Photovoltaic Hot-Spot Automated Diagnostics from Partial Noisy Sensor Data: A Physics based Approach using Splines and Stochastics

Shahar Ben-Menahem and Abraham K. Ishihara


photovoltaics, partial shading, hot-spots, Bayesian


We use an integrated optical, electrical and thermal plant physics model of a photovoltaic (PV) array, in conjunction with machine learning Sys ID- and diagnostics algorithms, to analyze partial, noisy sensor data at various granularity levels, acquired from our rooftop array, and to detect, characterize and predict the future evolution of hot-spots, caused by partial shading and other mismatches. Data are taken from electrical (current and voltage), temperature, irradiance and wind sensors. Sensor data from electric load sweeps at two different timescales, as well as downloaded online astronomical and weather data, are fed to our suite of algorithms. Baseline system ID is performed, and our (Bayesian MAP, path-integral based) diagnostics algorithm is used to detect artificial partial cell shading (when present) and prognosticate the hot spot which it may cause. Our machine learning suite of algorithms use a simplified, splines-based version of the integrated plant model (for rapid parameter estimation and hypothesis evaluation). These algorithms were tested on simulated and partially simulated data (generated via our integrated plant model), before being applied to analyze empirical data.

Important Links:

Go Back