Logiciel Contact Us
Success Stories Tech News Contact Us
whitepaper

AI Helps The Grid Only Where The Control Room Can Use It.

Two decades of flat demand planning just ended, and utilities are buying models faster than they can deploy them. This report maps the four domains where AI actually pays, the OT boundary that decides the architecture, and the evidence trail a rate proceeding will ask for years later.

In depth

Demand Is Rising Into A System Planned For Flat Load.

01

The trap most utilities walked into: buy the model first, hit the accuracy target, demo it to the executive committee, then lose eighteen months to data lineage, asset identity and a NERC CIP assessment nobody scoped, while GIS, the outage system and work management still disagree about which circuit the transformer sits on.

02

What the utilities doing better do instead: name the decision and the person who makes it (a dispatcher, a trader, a vegetation planner, an asset engineer), settle the NERC CIP boundary in week one, then buy distributions rather than point forecasts and score them where the cost actually sits, in the tail.

The detail

What Separates Grid AI That Reaches The Control Room.

Zone · 01

Models Stay Outside The Perimeter

Line differential and distance protection operate in three to five cycles, and the correctness argument for that equipment is a settings file and a test report, not a validation dataset. Anything statistical or periodically retrained belongs outside the electronic security perimeter, reading a one-way replicated feed. That boundary sets the architecture and half the cost.

Zone · 02

Distributions

Not Point Forecasts

Net load is harder to forecast than gross load was, because behind-the-meter solar makes a large share of supply invisible to SCADA. A single-value day-ahead forecast tells a commitment engine nothing about the tail, and the tail is where reserve, start-up and curtailment costs land. Require quantiles or ensembles, scored with pinball loss or CRPS rather than MAPE.

Zone · 03

An Asset Register Somebody Owns

Training needs failure history joined to asset identity, and that join is the real work. GIS says a transformer sits on one circuit, the outage system says another, and the field photograph disagrees with both. Fund it as an asset data programme with an owner and a dated deliverable. Calling it data preparation is how it gets cut.

By the numbers

The figures that make it a board-level conversation.

132
GW added to the ten-year North American summer peak forecast, more than double the prior assessment, with over half the continent at elevated or high risk (NERC)
945
TWh of projected global data centre electricity use by 2030, roughly double the 415 TWh of 2024, arriving faster than transmission can be built
2,300
GW of active capacity sitting in US interconnection queues, about five years from request to commercial operation, and only around 14% of past requests were ever built
80-160
GW of virtual power plant capacity needed by 2030, against roughly 30 to 60 GW today
Inside the report

What you'll take away.

01

Scope NERC CIP before you scope the model

Settle in week one whether the system touches BES cyber systems and who signs. That answer sets the architecture and roughly half the cost, and discovering it in month nine is what stalls these programmes for eighteen months.

02

Reconcile the asset register first

Circuit and structure identifiers across GIS, outage history and work management, with a named owner and a dated deliverable. One utility built a working vegetation risk model in eleven weeks, then spent nine months on the identifiers.

03

Start with asset health

Dissolved gas analysis, partial discharge, thermal history and breaker counts are already collected, the failure modes are physically understood, and the counterfactual is one finance recognises: an avoided outage or a transformer replacement deferred against a three to five year lead time.

04

log every recommendation for the rate case

Model version, input snapshot, output, who acted and what they did, retained as long as the capital and outage records they support. The algorithm ranked it 47th is not a prudency defence in front of intervenors. An ordinary engineering record is.

Questions

Frequently asked.

Can we put a model inside the control system if it improves response time?

Not on the deterministic path. Protection operates in three to five cycles and IEC 61850 GOOSE messaging assumes millisecond determinism. NERC CIP also imposes perimeter, patching and change controls that a weekly retraining
cadence cannot satisfy. Advise from outside instead.

Our forecast accuracy already beats the vendor benchmark, so what changes?

The scoring does. MAPE rewards a sharp point forecast and hides tail behaviour, which is where reserve, start-up and curtailment costs land. Score quantiles with pinball loss or CRPS against persistence and a weather model, and track calibration by hour in production.

Why does a machine learning project turn into an asset data programme?

Training needs failure history joined to asset identity, and that join is the real work. GIS, the outage system and work management routinely disagree about which circuit an asset sits on. One utility built a working vegetation model in eleven weeks, then spent nine months reconciling structure IDs.

Is virtual power plant participation not already solved by FERC Order 2222?

Market access is solved. Measurement is not. Advanced metering covers about three quarters of US customers, but interval data arrives on a daily backhaul, so dispatch is minutes old and verification is a day old. Settlement needs telemetry most aggregations do not have.

How do we defend a deferred transformer replacement years later?

With an ordinary engineering record, not a ranking. Keep the condition data, the model output as one input among several, the named engineer who agreed, and the reasoning. Retain versions and snapshots for as long as the capital records they support.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send AI Helps The Grid Only Where The Control Room Can Use It straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Review the boundary first.

Bring one control room or trading desk decision to a working session with our engineering leads, and we will map the OT boundary, the data work and the evidence trail.

Book an architecture review