A Novel Distributed PV Data Virtual Collection With Continuous-Binary Denoising Auto-Encoders

The massive deployment of distributed photovoltaic (PV) installed on residential rooftops makes PV data collection and user privacy protection more prominent. To reduce number of devices and improve the ability to combat data loss in complex terrain environment, this paper proposes a deep learning-based virtual collection method. The proposed method employs a continuous-binary denoising auto-encoder (CB-DAE) to analyze the static characteristics of a multi-photovoltaic system (MPVS) and select reference PV (RPV). A data anomaly detector is constructed using ensemble learning to assist CB-DAE in data blind cleaning and decouple the dynamic characteristics from MPVS. To train the CB-DAE, a joint optimization model of data cleaning - RPV selection - the virtual collection is established, and the sparsity constraint of
$\mathbf {L}_{\mathbf {0}}$
-norm is expressed as
$\mathbf {L}_{\mathbf {1}}$
-norm constraint. Furthermore, an improved split Bregman using
$\mathbf {L}_{\mathbf {p}}$
-quasinorm approach is proposed to accelerate the solution. The proposed method is validated using data from Jan. 1 to Apr. 15, 2022, of 110 PVs from Badong County, Hubei Province, China. The impact of the scale of MPVS and the number of RPV on accuracy is also explored. Experimental results demonstrate the effectiveness and superiority of the proposed method.