Irradiance Forecasting through Analysis of Sky Images

The pressing need to reduce global emissions in the next decades is forcing countries to strive for a more sustainable model of development. Besides introducing necessary demand-side mitigation options, international panels recommend that governments carry out a deep energy transition, leading to a gradual decarbonisation of the supply-side. For this reason, a growing number of countries and energy companies are investing in more sustainable sources of energy, and in particular in solar energy, which is expected to become increasingly prevalent in renewable energy mixes.

The production of electricity from solar panels suffers, however, from high variability due to the discontinuity of the energy generation caused by sparse cloud cover. This currently limits the full development of solar energy. To address this problem, we would like to forecast irradiance (solar flux), and more broadly electricity production, from a few minutes to several months ahead. This would improve the following: plant and grid operations, quality of the power supply, grid planning, network balance, production optimisation and electricity trading.

With this in mind, the analysis of ground-based images (180 sky images taken by hemispherical cameras on the ground) is a promising starting point for local short-term irradiance forecasts in a 20 minutes time window, over an area imposed by the skyline.

Different approaches can be taken to such forecasting of irradiance from sequences of these images taken at fixed time intervals. One existing method is to generate, at each time step, a 3D model of the cloud cover from sky images and estimate future changes from observed displacements. Physical models are then exploited to approximate irradiance components at a given point from the predicted shadow on the ground.

This hand-crafted approach does, however, involve some approximations which undermine the quality of the forecast. The main issue being the difficulty of integrating the complexity of cloud structure and cloud movement into models.

The approach we will initially adopt involves the application of state-of-the-art Machine Learning techniques to tackle this forecasting problem. The understanding of the scene will be provided by Convolutional Neural Networks (CNN), which have proved to be a very efficient tool in computer vision for extracting relevant features from images at different levels of abstraction. The prediction step will be achieved by a Recurrent Neural Network (RNN), which enables the model to extract additional information from sequences, i.e. adds memory to the model. The idea is also to quantify forecast uncertainties from these deep neural networks through, for example, MC-Dropout or Gaussian Processes. Recently, a similar approach was successfully applied to forecasting rain forest conditions from satellite images.

The forecasting power of the model will be further improved to increase its temporal and spatial validity by using other tools from the fields of Statistics and Machine Learning and by exploiting additional data such as meteorological data or satellites images.