The challenge
The organisation already had utility monitoring, Power BI dashboards, on-premises SQL Server data sources, and real-time readings across energy, water, and gas. The challenge was not a lack of data, but the difficulty of turning sensor readings into meaningful, prioritised alerts. Static thresholds created alert noise, high consumption was not always abnormal when production was high, and gradual changes or out-of-hours consumption could be missed without production, shift, temperature, downtime, and cost context.
The approach
Scoped a focused solution around one site initially, with a defined set of energy sensors, and twelve months of historical readings. The solution used Microsoft Fabric as the governed data and analytics foundation, ingesting sensor readings, metadata, production context, operational calendar, temperature history, and tariff data where available. Data was organised through a medallion architecture, with Bronze raw data, Silver cleaned and aligned datasets, and Gold expected-vs-actual consumption and alert-ready datasets. An XGBoost-based model was proposed to learn expected energy consumption from context such as time, temperature, production volume, shifts, machine status, downtime, maintenance windows, and historical sensor behaviour. Power BI provided operational insight, while Fabric Data Activator triggered alerts for prioritised exceptions.
The outcome
Created a clear path from static utility reporting to AI-enabled operational intelligence. The solution design established dynamic energy baselines, anomaly scoring, actual-versus-expected consumption analysis, severity classification, cost impact estimation where tariff data was available, and Power BI-based consumption insights. The approach reduced dependency on fixed threshold alerts and provided a scalable foundation for future utility optimisation and AI use cases across manufacturing operations.
