Vision-Based Solar Forecasting with Deep Learning

Abstract

Solar power is expected to play a leading role in the current electrification of our economy and its shift towards a low-carbon energy supply. This source of energy has numerous advantages including a wide availability and low costs, but also some limitations such as space and material usage. In addition, the inherent spatio-temporal variability induced by the cloud cover dynamics causes a large uncertainty in solar power supply, limiting its contribution to the energy mix. The diverse solutions developed to increase the reliability of solar energy, including improved storage systems or demand flexibility, require accurate estimations of the future solar energy yield. At a short-term scale, the main source of variability can be modelled by observing the cloud cover dynamics from sky cameras or satellites. Besides traditional physics-based algorithms, neural networks have recently shown considerable potential in tackling this computer vision task. Benefiting from a wider availability of resources and rapid progress in the field of machine learning, this interest is expected to grow together with the value of solar power forecasting and the addition of new solar capacities. In that respect, this thesis aims to evaluate, interpret and, thus, advance the applicability of machine learning to vision-based solar forecasting. Observed limitations of this data-driven approach are first quantified and illustrated. To address the difficulty of deep learning models to predict critical events on time, a novel spatio-temporal deep learning architecture named ECLIPSE is introduced and compared with existing models. Further, several scene representation and data augmentation strategies applied to cloud coverage modelling with neural networks are shown to improve predictions. Following this, a hybrid approach combining sky and satellite observations in a single machine learning framework is tested for intra-hour solar forecasting. This novel approach shows improved forecasts from a 25-min forecast horizon and realistic probabilistic predictions in various weather conditions. Finally, the possibility of combining diverse sky image datasets collected at different locations to improve training via transfer learning and dataset integration is assessed. Despite several ongoing challenges such as the generalisation of algorithms to new solar sites and the development of foundation models based on multilocation data, this thesis shows that the application of deep learning to cloud cover observations has a considerable potential to improve solar power forecasting. This, in turn, would facilitate the integration of solar energy into power systems by increasing its reliability and thus contribute to the ongoing energy transition.