Meter Reads - Machine Learning Validation

Machine Learning normalcy checks validate meter readings by processing incoming initial measurement data using Artificial Intelligence and Machine Learning neural network models to recognize "normal" versus "anomalous" measurement data. Measurement data passing the normalcy check moves to the Final Measurement table for billing while anomalous data gets processed by VEE rules. In this release, the normalcy checks validate residential customer data for "anomalous measurement data".

Implementations using traditional algorithm-based VEE rules often set the tolerance of the rules either too high or too low, resulting to a large number of To Do entries or truck roll requests that drives up customer costs. Machine learning helps alleviate such problems by providing more accurate prediction of anomalies through its "Deep Learning Model", which is trained on millions of data points to recognize normal versus anomalous usage problems. The model processes the data using the incoming measurement data curve and weather data to calculate a predicted curve based and generates an "Anomaly Score" for each measurement by comparing the incoming measuring data curve to the calculated predicted curve.

If any of the Anomaly Scores for any of the measurement values exceeds the "Anomaly Score Threshold":

  1. The initial measurement is deemed "anomalous".
  2. An exception is created.
  3. The initial measurement data proceeds through the normal lifecycle, including VEE processing and creation of final measurements.

If none of the Anomaly Scores for any of the measurement values exceeds the Anomaly Score Threshold, the initial measurement moves directly to the creation of final measurements.

This optimizes performance and reduces user exceptions.

Steps to Enable

To enable this feature, refer to the Configuring Machine Learning Anomaly Scoring section of Administrative User Guide for more information.

Tips And Considerations

This feature will be available to select early adopter SaaS customers in the United States. The current model is trained for residential electric AMI accounts with meters programmed for 15 and 60 minute interval data as well as daily scalar data.