AnyCast: Efficient Graph Learning for Large-Scale PV Power Forecasting with Extreme Missing Data

초록

Recently, the massive penetration of renewable energy has been a major concern in power systems for stable operation. To minimize the negative impact of renewable energy in power systems, accurate, scalable, and stable forecasting for renewable energy can be considered. In this paper, we propose an efficient graph learning method called AnyCast for probabilistic multi-site photovoltaic (PV) power forecasting, especially focused on the scalability of a large-scale problem and robustness under missing data. By leveraging the graph neural network (GNN), the proposed random coarse graph attention and probabilistic spatio-temporal learning perform effectively for large-scale PV sites in terms of forecasting accuracy and model training complexity. Simultaneously, our method adaptively imputes the missing PV data, ensuring stable forecasting even with the extreme missing data. Simulation results show that AnyCast achieves 7-12% improvement of mean absolute error and 79-82% improvement of total GPU usage in model training complexity for over 1600 PV sites and achieves 4-36% improvement of mean absolute error for 364 PV sites under the 90% missing data.

키워드