PV Monitoring Based on Linear Regression
summary
Proposes replacing moving averages with robust linear regression using least absolute deviations solved by iterative reweighted least squares, initializing weights from local slopes and falling back to prior parameters when iterations diverge; converts residuals to a scale-free z score via a Poisson model with normal approximation (Poisson → normal z-score) and tests against a probability-derived threshold. Handles PVs in the hundreds of thousands, observes a ~4× day/night scale difference, and reports >80% precision and >95% recall with a web-based open source labeling tool.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:00Difficulties
- 4:47Moving Average
- 6:39Linear Regression
- 10:18Initial weights
- 11:24Residue
- 12:27PV Number
- 16:15Results
- 16:45Evaluation
related talks
36mThe ML Monitoring Flow for Models Deployed to Production
32mPopmon: Population Shift Monitoring Made Easy
41mReliably forecasting time-series in real-time
30mRealtime Time Series Anomaly Detection in Production
30mPredictive Maintenance and Anomaly Detection for Wind Energy
12mModel Monitoring: Detecting and Analyzing Data Issues
01h 19mModel criticism with ArviZ
04mLightning Talk: Anomaly Detection - Robust Strategy