Seasonal signatures: catch deviations before the invoice
A consumption spike on Tuesday at 2 p.m. is not an anomaly if it matches your usual profile. Useful detection starts by modeling what normal looks like.
Octowise Team
Fixed thresholds (“alert if power exceeds 80 kW”) create a lot of noise on tertiary portfolios. A spike at 2 p.m. on a Tuesday in winter can be perfectly normal; the same spike on a Sunday in August is not. Octowise trains a seasonal signature per data source (month × weekday × hour) so it only triggers on meaningful deviations from the site’s usual profile.
Why the calendar matters more than a threshold
A tertiary building does not have an “average” consumption. It has regimes: weekday occupancy, weekend idle, winter preheating, summer cooling, school holidays, one-off events. Comparing the current reading to a rolling average hides those cycles and creates two symmetric errors: false alerts on predictable peaks, and silence on drifts that stay under the threshold while being abnormal for that hour and day.
Energy managers already know this empirically: HVAC starting earlier in January, night-time safety lighting, canteen load at noon. What is often missing is a model that captures this business recurrence instead of treating it as statistical noise.
How the signature is built
Octowise learns a signature per data source, not a single curve for the whole building. The model crosses month, weekday (week starting Monday), and hour. The profile of an electricity meter on a Wednesday in February at 8 a.m. is not mixed with a Saturday in August at 8 a.m.
Training relies on actually collected history. The more stable and continuous the source, the more discriminating the signature. A recently connected source, or one that drops too often, yields a more conservative model: fewer alerts beat unactionable ones.
- A distinct signature per source (meter, BMS, API), not an overly broad building average.
- Month × weekday × hour granularity, aligned with operating rhythms.
- Alerts only when the gap versus the usual profile is significant, not on the first overshoot.
From drift to health score
Each identified deviation is not an end in itself. It feeds the building health score and can trigger an in-app, email, or mobile push alert according to team preferences. The goal is not to add another BMS alarm layer, but to give managers a prioritized signal: this site is leaving its usual regime, at this hour, on this source.
It is also a portfolio review lever. On the portfolio map, the AI score and alert volume show which buildings are actually drifting, beyond DPE letters, which only change with each new diagnostic.
What the signature does not replace
A seasonal signature detects the unusual. It does not say whether a business setpoint is being respected. For that, operating conditions remain the right tool: percentage variation, minimum duration, time window, recurrence, sensor groups.
The two approaches complement each other: the signature draws attention to a drift nobody encoded as a rule; operating conditions formalize a known operating requirement (no heating on weekends, chilled-water setpoint, and so on).