Anomaly detection in consumption data

The task

The project analysed heating, water and electricity consumption in buildings. Monthly consumption values need to be understood in the context of the device, the unit it serves and historical patterns. Operational review needs to distinguish unusually high or low values from expected variation.

The task was to assess consumption data systematically and make relevant findings available for further review. Anomaly detection adds a model-based assessment to the operational review process.

A reliable data foundation

The processing combines consumption values with device information, the associated unit and available historical readings. Clear relationships and a consistent monthly time reference form the basis for assessment.

Recorded zero consumption remains a valid observation. Missing values and separately generated estimates are treated as distinct data states.

• Align monthly values with their time period and business context
• Account for device type, usage context and available history
• Keep observed consumption distinguishable from estimates

Assess deviations with machine learning

The solution calculates expected consumption and compares it with the recorded value. The deviation and a calibrated threshold produce a score that helps prioritise findings.

Different model variants are used depending on the available history. Unusually high and low consumption, as well as negative values, are reported as distinct finding types.

• Compare observed and expected values
• Provide the deviation, score and severity
• Identify the model and model version used

Integration into existing applications

An API connects anomaly detection to the data processing workflow. The application receives structured results containing expected values, deviations, severity and model information.

Findings can be retrieved by building, unit, device and severity. This allows the responsible teams to handle relevant cases in their existing processes.

Production use and operational value

The solution is in production. It supports the review of monthly consumption data through consistent assessment and prioritisation of unusual cases.

An anomaly is a prompt for further review. Its cause needs to be assessed in context. The model alone does not determine whether a deviation reflects a change in usage, a data issue or a technical fault.