Solar Radiation, Photovoltaic (PV), Forecasting, Machine Learning
Summary / Abstract:
In this paper we present recent advances at CSEM in multi-site PV forecasting that rely entirely on past production data. The presented methods take a graph signal processing perspective and exploit the intuition that a dense network of PV stations provides enough information for short-term prediction horizons (up to six hours ahead) with a high temporal and spatial resolution. The proposed forecasting models are based on recent advances in graph machine learning to capture the spatio-temporal dependencies of the PV production data. The proposed methods were evaluated in two data sets for an entire year: 1) production data from 304 real PV systems, and 2) simulated production of 1000 PV systems, in both cases distributed over Switzerland. The best of the proposed methods achieves a daytime NRMSE (over all sites) of 15.5% for the real data set and 12% for the synthetic data set for 6 hours ahead prediction and 7.5% for the real data set and 3.5% for the synthetic data set for 15 minutes ahead prediction. Furthermore, the proposed nonlinear methods outperform state-of-the-art methods with NWP as inputs on time horizons between 0 h and 5 h.