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.
Three things you write every week. Watch what the model gets right, and what it invents.
Flag every claim you would not sign. One claim in this note is sound — leaving it alone is part of the score.
One day-ahead forecasting problem, four ways of looking at it — the artefact this whole drill is about, opened up so the room can argue with it.
Eight comparable days. Error is stated as absolute percentage error on the day's maximum demand block, against metered actuals.
It is: the 10th-to-90th percentile spread of actual demand across the eight most similar historical days, block by block. It describes how much days like this one have varied in the past.
It is not: a bound on the model's error. The actual can and does fall outside it — on 14 April in the table below it did so for nine consecutive blocks. A band drawn from history says nothing about a condition history has not seen.
This is the same decision log as the role panels, pre-filled with the scenario you just ran. Complete it and it joins the printable log at the bottom of the page.
21:11 on the same day. An alarm-management layer collapses the flood into three probable causes and suppresses the rest. Judge each collapse — then open the bin.
Eight things that cross a despatch desk. For each, decide whether it can go into an AI tool at all.
This is not a new imposition. TGTRANSCO already runs Cyber Suraksha as a standing channel for employee IT-security awareness — the organisation long ago accepted that an employee's everyday computing habits are a security matter. This module extends that instinct by exactly one step: the browser tab is now also an exit route for control-room data.
This module teaches the instinct, not the policy. Where any specific item sits under TGTRANSCO's own classification is for the IT wing to say. What is being drilled here is the habit of asking the question before the paste, so that when the policy does land you already know why it says what it says.
Your call, your reason, your name. The log is where the name goes.
AI flags, ranks, drafts and estimates. A human despatches.
Entries you generated today, in the register of an SLDC internal log. Print this and keep it.
Every accuracy figure carries four things. Without all four it is not a number.
Five questions in order, then the call. Works on an alert, an anomaly flag or a forecast revision.
Eleven terms. Drift and suppression are the two nobody explains, and both cost you quietly.
| Term | In plain language |
|---|---|
| Model | A system that has learned patterns from past data. It has no physics in it unless someone put physics in it. |
| Training data | The history it learned from. If the system it learned on has changed, what it learned is now partly wrong. |
| Drift | The model quietly getting worse as the grid changes underneath it. Nothing alarms. The forecast simply degrades. Somebody has to own noticing. |
| Confidence band | The spread of comparable historical days. Not a bound on the model's error. |
| MAPE / MAE | Average percentage miss / average MW miss. |
| Anomaly detection | Flagging a pattern unlike the past. Unlike the past is not the same as wrong, and not the same as dangerous. |
| False positive | An alert on nothing. Costs attention. |
| False negative | The thing it did not flag. Costs everything else. |
| Suppression | Alarms the system decided not to show you. Invisible by design. Open the list after every event. |
| Hallucination | A stated fact the model produced because the sentence needed one. In this track it appears as an invented cause. |
| Human in the loop | AI flags, ranks, drafts and estimates. A named officer despatches. |
Five instructions. The first one is the sentence that keeps a manufactured cause out of a settlement document.