A Hybrid Numerical Method and Comparative Study for Solar Irradiance Forecasting

Renewable energy sources like solar energy and wind power, which are not available on-demand must be stored or used instantly. The solar irradiance is strongly dependent on unpredictable atmospheric conditions. Thus, the performance and the accuracy of solar forecasting model is highly influenced by the complex and stochastic atmospheric conditions. The forecasting challenges of solar energy augment with increasing deployment of photovoltaic (PV) system into electrical grids. As the global horizontal irradiance (GHI) is the most important key input in PV power prediction systems, increasing effort is spent on research on forecasting of GHI as a basis for corresponding power forecasts.
The proposed approach in this paper is based on forecast of the global numerical weather prediction (NWP) model Global Forecast System (GFS). The proposed model uses day-type classification benchmark based on hourly vectors generated using clear sky index and hybrid numerical formulation. A novel GHI prediction system using input data from an GFS and two levels ensemble is presented and evaluated by providing day ahead forecasts with an hourly time resolution. The methodology applies various machine learning (ML) and recurrent neural network (RNN) models along with a parametric model as the basis to build an ensemble model. The focus of this paper is the comparison of the capability and performance of developed models and build a hybrid ensemble model for producing reliable day-ahead forecast horizon and an hourly resolution of GHI. The detailed study and analysis of probabilistic time series forecast is presented. The reliability and the sharpness of the forecast
models is evaluated by using various skill scores of probabilistic forecasts.
Results show that the hourly irradiance classification approach reduces the forecasting error by a considerable margin and the optimized hybrid ensemble model outperforms all other models. These findings also emphasize that all
the models perform better than GFS.