One-Year Real-Time Operational Prediction Intervals for Direct Normal Irradiance

Abstract:

This work describes an algorithm to generate intra-hour prediction intervals (PIs) for the highly-variable direct normal irradiance, which is the energy source for the concentrated solar power technologies. The prediction intervals are generated using a Multi-layer Stochastic-Learning Model (MSLM), which is developed based on methods such as: sky imaging techniques, support vector machine and artificial neural network. The MSLM is trained using one year of co-located, high-quality irradiance and sky image recording in Folsom, California. In addition to being validated with historical data, the algorithm has been generating operational PI forecasts in real-time for that observatory since July 1^st 2014. In the real-time scenario, without re-training or significant maintenance, the hybrid model consistently provides valid PI (PICP > 92%) and outperforms the reference persistence model (PICP ~ 85%) regardless of weather condition. This work has great impact in the field of solar energy to potentially facilitate the level of solar penetration in the grid with significantly reduced integration costs.