Improving Photovoltaic Energy Yield Forecasting Accuracy Using Neural Networks

Abstract

The ongoing growth of renewable energy installations is slowly shifting the energy mix for grids worldwide. Photovoltaic (PV) installations in particular continuously increase their share, and according to IEA (International Energy Agency) could account for around 11% of global electricity in the near future.One of the biggest challenges is the fluctuation of the provided solar energy to the grid due to the stochastic nature of meteorological phenomena, in particular solar radiation. The ability to forecast such changes is becoming more and more relevant, for both interconnected and isolated grids. Especially in short horizons (< 30 minutes for single camera installations), forecasting of future energy output will help regulating the grid accordingly and decrease fuel consumption for the stand-by generators. For short-term forecasting of available solar radiation in an area of interest, usage of a sky-imaging system is appropriate. In this thesis, extracted sky- image information, provided by temporal-spatial resolution sky imaging forecasts for a 15 minute horizon, with measurements for PV output and solar irradiance for the zero horizon are used as inputs to NARX neural networks. The final outcome of this coupling is a forecast of the actual energy yield of the installation for horizons of typically 15 minutes. The extracted image data and all other measurements were provided by the University of Oldenburg,Germany. In addition, a comparison between the implemented NARX model and the University of Oldenburg's state-of-art short-term irradiance forecasting model is presented.