Context
Many regions continue to rely on traditional meters or smart meters that communicate infrequently, creating a need to estimate customer electricity consumption to enable timely billing by energy retailers.
Challenge
Current heuristic, benchmark-based estimation approaches have a stated performance ceiling, leading to bill shock, customer churn, increased cost to serve, and revenue leakage worth millions for energy retailers.
AI Solution
Kaluza applies ML to generate consumption estimates by learning from historical meter reads across a large cohort of over 1 million traditional and smart-meter customers , with delivery through standard integration using APIs (Application Programming Interfaces), SDKs (Software Development Kits), or connectors and integration into meter read, billing, and CRM (Customer Relationship Management) systems.
Impact & evidence
Success and impact are measured using metrics including MAPE (mean absolute percentage error) reduction in estimates, cancel and rebill rate, contact rate, churn rate, and revenue leakage reduction across cohorts with model deployment compared to without.
Implementation / Adoption
Replication requires access to large-scale anonymised historical consumption datasets, anonymised contact data, integration with meter read, billing and CRM systems, and ML Operations (MLOps, machine learning operations) infrastructure, with optional weather and property datasets. Prerequisites also include significant client-side data preparation. Success of the solution would depend on the scale of training data (across diverse geographies and property types), continuous and iterative model retraining as new consumption data becomes available, integration with smart meter feeds for validation and ability to control rollout across cohorts to measure and validate impact. Kaluza implemented the solution through an initial proof of concept (POC) phase lasting a quarter, operated in shadow mode alongside an existing heuristic approach to validate parity before exposing outputs to real customers. This is scheduled to be followed by a multi‑month cohort expansion, with gradual rollout segmented by tariff type, meter reading frequency, and customer tenure.
Data
Customer-provided data includes historical consumption data and meter reading data (half-hourly, daily, or manual submissions), anonymised contact data, and integrations with meter read, billing, and CRM systems, with optional property data such as EPC (Energy Performance Certificate) ratings and land registry information and optional weather data via Weather API integration.
Future enhancements
Depending on model success, the model and pipeline may be iterated to additional use cases, including anomaly detection to flag meter reads that are out of reasonable bounds, improved direct debit projections, and identifying accounts for proactive outreach for meter reads.