Company case
Google DeepMind: Predictive ML
What the published sources confirm
In 2016 DeepMind and Google described neural networks trained on historical data-centre sensor data to predict energy efficiency (PUE) and future temperature and pressure, used to recommend cooling actions within safety constraints.
Business problem and AI concept
- Problem
- Cooling systems are complex and interacting; fixed rules leave energy savings unrealised.
- Concept
- Supervised prediction from time-series sensors, used to rank safe control actions.
Reported results
The 2016 post reports up to a 40% reduction in energy used for cooling, equivalent to a 15% reduction in overall PUE overhead.
Google / DeepMind blog, 20 Jul 2016. Caveat: Historical result from specific facilities; not a general guarantee.
Suggested workflow
Written for learning, not the company's system. This flow shows how you could build something similar. It is not a description of Google DeepMind's private architecture.
Step 1: Gather historical sensor readings and operating settings
Step 2: Train models to predict efficiency and near-future conditions
Step 3: Compare against the existing rule-based baseline
Step 4: Generate candidate settings and predict their outcome
Step 5: Reject any candidate that breaks hard safety limits
Step 6: Operators review and monitor results
Suggested technologies and alternatives
| Layer | Suggested | Alternatives |
|---|---|---|
| Modelling | scikit-learn / gradient boosting, then PyTorch | XGBoost, LightGBM |
| Time series | pandas feature windows | Feature store |
| Tracking | MLflow | Weights & Biases |
| Dashboard | Grafana | ASP.NET Core + Blazor |
Small example
- Input
- Last 60 minutes of temperatures, pump speeds and IT load.
- Output
- Predicted PUE for three candidate settings; the best safe one recommended, with predicted max temperature.
Rebuild it yourself
Build a cooling-energy predictor on synthetic sensor data with a safety-constrained recommendation simulator.
Definition of done
- Time-based split (train on earlier weeks, test on later)
- Beats a persistence baseline
- Never recommends settings outside limits in tests
Security and privacy risks
- Unsafe control actions
- Distribution shift after hardware changes
- Leakage from future data in features
Sources
Listed by title and publisher. This guide contains no external links; search for the title if you want to read the original.
- DeepMind AI reduces Google data centre cooling bill by 40%. Google DeepMind, 20 Jul 2016. Company research blog.
- Google unleashes DeepMind on energy-hungry data center. TechCrunch, 20 Jul 2016. Independent reporting.