Convolutional Neural Networks applied to sky images for short-term solar irradiance forecasting

Abstract

Context. Reducing carbon emissions is critical to meeting international targets aimed at mitigating climate change. One way to reduce emissions is for countries to decarbonise their energy production in the upcoming decades. For this reason, the share of renewables is expected to increase in the global energy mix. In particular, the importance of solar energy for electricity production is set to rise. However, despite seamless advances in the estimation of the solar resource, there is still a need for better solar forecasting to improve its integration into the energy supply. Fish-eye cameras are emerging in-situ meteorological sensors that have already demonstrated promising and interesting results for high temporal resolution and very short term solar forecasting. The many innovations in the fields of machine learning and computer vision are likely to improve upon the performance of current solar forecasting techniques; this is to be expected given the complexity of the sky and cloud movement when viewed in high resolution. This work presents preliminary results of the use of deep Convolutional Neural Networks (CNNs) for short-term solar irradiance forecasting using sky images.Aim and Approach. To facilitate its integration and increase its economic value, solar energy would benefit from improvements in the forecasting of electricity generation at short (10 minutes), medium (1 hour) and long (1 day) term timescales. Tools available range from satellite imagery analysis and statistical modelling to ground-based sky image analysis: the latter provides very short term information on cloud cover changes which can be used, for example, to predict clouds hiding the sun. However, current approaches to model the cloud cover dynamics from sky images still lack precision regarding the spatial configuration of clouds, their temporal dynamics and their physical interaction with solar radiation.The approach introduced by Zhang (2018), Siddiqui (2019), Zhao (2019), which we follow here, is to apply recent computer vision techniques from the field of Deep Learning to forecast solar radiation up to 20 minutes in advance, from a range of data including sequences of past ground-based sky images. Artificial neural networks are implemented to extract relevant features from a dataset of around 64,000 images and in-situ pyranometric measurements taken from February to September 2018 in Palaiseau with a temporal resolution of 2-min (Haeffelin 2005).Scientific innovation and relevance. This work aims at bringing innovative insights via a novel approach to irradiance forecasting using the Deep Learning framework, which constitutes an effective environment for a richer modelling of the cloud cover and its dynamics. The model performance has been evaluated using the Mean Square Error (MSE) as error metric and a persistence of the clear-sky index as a reference to calculate the corresponding skill score. Also, visualisation methods were implemented to understand what the model has learnt during training and what region of the image it is ...

ECLIPSE : Envisioning CLoud Induced Perturbations in Solar Energy

“…Artificial Neural Networks (ANNs) were first successfully trained to translate an image of the sky into its corresponding simultaneous irradiance value [43,40]. Convolutional Neural Network (CNN) models have been shown to recognise specific cloud patterns and adjust their prediction accordingly [47,30]. Regarding the more challenging task of predicting future solar flux or solar energy production, numerous DL architectures have been shown to reach high quantitative performances: CNN [41,9], CNN + Long short Term Memory (LSTM) networks [50,10,30,47,38,29], 3D-CNN [51,29], implicit layers [24], Convolutional LSTM [21,29].…”

“…We compare Eclipse with the four DL architectures benchmarked in [29] on the 2, 4, 6, 8 and 10-min ahead predictions. These are representative of the different models published in recent studies: CNN [10,30,47], CNN+LSTM [50,38], 3D-CNN [51] and Convolutional LSTM [29]. Contrary to ECLIPSE, these models only predict the future irradiance level and not the corresponding future segmentations.…”

A moment in the sun

“…Finally, another line of research [28,30,36,43,44] utilizes skyimages collected from specialized cameras installed directly at a solar site, which continuously captures images of the sky at high resolution. Of course, this approach is not applicable to any site without such a specialized sky-imager.…”

“…Of course, this approach is not applicable to any site without such a specialized sky-imager. Indeed, prior work on such methods has only considered a very small number of solar sites, for instance, a maximum of only 2 solar sites in [28,30,36,43,44]. In contrast, our approach uses satellite data that is readily available for any location, and thus is widely applicable to any location within the GOES-R coverage area, i.e., much of North America.…”

“…Real-time data from satellites, such as GOES-16, presents an untapped opportunity for solar nowcasting at global scales using data-intensive deep learning techniques. Prior work has primarily focused on solar nowcasting using surface-level sky-imagery data [30,36,43,44], which predicts near-term solar generation from images of the sky taken at a solar site's location. Thus, this technique is "site-specic," and requires specialized hardware to be deployed at a particular location.…”