What actually counts as an actionable alarm
Most monitoring platforms generate noise. A practical framework for alarm thresholds that get acted on instead of ignored.
5 min readThe problem with default thresholds
Most monitoring platforms ship with default alarm thresholds tuned to avoid missing anything, which in practice means they fire constantly. An engineer who gets a dozen low-priority alerts a day learns, reasonably, to stop reading them closely. By the time a genuine fault shows up, it's just one more notification in a list that's already being skimmed.
Three tests for whether an alarm belongs
Would a specific person do something different because of this alert, right now? If the honest answer is "note it and move on," it's not an alarm, it's a log entry, and it should be routed there instead.
Is the threshold set against this plant's own baseline, or a generic default? A performance-ratio drop that's meaningful on a well-maintained plant might be normal variance on one with a higher baseline soiling rate. Thresholds copied from a template rather than derived from the plant's own history are a common source of noise.
Does the alarm distinguish a real fault from a known, explainable condition? Grid curtailment, scheduled maintenance downtime, and a cloudy afternoon can all look like a performance drop if the system doesn't know to expect them. An alarm that fires during expected downtime trains people to ignore the next one too.
A structure that tends to work
In practice, a small set of high-confidence alarms, inverter offline, performance ratio drop beyond a plant-specific threshold sustained over a set window, and string-level anomalies, catches almost everything that actually matters. Everything else, minor fluctuations, single-day dips, informational events, belongs in a log an engineer can review on their own schedule, not a push notification competing for attention.
Routing matters as much as thresholds
An alarm that's correctly tuned but sent to a shared inbox nobody owns is functionally the same as no alarm at all. Every alarm needs a named recipient and an escalation path if it goes unacknowledged. That's a process decision, not a technical one, but it's the part that determines whether the alarm actually gets acted on.
Want this run on your own plant?
Send us your generation data and we'll tell you where you actually stand, not just what the theory predicts.