Prediction and AMI Events

The Responder Prediction Engine uses outage information to determine the potential source of an outage. This outage information may come in various forms, including phone calls from customers and events from AMI devices. Calls and AMI events are processed by the Prediction Engine in much the same way with some differences. Both utilize enhanced Prediction to ensure faster performance. Prediction uses partitioning to optimize performance when processing AMI events.

Enhanced Prediction requires that AMI events hold in the state data store for a set amount of time. The Post Event Dispatcher queries the state data store for events that have been held for the required amount of time. It passes these events to the Post Event Processor pipeline where Responder incidents are created for partial power events. The remaining events are used to update their duration in the state data store to "stewed" which indicates they have been held in the state data store for the appropriate amount of time.

At regular intervals, the Prediction Service queries the state data store via the State Service for events that have been held for a set amount of time and have a status of "stewed." It receives all events that meet the criteria, up to 1000 by default. This limit was chosen to ensure optimal performance. You may modify this limit in the Prediction Services configuration file.

Next, any events that can be associated with an existing incident are assigned in Responder and removed from the original list received from the State Service. For the remaining events (from the original 1000 or fewer received), the Prediction Engine uses partitioning to increase efficiency in processing the AMI events.

Partitioning

Partitioning manages how the Prediction Engine processes AMI events. The Prediction Engine selects an AMI event from the remainder of the original 1000 received and locates it in the network, then traces upstream to the feeder source. This allows it to focus processing on the entire feeder. It then traces downstream and sends this collection of traced load points to the State Service. The State Service queries the state data store for events that reside on the load points in the collection and that have set for the requisite amount of time. These events are sent to Prediction and marked with a state of "predicting." The Prediction Engine can then predict on the entire feeder and mark them as predicted out in the state data store.

Prediction and Canary Devices

If your network makes use of canary devices and they are configured correctly, the Prediction Engine will respect these when processing AMI events. When Prediction encounters an event from a canary device, it traces to the nearest upstream protective device and predicts on that device. When an incident is created or detected for that protective device, all devices below it (including the canary device) are predicted out.