A new sky imager based global irradiance forecasting model with analyses of cirrus situations
High resolution irradiance forecasts based on sky imagers are valuable for applications that require short term decisions based on ramps of solar irradiance. Here, we present our sky imager based forecasting algorithm, using images of a low cost surveillance camera. Model development and evaluation is done separately for the different steps in sky imager forecasting, starting with cloud detection, followed by estimation and extrapolation of cloud movement, and finally deriving irradiance forecasts from the predicted cloud images. We distinguish between clear and cloudy conditions and especially evaluate the effect of cirrus situations on the different forecasting steps. To create binary cloud masks, we adapted a pixel value based cloud decision algorithm using a set of manually classified pixels. In an independent validation dataset 90,3% of the pixels are classified correctly. For the circumsolar region, where cloud decision is known to be especially challenging, we introduce a correction procedure using real time irradiance measurements and object recognition methods. Applying this method we can significantly reduce the cloud detection in the circumsolar region and increase the forecast skill of the cloud decision forecast. The development of the irradiance algorithm is a special focus of this paper. Real-time irradiance measurements and cloud decision information are used as input to our irradiance algorithm. The algorithm is developed and optimized using cloud decision information derived from measurements instead of sky imager cloud decision forecasts in order to exclude the influence of errors in cloud decision and cloud motion methods for model development. Afterwards, the irradiance algorithm is applied to sky imager based cloud decision forecast. Even though we start with a binary cloud decision algorithm, the distribution of the clear sky index from our forecasts is in very good agreement with the distribution of the measurements. In a validation dataset of 46 days, we receive a positive forecast skill for all forecast horizons larger than 1 min. We also apply our forecast chain to a dataset of two month from an independent measurement station resulting in a comparable forecasting performance.