Despatch Desk TGSLDC · TGTRANSCO · in association with NI MSME · delivered by TechBaton

Track 2 · Grid Operations & Forecasting · 1 day
Start here

Same grid. Six distances from the moment it matters.

Five drills. You leave with a printed decision log, and four things for the control room wall.

All figures are synthetic. Nothing here is quotable or fileable.
The despatch clock · six desks, one gridselect your desk
Every desk works the same system. What separates them is how far they sit from real time.
The rule we teach before anything else

AI flags, ranks, drafts and estimates. A human despatches.

No model closes a breaker, clears an outage, issues a schedule, states a cause in a settlement, or puts a figure into a filing. Every one of those is an act by a named officer, and the log is where the name goes.
Take it back to the control room

My decision log

Entries you generated today, in the register of an SLDC internal log. Print this and keep it.

The control room line — three checks, every single time

  1. Nothing that is live, internal or identifying goes into a tool outside the control room. Published means published; everything else is a no.
  2. Every figure carries its denominator. An accuracy number without a period, an aggregation level and an exclusion list is not a number, it is a mood.
  3. The model advises and I decide, and the log records the reason — not the recommendation.
Four take-homes

Four things for the control room wall

Your decision log is the personal one. These four are the same for every desk in the room, and every one of them comes out of the printer.

Take-home 01Forecast accuracy quick-guideThe four things every accuracy figure carries, what the common measures actually mean, and three questions for any claim.
Take-home 02Alert triageFive questions in order for every alert, and the trust / verify / reject call at the end of them.
Take-home 03Operations AI glossaryEleven terms, said plainly. Drift and suppression are the two nobody explains and both cost you quietly.
Take-home 04Prompt libraryFive guardrails that survive the day. The deviation one is the sentence that keeps a cause out of a settlement.
Take-home 01

Forecast accuracy quick-guide

Every accuracy figure carries four things. Without all four it is not a number.

What a number has to carry
All four, or it is a mood
  1. The period. A month, a season, a fortnight of one weather type?
  2. The aggregation. Averaged over blocks, over days, or over the peak only? A daily average hides a bad morning inside a good night.
  3. The exclusions. Which days were left out, and who decided, and when?
  4. The measure. Percentage of what denominator — actual, scheduled, or installed capacity?
What the common measures mean
Plainly
Three questions for any accuracy claim
Ask them in the room, out loud
The band is not the error. A confidence band drawn from historical spread describes how comparable days have varied. It says nothing about a condition history has not seen.
Take-home 02

AI alert triage

Five questions in order, then the call. Works on an alert, an anomaly flag or a forecast revision.

Every alert, in this order
The order matters — question 1 kills more alerts than the rest together
  1. What data is this built on, and is that data healthy right now? A failed sensor produces a confident alert.
  2. Is this alert new, or is it the same alert re-firing? Re-fires are the largest single source of alarm fatigue.
  3. What does the alert claim, and what does it merely imply? Models flag; they do not diagnose.
  4. What is the cost of acting if it is wrong, against the cost of not acting if it is right? Asymmetric costs are the whole basis of triage.
  5. If I suppress or defer this, where does it go, and who opens that list?
After every event, before the log closes: open the suppressed list. No exceptions, whatever the shift workload. Grouping is visible and reversible. Suppression is invisible by design — that is the whole point of it.
Trust · verify · reject
Where the alert lands once you have asked
Trust
The claim states its mechanism, its input and its direction, and every one is checkable. A revision you can trace is a revision you can use.
Verify
The output is plausible and consequential, and the check is cheaper than the error. Almost everything lives here.
Reject
The claim carries no denominator; or it asserts a cause; or it is contradicted by data already in front of you.
And the fourth state, which is not on anybody's list and should be: accept and log the doubt. You acted on it because the clock ran out, and you wrote down that you were not sure. The decision log exists for this state.
Take-home 03

Operations AI glossary

Eleven terms. Drift and suppression are the two nobody explains, and both cost you quietly.

TermIn plain language
ModelA system that has learned patterns from past data. It has no physics in it unless someone put physics in it.
Training dataThe history it learned from. If the system it learned on has changed, what it learned is now partly wrong.
DriftThe model quietly getting worse as the grid changes underneath it. Nothing alarms. The forecast simply degrades. Somebody has to own noticing.
Confidence bandThe spread of comparable historical days. Not a bound on the model's error.
MAPE / MAEAverage percentage miss / average MW miss.
Anomaly detectionFlagging a pattern unlike the past. Unlike the past is not the same as wrong, and not the same as dangerous.
False positiveAn alert on nothing. Costs attention.
False negativeThe thing it did not flag. Costs everything else.
SuppressionAlarms the system decided not to show you. Invisible by design. Open the list after every event.
HallucinationA stated fact the model produced because the sentence needed one. In this track it appears as an invented cause.
Human in the loopAI flags, ranks, drafts and estimates. A named officer despatches.
Take-home 04

Prompt library — the guardrails that survive the day

Five instructions. The first one is the sentence that keeps a manufactured cause out of a settlement document.

Decision log · 0 entries · not printed