Anomaly detection
Article 3 of 4 · The analyses in depthWhat is anomaly detection in energy data? Spotting deviations before they cost money
Direct answer
Anomaly detection is the automatic recognition of consumption that does not match a meter's normal pattern: an installation that keeps running, a clock set wrong, a meter reporting strange values. Because no one can review every chart every day, the detection does the watching and you only receive a signal when something genuinely deviates.
- Deviations visible immediately
- From chance to system
- Data quality guarded too
- Signal plus next step
Anomaly detection versus manual checking
The most expensive energy problems are rarely spectacular. It is the heating that keeps firing in an empty building, the cooling that stayed in manual mode after a fault, the time schedule that shifted an hour at the switch to summer time and never shifted back. Each of them invisible in daily work, and only felt months later on the statement, by which time it is impossible to trace when it started. Anomaly detection is the answer to precisely this problem. It learns per meter what is normal and raises a flag as soon as actual consumption deviates from it. No daily chart reviews, no lucky finds, but a system that watches continuously and only rings the bell when something is genuinely going on. This page explains why that matters more than it sounds, which components interact within it and how detection works under the bonnet, from the first alert to the statistics behind it.
Key points
Moment of discovery
Traditional approach
At the annual statement or by chance.
Modern approach
As soon as the pattern deviates, often the same day.
Coverage
Traditional approach
The meters someone happens to look at.
Modern approach
All meters and sites, every day anew.
Workload
Traditional approach
Reviewing charts takes structural time.
Modern approach
Only genuine signals demand attention.
Evidence
Traditional approach
Reconstructing after the fact when things went wrong.
Modern approach
The starting moment and the pattern are immediately traceable.
Why deviations remain invisible for so long
Energy consumption has a property that makes it dangerous: it always keeps running, whether anyone is watching or not. A faulty valve control does not complain, a wrong time schedule gives no notification, and a building that is heated too warm at night is simply comfortable during the day. Moreover, at most organisations no one has the task of reviewing meter data daily, and that is entirely understandable. The result: deviations live unnoticed for months on average, and by the time the statement raises questions, the cause can no longer be traced. Anomaly detection breaks that pattern by reversing the roles. It is not the human who searches the data, but the data that reports to the human, only when there is something to report. That makes the difference between structurally throwing money away and a short repair at the right moment.
- Deviations give no notification themselves; they hide in normal-looking days.
- No one has time to review all meter data daily, and no one needs to.
- The later the discovery, the harder it is to trace the cause and the starting moment.
- Detection reverses the roles: the data only reports when something is going on.
The components of good anomaly detection
The heart of anomaly detection is the expectation pattern: per meter, a picture of what is normal for the time of day, the day of the week and the season. Actual consumption is continuously checked against it. The second building block is context: the same deviation means something different in a round-the-clock care facility than in an office that closes at the end of the working day, so the detection must know the nature of the site. The third building block is data quality monitoring: missing quarter-hours, duplicate or implausible values and meters that have gone silent are anomalies themselves, and whoever fails to catch them ends up chasing ghosts in polluted data. The fourth and most important building block is follow-up: a signal only becomes value when someone picks it up. That is why every alert comes with a logical next step, and every recurring alert with the question of whether the pattern itself needs adjusting.
- Expectation pattern per meter: what is normal for this moment, this day, this season?
- Context per site: a care facility and an office have different pictures of normal.
- Data quality monitoring: gaps and implausible values are signals themselves.
- Follow-up: every alert gets an owner and a next step, otherwise the value evaporates.
