Benchmarking of Deep Learning Irradiance Forecasting Models from Sky Images -- an in-depth Analysis

Abstract

A number of industrial applications, such as smart grids, power plant operation, hybrid system management or energy trading, could benefit from improved short-term solar forecasting, addressing the intermittent energy production from solar panels. However, current approaches to modelling the cloud cover dynamics from sky images still lack precision regarding the spatial configuration of clouds, their temporal dynamics and physical interactions with solar radiation. Benefiting from a growing number of large datasets, data driven methods are being developed to address these limitations with promising results. In this study, we compare four commonly used Deep Learning architectures trained to forecast solar irradiance from sequences of hemispherical sky images and exogenous variables. To assess the relative performance of each model, we used the Forecast Skill metric based on the smart persistence model, as well as ramp and time distortion metrics. The results show that encoding spatiotemporal aspects of the sequence of sky images greatly improved the predictions with 10 min ahead Forecast Skill reaching 20.4% on the test year. However, based on the experimental data, we conclude that, with a common setup, Deep Learning models tend to behave just as a 'very smart persistence model', temporally aligned with the persistence model while mitigating its most penalising errors. Thus, despite being captured by the sky cameras, models often miss fundamental events causing large irradiance changes such as clouds obscuring the sun. We hope that our work will contribute to a shift of this approach to irradiance forecasting, from reactive to anticipatory.

ECLIPSE : Envisioning CLoud Induced Perturbations in Solar Energy

“…For instance, statistical approaches reach high RMSE FS by avoiding the largest errors despite no anticipation skill. This behaviour is notably common with current DL approaches as observed by [29]. Behaving similarly as a 'very smart persistence' model, DL models tend to suffer from a consistent time lag, thus frequently missing critical events such as a cloud hiding the sun.…”

“…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].…”

“…Despite promising results based on common metrics such as the root Mean Square Error (RMSE), DL approaches applied to irradiance forecasting still face major limitations. In particular, traditional architectures suffer from a systematic forecast time lag, hence low anticipation skills [21,29]. Instead of actively anticipating future events (high irradiance shifts), neural networks tend to react to the current weather conditions in order to minimise the future risk of large errors.…”

“…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.…”

SolarGAN: Synthetic Annual Solar Irradiance Time Series on Urban Building Facades via Deep Generative Networks

“…To the best of our knowledge, this multimodal architecture has no precedent in data-driven urban solar study, but its design is coherent to the basic concept of crossmodal translation [32]. We also noticed the work of Paletta et al [33,34] that use fisheye photos obtained from the sky imager to predict ultra-short-term solar irradiance exhibit similar mindset and data pipeline, although they did not apply DGNs. The detailed explanation of DGNs applied in our proposed model could be found in Section 3.3 in the section of methodology.…”