Anomaly detection in PV fleet data via interpretable machine learning

We present a data-driven methodology for inferring anomalies in photovoltaic (PV) power generation signals, based on a set of nearby generators. Our approach models correlations between nearby PV generators over a sliding time window. Our novel, white-box machine learning approach has three important components: (1) a time-dependent, multi-periodic quantile model of the individual power signals, which is used for marginal normalization of observed signals, (2) a linear regression model that predicts the normalized power output of a system at a given time based on a "neighborhood" of measurements across systems and time, and (3) a traditional binary classification algorithm. Model fitting is achieved via convex optimization, which provides globally optimal solutions in polynomial time.