May 22, 2025 • Edward Mellor • 4 min read
Alerts are only useful if they lead to action. Traditional systems swamp teams with threshold breaches that all look equally urgent. Ranking anomalies by estimated financial impact changes the game: the most expensive problems rise to the top, the noise sinks, and limited resources go where they matter most.
Estimate the cost of each anomaly by comparing actual usage to a learned baseline for the same conditions—weather, occupancy, schedule, and tariff. Multiply the excess kWh by time‑of‑use rates to capture real £ impact. Sort the queue by this value so an overnight baseload spike beats a brief daytime blip every time.
Context matters. Attach likely causes and recommended actions to each alert—“after‑hours HVAC runtime in Zone B; check schedule and fan start relay”—so engineers can move straight to resolution. Where confidence is low, flag as “needs validation” to avoid wild goose chases.
Suppress duplicates by grouping similar events into one incident, and dampen flapping with hysteresis and minimum duration rules. Silence known, low‑value patterns (e.g., brief door‑heater spikes) and auto‑close events that self‑resolve within a defined window. The aim is a short daily list that a human can realistically clear.
Ownership and routing are key. Send refrigeration issues to the refrigeration team, BMS faults to controls, and procurement anomalies to finance. With fewer, richer alerts, each team sees only what it can fix.
Turn alerts into tracked actions with owners and due dates. Measure time‑to‑acknowledge and time‑to‑resolve, and surface blockers in a weekly review. Close the loop by verifying savings post‑fix using the same baseline method—this keeps the ranking model honest and builds trust.
Over time, the system learns. Fix feedback tunes models, seasonal patterns adjust automatically, and repetitive issues get playbooks. The outcome is fewer alerts, higher signal‑to‑noise, and a faster cadence of verified savings.