁴

How does it compare with the usual ways?

Range width, same hit rate (share of real readings inside the range). Shorter is better.

Dominus1x

Off-the-shelf forecasting tool alone1.42x

Textbook physics alone10.1x

Against textbook physics, a reading has to move about a tenth as far to stand out.

Alarms per 12-hour shift, same limit⁵

Textbook physics2.16

Dominus0.37

Across every shared limit, physics alone sets off 2x to 6x the alarms.

Next-week readings in range⁶

Updated model95.2%

Fixed model89.3%

What it learned about this heater

On the early data:

All 21 big pressure jumps fit the pattern of operators adding or pulling burners (found on that same data).
A burner law (pressure versus gas flow) measured from this heater cut what-if error by 42% to 51% against the textbook law.

Request a demo

Want to go deeper?

How does it handle uncertainty?

Safe limit Range Reading

Keep running Act now

Illustration: dotted means outside the range; calls plan on the top edge.

Under 0.18%³

upper bound on the rate of wrong safe calls, at 95% confidence, in practice runs like these.

Each forecast includes an uncertainty range.

Recommendations use the upper edge of that range.

What does one trip put at stake?

$500,000⁷

at stake per heater trip, with example inputs.

For your engineers: the setup

Item	Spec
Advises on	Standard process control
Also advises on	APC, RTO, trips, shutdown calls
Where it runs	On-premises

IRIS provides advice without writing to your control, alarm or safety systems.

Want to see it live?

We show what caused a simulated trip and which procedure step was running.

Request a demo