Anomaly DetectionMarch 1st, 2017 by Matthew Ross
Part 1: Avoiding False Positives
In our last blog we discussed the importance of Data Quality. Now we dive deeper into the various methods of identifying anomalies while reducing the number of False Positives.
Poor anomaly detection results in datasets full of errors – confusing systems and analysts alike
Implications of false positives
When it comes to finding outliers and anomalies, it’s better to be safe than sorry, right? Not exactly…
Energy usage data, especially from smart meters, is “noisy” and contains a variety of sudden changes that would trigger most simple anomaly detection algorithms. These “false positives” can directly impact revenue in a variety of ways.
– Auto correction of these “outliers” can alter the load shape and impact product selections by losing the natural volatility of a customer’s load
– Unnecessary changes lead to wrong usage totals affecting calculations and predictions
– During forecasting, this can mean a less competitive offer or loss of margin
– For billing, incorrect bills can frustrate customers, increase customer service calls and potentially miss revenue that was actually valid
On a practical note, too many errors can choke up (billing) systems, stopping the “happy path” right in its tracks. The next step is usually manual attention which is costly and time consuming. Ultimately, this time translates to unbilled customers, missed opportunity and revenue sitting on the table.
Uncovering anomalies and nothing more
The key to accurate anomaly detection is to detect the various changes in load shape allowing the right algorithms and parameters to be applied at the correct time. Some algorithms can adjust automatically to changes but unfortunately there is no “one size fits all” anomaly detection method. Instead, we utilize a variety of automated strategies that are optimized for different load shapes.
This approach can be enhanced further by normalizing the data, taking into account weather, seasonality, ramps up/down and step changes for example. The result is a more “stationary” load shape, making outliers much easier to catch and correct.
Learn the truth by adding context
Load detection and normalization makes anomaly detection much more effective, but you’re still not safe from false positives until you add context to the dataset. Let’s use the chart below as an example.
At face level, it looks like there are three “dips” in the dataset that should be flagged as outliers and corrected. However, zooming in, we see a rather obvious pattern. The dips occur on three U.S. holidays: Memorial Day, Independence Day and Labor Day. Energyworx anomaly detection methods account for holidays, thus reducing this common source of false positives.
The same methods also contain similar logic to capture severe weather, outages and local events that may impact usage – creating outliers – but are actually “true” and should be billed and forecasted as such.
Are false positives holding you back? Contact us to experience the latest techniques for anomaly detection.
Still can’t get enough anomaly detection in your life? Then join us at Google Next in San Francisco to watch Founder & President, Edwin Poot, demo our anomaly detection methods live on stage!