Towards Improving Solar Irradiance Forecasts with Methods from Computer Vision

Modern economies turn towards renewable energy to lower the dependency on fossil fuels and nuclear power. The integration of these power sources into the power grid poses new challenges. To enable the economic exploitation of solar power, short-term forecasts of solar irradiance are required. We try to establish such forecasts using a ground-based camera. We aim to predict the movement of clouds with an image registration approach. First results are encouraging, yielding an improvement over prior work of 19%. The output of solar and wind power plants is dependent on the local weather and can fluctuate at short notice. The power grid, however, requires a stable supply. Therefore, production peaks have to be stored, and periods of low production have to be bridged. The switching between producers requires precise forecasts. Prediction errors can lead to a waste of power, instabilities, or even a collapse of the grid. We focus on the prediction of power output for solar power plants. Occlusions of the sun by single clouds are the main cause of drop-offs in power production. Traditional prediction methods for weather and cloud coverage, e. g. satellite imagery or numerical models [1], lack the necessary spatial and temporal resolution to predict these occlusions. To extend the existing prediction methods, we use a ground-based camera to monitor the sky. The feasibility of this approach for irradiance prediction has been shown by Chow et al. [2]. We refine their work to better address the difficult dynamics of cloud movement. Not only can clouds deform over time, but also clouds at different heights can move at different speeds and directions. Clouds can merge, existing clouds can dissipate and new clouds can form. Thus, the choice of the motion model is an important decision. Chow et al. proposed to use a rigid motion model by employing a block matching strategy to detect the cloud movement. This approach does not incorporate all of the mentioned aspects of cloud movement. Therefore, it has drawbacks when parts of clouds change their appearance over time, or when regions with different wind speeds are present. We address these limitations by using non-rigid registration to model complex dynamics of the cloud motion. In detail, we investigate two different approaches to non-rigid registration that employ different regularisation and matching strategies. – Thirions Demons [3] algorithm is based on optical flow. It iteratively calculates forces that deform the template image to match it to a reference image. Spatial (a) 0 800 1,600 2,400 3,200 2 3 4 5 t [s] e c Block Matching Demons Variational