Feeder: the design in ten minutes
Feeder is a governed operator for a distribution utility's outages and meters. It reads what the network already reports (meter last gasps and restores, feeder and transformer readings, tickets, crews, booked work, daily meter totals), finds what has gone wrong, acts inside written authority, and writes down why. This page is the short version; the full design has the detail. Everything shown runs on a synthetic network; known gaps says what that does and does not prove.
1. The problem, in one storm evening
A front crosses the region between 17:00 and 02:00. Trees come down on rural lines, fuses blow, reclosers lock out. The meters know first: each one that loses supply sends a last-gasp message within seconds. The control room usually learns from customers, twenty minutes or more later, when the second call about a street comes in. In those twenty minutes nobody has opened a ticket, so nobody has told the customers, nobody has checked who on the street depends on electricity for their health, and the crews are routed on yesterday's picture.
The same evening, a crew drives to a fault that remote switching has already cleared, planned work booked for the morning sits on a feeder whose backup is down, and transformers loaded by new heat pumps run past their rating in the cold. Away from the storm, the meter data has its own slow faults: registers stuck at zero, power flowing out of premises with no solar registered, meters that stopped talking just before their bill, and meters the network model places on the wrong feeder.
2. What runs
- Detectors are definitions in the dictionary: SQL with an owner and a version. The network ones run every tick of the logical clock; transformer loading and network-model checks hourly; meter data once a day after the overnight meter data run.
- A worker proposes one action for each new incident. A fixed rule decides when the runbook leaves one choice; a model is asked only where there is a real choice and is never required.
- The controller checks the proposal against
policy.yaml: is the action listed for this kind, which tier, do the arguments name only what the evidence names, is a cooldown or a budget in the way. It is separate from the worker and the model cannot change it. - Tools stand for one system call each (OMS ticket, outage notice, priority flag, meter re-read, metering case, work order, stand-down, switching request, load transfer, planned-work postponement). Each records the rows it changed so it can be undone.
- The ledger keeps every decision: incident, evidence, definition and version, rules or model, the call, the tier, who approved, and how to undo it. It is append-only.
3. Authority, in four tiers
| Tier | Meaning | Examples |
|---|---|---|
| 0 | Just do it | open the outage ticket the meters ask for, notify its customers, flag a priority customer, re-read a meter, open a metering case |
| 1 | Do it and tell | send a crew to inspect, stand down a crew driving to restored supply |
| 2 | Ask a person | switching to a tie feeder, moving load off a transformer, postponing booked work, a customer notice promising more than twelve hours |
| 3 | Never | protection settings, anything on the transmission side, disconnecting a customer, editing reads, tariffs |
Budgets cap automatic actions per feeder per hour and across the network per minute; the metering desk has its own. A budget that runs out sends work to a person, not into silence.
4. One vocabulary
Outage, sustained outage, restored, affected customers, priority customer and feeder load are written once and used by the detectors, the operator and the Ask screen. Where a term has two valid meanings (a sustained outage is over three minutes for the regulator and over five for network operations), Ask shows both answers with their owners and does not pick one.
5. Measured against the answer key
The generator plants ten kinds of fault and writes each one, with its correct handling, to an answer key. A replay runs the operator over the storm evening and the six days before it, with scripted on-call people answering the queue, and scores every detection: precision and recall per kind, the share handled without a person, reverted actions, time to resolve, and the weakest detector, which is named rather than hidden. Thresholds are tuned on two seeds and the published numbers come from a third.
6. The map and play mode
The main screen draws the network from the database: substations, feeders, transformers sized by meters, and a heat layer of dark meters, split into clusters with and without a ticket. The counters (customers off supply, open tickets, meters dark, dark without a ticket) come from the same dictionary terms the operator uses (affected_customers, outage), evaluated at the moment on the storm strip. Crews are placed from their last event: at depot, en route (interpolated along the work order), on site, or just cleared. The right rail lists the operator's own ledger rows up to that moment. Play mode copies the world to out/play, replays the storm from 17:00 with the operator running live at x60, x180 or x360, and the map reads that copy, so the recorded run is never touched.
7. Honest limits
The network, its faults and its people are synthetic; the physics is radial and simple; no adapter calls a real system. The way into a real utility is shadow mode: the same operator on a read-only feed, writing "would do" rows for the control room to review, before any action touches the network.