Power Grid Monitoring: Tennessee Valley (TVA)

Findings report · scored 2020-01-08 to 2026-04-01 · hourly · all times UTC

Summary

Over 6.2 years of hourly TVA data the monitor fired 10 alerts. Of 1 target event: 1 detected (Elliott, 7.8 h lag); 0 missed. 8 alerts fell outside every target window: 0 inside explanatory windows declared in the run config and 8 unexplained.

At a glance

Scored period6.2 years54,383 hours

Target events1

Predicted0 / 1

Detected1 / 1

Missed0

Alerts fired101 first, 1 repeat

Explained0hindsight label

Unexplained81.3/yr; 1.3/yr with explained

How it is scored

An event is PREDICTED if the first alert falls within 168 hours before onset, DETECTED if it falls within 48 hours after onset, and MISSED otherwise. The same windows apply to every event. The reasoning is in “Too early is the same as wrong”.
Onset is the time the grid operator declared the highest emergency level the event reached, converted to UTC.
Consecutive alerting hours count as one alert. A later alert inside a target window is a repeat, with no credit and no penalty.
An alert outside every target window is “explained” if it starts inside an explanatory window declared in the run config, and “unexplained” otherwise.

Target events

Event	Outcome	Lead / lag	First alert (UTC)	Onset (UTC)	Onset basis and source
Winter Storm Elliott EEA 3, first rolling blackouts in TVA's history	DETECTED	7.8 h lag	2022-12-23 19:00	2022-12-23 11:12 2022-12-23 06:12 local	TVA declared EEA 3. FERC, NERC and Regional Entities, Inquiry into Bulk-Power System Operations During December 2022 Winter Storm Elliott (Oct 2023), p. 10 and fn. 26

Winter Storm Elliott DETECTED

Scoring window: 2022-12-16 11:12 to 2022-12-25 11:12 UTC (168 h before onset to 48 h after).

First alert at 2022-12-23 19:00 UTC, 7.8 h lag. Models flagging in that hour: forecast error.

Repeat alerts inside the window: 2022-12-24 09:00 UTC.

Every alert

All 10 alert episodes in the scored period, in order. Nothing is left out.

#	Start (UTC)	End (UTC)	Hours	Class	Matched to	Models flagging at start
1	2020-04-21 17:00	2020-04-21 22:00	6	Unexplained		demand vs normal, forecast error, net generation vs normal
2	2020-12-01 00:00	2020-12-01 02:00	3	Unexplained		demand vs normal, forecast error, net generation vs normal
3	2021-11-29 03:00	2021-11-29 10:00	8	Unexplained		demand vs normal, forecast error, net generation vs normal
4	2022-06-14 02:00	2022-06-14 05:00	4	Unexplained		forecast error, net generation vs normal
5	2022-12-23 19:00	2022-12-23 20:00	2	First alert	Winter Storm Elliott	forecast error
6	2022-12-24 09:00	2022-12-24 10:00	2	Repeat	Winter Storm Elliott	forecast error, net generation vs normal
7	2023-05-10 02:00	2023-05-10 07:00	6	Unexplained		forecast error, net generation vs normal
8	2024-01-19 06:00	2024-01-19 08:00	3	Unexplained		demand vs normal, net generation vs normal
9	2025-01-07 01:00	2025-01-07 04:00	4	Unexplained		demand vs normal, forecast error
10	2025-12-07 15:00	2025-12-07 20:00	6	Unexplained		demand vs normal, forecast error, net generation vs normal

Configuration of this run

Signals, one HTM model each: forecast error, demand vs normal, net generation vs normal.
A model flags an hour when its anomaly likelihood is at or above 0.995. A model's warmth is the share of flagged hours over its last 8 hours.
Signals are grouped (table below). The system alerts when 2 of 2 groups are hot for at least 3 consecutive hours.
Warm-up: the first 168 hours are not scored. Learning stays on for the whole run; no labels are used.

Group	Signals	Hot when
structure	demand vs normal, net generation vs normal	2 of its 2 signals flag in the same hour, or the group's mean warmth is at least 0.45
residual	forecast error	1 of its 1 signal flags in the same hour, or the group's mean warmth is at least 0.5

Limits

One target event is a case study. It shows how the detector behaves on this event; it does not establish a detection rate.
The replay treats each hourly value as available at its timestamp. EIA publishes demand about an hour after the operating hour and net generation the next day, so a live deployment at these lead times would need the operator's own telemetry.
"Explained" is assigned in hindsight from explanatory windows declared in the run config. It never counts as a catch, and the alert rate is given both with and without those alerts.
The alert-rule settings listed above were set per region. Only the HTM model parameters, the encoder shape, the one-week warm-up and the consensus method are shared between regions.
218 hours are absent from the run's timebase because at least one input signal was missing in the source data.