Ultra-short-term distributed PV power forecasting for virtual power plant considering data-scarce scenarios
Yuqing Wang, Wenjie Fu, Junlong Wang, Zhao Zhen, Fei Wang
Abstract
Accurate forecasting of distributed photovoltaic (DPV) power plays a crucial role in enabling a virtual power plant (VPP) to grasp its internal DPV characteristics and support the development of optimized internal scheduling strategies. However, majority of the existing forecasting methods are developed rely on sufficient power data samples. Power data scarcity caused by site construction, upgrading or limitation of data share in VPP will lead to poor forecasting accuracy of DPV. To conquer this limitation, the adversarial Graph Neural Network based ultra-short-term DPV power forecasting method for VPP considering the scenarios of data-scarcity is proposed. In this method, the forecasting model is developed on the graph structure data of multi-sites, which is constructed from data-rich region and data-scarce region, respectively. Specifically, graph representation learning and adversarial domain adaptation techniques are utilized to learn domain-invariant graph embedding features, which are further incorporated into the spatiotemporal modeling of distributed photovoltaic power data. As a result, the power forecasting accuracy of sites within the data-scarce region is improved by transferring “forecasting-related knowledge” from distributed photovoltaic sites of data-rich region with different data distributions and graph structures. The simulation experiment proves that the forecasting accuracy can be further improved by integrating the domain-invariant features into the power forecasting of distributed photovoltaic in the data-scarce region via a real power dataset.