Permutation test and bootstrap methods for unsupervised detection and estimation of behind-the-meter photovoltaic generation

The penetration of unmonitored behind-the-meter photovoltaic systems has increased rapidly over the past decade. While unsupervised methods have been proposed to estimate behind-the-meter photovoltaic generation, their performance is significantly affected by hyperparameters. Existing methods for choosing hyperparameters require labelled data, which are unavailable to system operators. In this paper, a data-driven method is proposed to generate labelled data to optimize the hyperparameters for unsupervised estimation of behind-the-meter photovoltaic generation. First, a permutation-test based method is developed to detect photovoltaic installation in an unsupervised way. Second, consumers without PV are combined with limited monitored photovoltaic sites to simulate the original system through the bootstrap method. The simulated system provides labelled data for hyperparameter optimization. Finally, the near-optimal hyperparameters are searched on the simulated system and applied to the original system's estimation. Through experiments on a smart meter dataset, the effectiveness of the proposed methodology is verified.

Automated summary

This paper proposes a data-driven method to generate labelled data for optimizing hyperparameters in unsupervised estimation of behind-the-meter photovoltaic generation. By using permutation tests and the bootstrap method, a simulated system is created to provide labelled data for hyperparameter optimization. Experiments on a smart meter dataset verify the effectiveness of the proposed methodology.