学术文献库

Satellite-based solar irradiation forecasting is useful for short-term intra-day time horizons, outperforming numerical weather predictions up to 3-4 hours ahead. The main techniques for solar satellite forecast are based on sophisticated cloud motion estimates from geostationary satellite images. This work explores the use of satellite information in a simpler way, namely spatial averages that require almost no preprocessing. Adaptive auto-regressive models are used to assess the impact of this information on the forecasting performance. A complete analysis regarding model selection, the satellite averaging window size and the inclusion of satellite past measurements is provided. It is shown that: (i) satellite spatial averages are useful inputs and the averaging window size is an important parameter, (ii) satellite lags are of limited utility and spatial averages are more useful than weighted time averages, and (iii) there is no value in fine-tuning the orders of auto-regressive models for each time horizon, as the same performance can be obtained by using a fixed well-selected order. These ideas are tested for a region with intermediate solar variability, and the models succeed to outperform a proposed optimal smart persistence, used here as an exigent performance benchmark.