How it works

The function of the Meter Bypass application is to detect meter bypass theft conditions. The Meter Bypass application is able to identify ongoing bypasses as well as new ones.

  1. Synthesis. Per-second data sampling of various features aggregated at the meter and sent to the back office.

  2. Inference. A machine learning model uses DI data to make scored recommendations daily.

  3. Action. The utility revenue assurance team then downloads recommendations through manual or automated methods.

Figure 1  Meter Bypass overview diagram

The statistical ranges provided in the agent configuration allow you to test the Meter Bypass machine learning algorithm when a bypass is added or is present. The range is derived from positive examples seen at other utilities. There will be some differences in applicability to your utility's territory given the locations and differences in weather conditions. If there are too many false positives, Itron will fine-tune the configuration model based on your own data (especially any findings of bypass in the field where data has been collected from Itron meters). Implementing a conservation voltage reduction (CVR) strategy via your advanced distribution management system (ADMS) can make the model even more effective.

The Meter Bypass machine learning model will identify existing bypass events, even if there is no decrease in the measured value.

If the voltage of a bypassed meter falls outside the configured range, it may still be identified, though this is less likely.

If the outcome is removed and reinstalled, or if the outcome is upgraded to a newer version, any bypass events detected before the removal, reinstallation, or upgrade will still be detected.

A minimum of seven days of historical data is required for the outcome to establish a baseline. Itron recommends using 14 days of historical data to establish your baseline. The machine learning model needs seven days of historical data to start running and seven more days to start flagging and identifying whether events are occurring daily or just one time.

Machine learning process

The model creation is based on physical and statistical insights. It is updated regularly.

Figure 2  Machine learning process diagram