What is non-technical loss and how does AI help detect it?

Non-technical loss (NTL) refers to energy billed for but not received as revenue, caused by meter tampering, billing errors, or theft. AI models compare AMI interval consumption data against billing records in the CIS to flag accounts where patterns suggest NTL. The CIS remains the record of billing; field verification confirms the flag.

Which analytics tools are purpose-built for utility customer energy data?

Oracle Opower, Bidgely, and Uplight are the three most widely deployed customer energy analytics platforms. All three ingest AMI interval data and return personalized insights or demand flexibility signals back through the CIS or customer portal.

Does AI improve demand forecasting compared to traditional statistical models?

For short-term load forecasting, machine learning models that incorporate weather variables, AMI load shapes, and economic indicators tend to outperform classical time-series methods, particularly during weather extremes. For long-range planning, the accuracy difference is smaller and governance of model assumptions matters more than model type.

AI Data Analytics in Utilities: Loss, Demand, and Asset Outcomes

Utility analytics conversations often stay generic: “AI will improve efficiency” without specifying which efficiency metric, whose data, and what the workflow looks like when a result is produced. This article takes three concrete outcome areas and traces the data and system connections behind each.

Non-Technical Loss Detection

Non-technical loss (NTL) sits at the intersection of AMI data and CIS billing records. The signal is a gap: a meter showing reduced or erratic consumption relative to its historical pattern, while the billing system shows no legitimate explanation such as a rate change, vacancy, or seasonal adjustment.

AI models trained on AMI interval data from Itron or Landis+Gyr meters can score every account in a service territory for NTL probability without requiring individual analyst review. The score is written back to the CIS as a work queue item. A billing investigator reviews accounts above a threshold, orders a field check, and the CIS records the outcome.

What AI does not do here is change the billing record autonomously. The CIS is the system of record for all account data; the analytics model is advisory. This matters for audit trail and regulatory compliance purposes.

Short-Term Load Forecasting and Demand Response

Accurate short-term load forecasting directly affects generation scheduling, capacity market positions, and the dispatch of demand response programs. Traditional regression models work reasonably well under normal conditions. They degrade when conditions depart from historical norms, which is precisely when forecast accuracy matters most.

Machine learning models that combine AMI load shapes, weather data, and calendar features handle non-linear interactions between variables more effectively. For distribution operations, feeder-level load forecasts built from AMI data give the ADMS better inputs for Volt/VAR optimization setpoints and predictive switching plans.

Oracle Opower and Bidgely both build customer-level consumption models from AMI data. Opower’s behavioral energy reports use those models to drive demand reduction. Bidgely’s disaggregation approach identifies which appliances are contributing to load, enabling targeted efficiency recommendations. Both platforms return insights through the customer portal or CIS, maintaining the utility as the interface with the customer.

The AI consumption optimization article covers the demand-side detail, and the smart grid and AI optimization post covers how forecast data feeds grid operations.

Asset Health Analytics

Distribution asset failure is expensive in two ways: emergency repair costs and the customer-minutes interrupted during an unplanned outage. Predictive asset analytics attempts to shift maintenance from calendar-based to condition-based, prioritizing the assets most likely to fail before the next planned cycle.

The data inputs vary by asset type:

Overhead lines and poles: storm damage history, inspection records, LiDAR point cloud data (for vegetation proximity), and age from the GIS and asset register in the ERP.
Substation transformers: oil analysis results, dissolved gas analysis trends, thermal model outputs from continuous monitoring, and loading history from SCADA.
Underground cables: insulation testing history, fault history by cable segment, and soil temperature data where available.

AI models on this data produce risk scores by asset. The output feeds the capital planning process and the annual maintenance schedule. A planner reviews the ranked list, adjusts for budget and crew constraints, and issues work orders through the ERP or work management system. The model recommends; the planner decides and the ERP records.

For utilities on SAP, this connects directly to the SAP IS-U plant maintenance integration and the broader utility billing ERP data environment. For Oracle environments, see the Oracle Utilities pillar for how Oracle’s asset management capabilities fit this picture.

The Common Thread Across All Three Areas

All three use cases share the same architectural principle: AI models read from the system of record, whether the CIS, AMI data management system, or ERP asset register, produce a scored or ranked output, and return it to a human workflow for confirmation. The data systems retain authority over billing, account status, and work order issuance.

Utilities that have not yet connected AMI data to the CIS analytics layer should prioritize that integration before building prediction models. Clean, reconciled interval data is the prerequisite for all three outcomes described here. Avansaber can help assess the readiness of your current data architecture.