Addressing data scarcity in local photovoltaic datasets: a GAN-based workflow

Digital twins are increasingly being implemented in smart cities, where they are designed to provide 3D digital representations for near-real time interactivity and status feedback of the twinned physical assets. The ability to provide accurate predictions in what-if scenarios plays a key role in driving the advances in resources optimization and risk mitigation. To this end, prediction models require large datasets to train, and such datasets are usually scarce at local level. In this paper, we propose a generative AI-based workflow to infer domestic photovoltaic energy production data, based on a process that uses Generative Adversarial Networks to generate an enriched meteorological dataset. The generated meteorological data captures essential statistical properties of real data that can be later used to infer realistic energy production data points. The resulting output dataset is validated against a real photovoltaic system, highlighting the ability to deliver high-fidelity time series out of scarce input datasets.