An Unsupervised Similarity-based Method for Estimating Behind-the-Meter Solar Generation

Accurate knowledge of solar generation in power distribution systems provides great values to utilities for efficient and reliable distribution system operation. However, many solar PV resources are installed behind-the-meter (BTM), and as a result only the net load measurements are available to the utilities. In this paper, a high-performance method for disaggregating BTM solar generation traces from net load traces is developed. The algorithm takes the net load data measured by smart meters and other widely available environmental measurements (e.g., publicly monitored solar irradiance and temperature) as inputs, and disaggregates the net load traces into BTM solar generation and load traces. Notably, the proposed method does not rely on any separately metered data of BTM solar generation. Rather, in a fully unsupervised fashion, the proposed method effectively exploits the self-similarity and cross-customer similarity of customer loads to achieve accurate BTM solar disaggregation. The developed unsupervised method is evaluated on two real-world smart meter data sets collected from New York and Texas, and exhibits very high performance that closely approaches the ideal performance bound from supervised learning.