PVInsight: Data Driven Approaches for Analyzing PV System Performance and Reliability
Friday, October 9, 2020
2:15 PM – 2:25 PM EDT
With over 500 GW of installed worldwide photovoltaic (PV) capacity and 30-40% of new capacity coming from distributed rooftop PV systems, we are seeing an increasing volume of unlabeled, heterogeneous PV system performance data. These data often have missing or corrupted values and are lacking correlated reference data, making it difficult or impossible to construct standard performance metrics such as the Performance Index. Our work under the PVInsight project is to develop algorithms for extracting performance and reliability information from this class of PV system performance data, namely the large number of internet-connected solar PV inverters, collecting real power generation data on 1-minute to 1-hour time scales. We start from the assumption that we do not know anything more about the site than its (possibly less-than-ideal) power generation signal—no location, site model, or reference data. We borrow from signal processing and unsupervised machine learning to automate the analysis of these PV systems, allowing for fleet-scale performance and reliability analysis of distributed rooftop PV systems.
In this talk, Bennet will discuss three aspects of this work: 1. Data preprocessing and filtering 2. Data-driven clear sky modeling 3. Long-term system degradation estimation
He will demonstrate the proposed algorithms on real-world PV system data sets and discuss the software implementation. Their code is fully open source and is available on GitHub, PyPI, and Anaconda under the names solar-data-tools and statistical-clear-sky. They believe strongly in the importance of developing approaches to data science that are repeatable and reusable, and so we make all our code open and free to other researchers, throughout the development process.