Time Series Forecasting & Anomaly Detection in Radio Networks
summary
Network telemetry arrives at scale — roughly 3,000 KPIs per cell per hour, 3k–60k cells per country and ~150 KPIs per call — and forecasting feeds an adaptive power scheme that aims to increase energy efficiency 80% by 2030. Final production uses XGBoost with two regressors to produce 7-day hourly sector forecasts (weekly runs) plus three apps (training, forecasting, diagnostics); anomaly detection sets IQR thresholds over 60-day windows, computes hourly/daily/frame scores with linear/sqrt/log transforms, and is rolled out across multiple technologies and opcos.
AI-written summary · about
chapters (8)
- 0:00Time series forecasting intro
- 3:00Telecom time series challenges
- 6:36ML for time series forecasting
- 9:45Energy-saving forecasting use case
- 13:26Anomaly detection intro
- 17:28Anomaly detection model core
- 24:19Calculating anomaly scores
- 30:09Summary and future steps
related talks
44mAnomaly Detection in Telecom with Spark
30mRealtime Time Series Anomaly Detection in Production
23mNetwork Anomaly Event Prediction and Optimal Resource Control in Cloud Native Network Functions
40mDeep Learning and Time Series Forecasting for Smarter Energy
32mPredicting Grid Faults Before They Happen: AI for Weather-Driven Grid Resilience
17mAI Weather Forecasting
45mForecasting at Scale
27mProject Zouwu: Scalable AutoML for Telco Time Series Analysis using Ray and Analytics Zoo