Recommendation System for Meter Replacement

An ML-driven analytics system using Azure Databricks and PySpark to prioritise water-meter replacement across a 10M+ device network, reaching 92% prediction accuracy.

Azure Databricks
PySpark
SQL
Machine Learning
DataOps
platform.internal/ml/meter-replacement

Device network

10M+

Prediction accuracy

92%

Incremental billing

€8.7M

Model accuracy: training epochs

92% val. accuracy

Replacement priority queue

MTR-118823
score 0.94High risk
MTR-118824
score 0.91High risk
MTR-118825
score 0.78Medium risk
MTR-118826
score 0.71Medium risk

Illustrative interface with fictitious sample data, shown in place of confidential client screens.

The problem

With a 10M+ device network, there was no reliable way to know which meters most needed physical replacement: ageing units under-report usage, which quietly costs the utility in lost billing until someone notices.

What I built

A machine-learning model trained on consumption patterns, meter age and fault history that scores every device by replacement priority, running on the same Databricks/PySpark foundation as the meter-closing pipeline so it consumes the already-clean reconciled data directly.

The model reaches 92% prediction accuracy on held-out data, and its output feeds a prioritised replacement queue that field teams work down instead of replacing meters on a fixed schedule regardless of actual condition.

Impact

Better-targeted meter replacement generated roughly €8.7M in annual incremental billing by catching under-reporting meters that a fixed replacement schedule would have missed.