PUL-DERs: Distributed Energy Resources Prediction Algorithm Via Pseudo-Labeling Unsupervised Learning

The prediction of Distributed Energy Resources (DERs) has become increasingly vital for future power networks to achieve net-zero carbon emissions. However, existing methods for this problem rely heavily on supervised learning with extensive labeled data for training and high computational costs, greatly restricting the applicability of traditional algorithms in DERs monitoring. To address these, we propose PUL-DERs, predicting the characteristic electrical profiles produced by electric vehicles (EV) and photovoltaic (PV) systems in the distribution network utilizing pseudo-labeling unsupervised learning. The optimized K-medoids clustering algorithm is applied to EV and PV prediction, incorporating cosine similarity as the distance metric. Our approach concatenates synthetically generated pseudo-labeling data that closely mimics real-world DERs distributions with smart meter data, thus enriching the dataset while maintaining relevance to actual values. Experimental results on the datasets demonstrate the advantage of PUL-DER's prediction accuracy, achieving 95.01% for EV usage and 98.15% for PV generation.