For the expert: how detection works under the bonnet
Technically, anomaly detection is a trade-off between sensitivity and calm. A simple fixed threshold, for example on night-time consumption above a limit value, is robust but misses everything that deviates within the threshold. Pattern-based detection compares the actual profile with an expectation built from historical behaviour, corrected for weekly and seasonal rhythm, and also flags subtler shifts: a base load that creeps upwards, a morning peak that consistently starts an hour earlier. The art lies in limiting false alerts, because a system that cries out too often gets ignored. That requires handling exceptions such as public holidays and holiday periods intelligently, thresholds that move with the spread of normal behaviour, and alerts that are bundled and prioritised rather than fired off individually. The goal is not as many signals as possible, but as few missed genuine deviations as possible per false alarm.
- Fixed thresholds are robust but coarse; pattern detection also catches creeping shifts.
- Expectation profiles correct for daily, weekly and seasonal rhythm per meter.
- False alerts are the biggest enemy: exceptions, adaptive thresholds and bundling keep the system credible.
- The yardstick is the ratio between missed genuine deviations and false alarms.
From signal to saving: the practice
In practice, anomaly detection almost always starts with the same first surprise: the picture of normal itself. Merely building the expectation patterns exposes what runs structurally outside opening hours, and that first picture often delivers the quickest win. After that, the value shifts to continuity: the detection quietly guards every meter, and the organisation gets used to a deviation being an alert instead of a surprise on the statement. The alerts themselves improve along the way, because every deviation that is followed up sharpens the picture of normal. Within COMCAM, this monitoring is part of the analysis: signals about abnormal consumption, night-time consumption and data quality come together in the same view as the load profile, so that interpretation and detection reinforce one another and every alert sits directly in context.
- The first win often already sits in the picture of normal: what runs structurally outside opening hours?
- After that, continuity becomes the value: every meter guarded, every day anew.
- Followed-up alerts sharpen the picture of normal ever further.
- Detection and consumption analysis reinforce one another: every signal sits directly in context.
Curious what this looks like with your own data?
In a no-obligation call, a specialist looks at your meters, sites and energy questions with you. Response within one business day.
Where this adds value directly
Healthcare
Guard continuous consumption and installations without extra workload for the team.
Hospitality
Receive a signal on abnormal night-time and cooling consumption, per branch.
Facility
Have time schedules, clocks and installation behaviour guarded automatically.
Frequently asked questions
Practical answers to common questions about this topic.
Question not covered here?
Ask a specialist directly. You will get a response within one business day.
Ask your questionWhat exactly is an anomaly in energy data?
Any consumption that does not fit the expected pattern of that meter at that moment: an installation running outside opening hours, a base load creeping upwards, a shifted time schedule, but also meter data containing gaps or implausible values.
Won't I be flooded with alerts?
No, that is precisely the core of good detection. Through expectation patterns per meter, exceptions for public holidays for example, and the bundling of signals, only what genuinely deserves attention remains. A system that cries out too often gets ignored; the design is built around avoiding that.
Does anomaly detection also work without extra sensors?
Yes. The detection works on the quarter-hourly data from your existing smart meters. The finer-grained the data, the sharper the detection, but no extra hardware is needed to start.
What happens after an alert?
Every alert shows the pattern, the starting moment and the context of the deviation, so the cause can be searched for in a targeted way. In practice the next step is usually small: correcting a time schedule, inspecting an installation or reporting a meter fault.
What is the core of energy anomaly detection?
Anomaly detection is the automatic recognition of consumption that does not match a meter's normal pattern: an installation that keeps running, a clock set wrong, a meter reporting strange values. Because no one can review every chart every day, the detection does the watching and you only receive a signal when something genuinely deviates.
What data do I need for energy anomaly detection?
Start with quarter-hour meter data, invoices, contract data and site characteristics. That makes energy anomaly detection concrete, comparable and easier to follow up, rather than just a separate report.
When does this topic become relevant for my organisation?
As soon as it touches costs, grid capacity, reporting or daily operations. Also consider related themes such as abnormal energy consumption, energy consumption alerts and deviations in energy data.
Series · The analyses in depth
All 4 articlesKeep reading
Useful next steps
Curious what this means for your organisation?
Book a no-obligation call and discover it with your own energy data.
