Cenovus Energy

Operational data was arriving from a mix of sensor systems, field readouts, and raw logs across several Cenovus facilities. The data was usable in principle and broken in practice. Sampling cadence drifted between sites. Units were inconsistent. Whole windows were missing.

Anything sitting on top of that layer (forecasting, predictive maintenance, anomaly detection) was going to inherit the noise. Before any modelling could be trusted, the data layer had to be made boring.

We built a secure, end-to-end pipeline covering ingestion, normalisation, feature engineering, training, evaluation, deployment, and reporting. Python ETL handled the unification of cadences, units, and missingness, and produced features tuned for the downstream forecasting and maintenance tasks.

For the modelling layer, we evaluated gradient-boosted trees and recurrent neural networks across cross-validation folds, on a per-task basis. Each task got the model class that earned it on validation, not the one we liked on paper. Final models were containerised and deployed to cloud.

Daily dashboards now surface forecasts and predictive maintenance flags to facility engineers, replacing manual rollups that used to consume real headcount. The reporting layer runs on schedule, the model layer runs on demand, and the data layer is no longer the bottleneck.

The team that used to spend a week assembling a report is now reviewing one.