Sky Images Research Articles

Short-term forecast of solar irradiance components using an alternative mathematical approach for the identification of cloud features

Solar energy technologies require precise solar forecasting to reduce power generation losses and protect equipment from irradiance fluctuations. This study introduces an alternative methodology for short-term forecasting of direct normal irradiance (DNI) and global horizontal irradiance (GHI) utilizing ground-based sky images captured by a single device. A low-cost all-sky imager (ASI) was developed, which implements an angular transformation and an optical flow technique to extract cloud features such as shape and velocity. A mathematical model calculates cloud transmittance based on pixel intensity, eliminating complex training steps. Results from a 30-day experimental campaign, incorporating diverse meteorological conditions, were compared against a secondary standard solarimetric station, a smart persistence model, and state-of-the-art approaches. The DNI forecast achieved an RMSE (relative error) of 46.79 W/m2 (11.99%) for 1-min intervals and 90.21 W/m2 (17.54%) for 10-min intervals, while GHI ranged from 31.73 W/m2 (4.68%) to 75.02 W/m2 (13.63%). Pearson correlation coefficients exceeded 0.9 overall, reaching 0.98 and 0.99 for the 1-min DNI and GHI forecasts, and 0.91 and 0.96 for the 10-min DNI and GHI forecasts, respectively, underscoring the system’s accuracy and robustness in complex meteorological scenarios.

An attention fused sequence -to-sequence convolutional neural network for accurate solar irradiance forecasting and prediction using sky images

The challenge of accurately forecasting ultra-short-term solar irradiance for photovoltaic systems is complicated by rapidly changing weather, and while ground-based sky images offer potential improvements, effectively extracting spatiotemporal data from these images remains a significant hurdle for current computer vision models. A new hybrid model, "An Attention Fused Sequence-to-Sequence Convolutional Neural Network," is being developed to address this challenge. The model predicts intra-hour GHI, DNI, and DHI with a 10-min lead time by combining a Convolutional Neural Network (for spatial feature extraction from sky images), an attention mechanism (to focus on relevant regions), and a sequence-to-sequence model (for temporal feature extraction from time-series data). The proposed Model is trained using the NREL Solar Radiation Research Laboratory Dataset while evaluating the model with the Mean Bias Error, Mean Absolute Error, Root Mean Squared Error, R Squared, and forecasting skill score. The 10-min and sequence length 2 interval is considered to be the best performing across most of the evaluation metrics with an MBE value is 2.321 W/m2, MAE value of 39.490 W/m2, RMSE value of 62.086 W/m2, R2 value is 0.909 W/m2, FSS value of 23.589 W/m2 and an MBE of 4.876 W/m2, MAE of 56.887 W/m2, RMSE of 85.346 W/m2, R2 of 0.834 and FSS of 24.881 W/m2 respectively. Furthermore, the sensitivity analysis reveals that the proposed model's performance is influenced by both the sequence length and the lead time. The proposed framework outperforms other techniques for ultra-short-term PV generation forecasting, demonstrating its potential for practical deployment in PV systems to improve grid reliability and energy management.

Retrieval of Solar Shortwave Irradiance from All-Sky Camera Images

The present work proposes a new model based on a convolutional neural network (CNN) to retrieve solar shortwave (SW) irradiance via the estimation of the cloud modification factor (CMF) from daytime sky images captured by all-sky cameras; this model is named CNN-CMF. To this end, a total of 237,669 sky images paired with SW irradiance measurements obtained by using pyranometers were selected at the following three sites: Valladolid and Izaña, Spain, and Lindenberg, Germany. This dataset was randomly split into training and testing sets, with the latter excluded from the training model in order to validate it using the same locations. Subsequently, the test dataset was compared with the corresponding SW irradiance measurements obtained by the pyranometers in scatter density plots. The linear fit shows a high determination coefficient (R2) of 0.99. Statistical analyses based on the mean bias error (MBE) values and the standard deviation (SD) of the SW irradiance differences yield results close to −2% and 9%, respectively. The MBE indicates a slight underestimation of the CNN-CMF model compared to the measurement values. After its validation, model performance was evaluated at the Antarctic station of Marambio (Argentina), a location not used in the training process. A similar comparison between the model-predicted SW irradiance and pyranometer measurements yielded R2=0.95, with an MBE of around 2% and an SD of approximately 26%. Although the precision provided by the SD at the Marambio station is lower, the MBE shows that the model’s accuracy is similar to previous results but with a slight overestimation of the SW irradiance. Finally, the determination coefficient improved to 0.99, and the MBE and SD are about 3% and 11%, respectively, when the CNN-CMF model is used to estimate daily SW irradiation values.

Sky images based photovoltaic power forecasting: A novel approach with optimized VMD and Vision Mamba

As the global demand for sustainable energy sources continues to grow, accurate prediction of photovoltaic power generation is crucial for optimizing the utilization of solar resources and enhancing the efficiency of photovoltaic systems. To improve the accuracy of photovoltaic power forecasting, this paper proposes a novel hybrid predictive model that integrates Optimized Variational Mode Decomposition (VMD), Vision Mamba (Vim) for extracting features from sky images, and advanced mechanisms like Patch Embedding and Variate-wise Cross-Attention. Initially, the proposed model employs SAO-optimized VMD to decompose the photovoltaic power series into high, medium, and low-frequency components. Subsequently, these components are patched to serve as input for the subsequent layers. In the third step, exogenous variables, including meteorological and image data, are introduced and processed through Variate Embedding combined with cross-attention mechanisms to capture the intricate interactions between these variables. Finally, by integrating the outputs from all processing steps through normalization and feed-forward layers, the final predictive results are produced. Experimental evaluations across different seasons demonstrate significant enhancements in forecasting accuracy, with the model achieving Root Mean Square Error (RMSE) values of 0.3587 in spring, 0.4376 in summer, 0.3544 in autumn, and 0.3493 in winter. Similarly, Mean Absolute Error (MAE) and Mean Squared Error (MSE) across these seasons underscore the model's effectiveness. This model offers new technical means for photovoltaic power forecasting and provides valuable decision support for the optimization and management of photovoltaic power systems.

LIGO-Virgo-KAGRA Limit on Relativistic Interstellar Objects near Earth

Relativistic interstellar objects (RISOs) are not constrained by astronomical sky surveys because they are smeared in sky images and at best appear in one frame. Here, I show that irrespective of their nature, RISOs more massive than ∼1014 g would have been detected by the LIGO-Virgo-KAGRA collaboration through their tidal gravitational signal at a frequency of f ∼ 50 Hz within a distance comparable to the Earth radius. This constrains the passage of relativistic primordial black holes or other exotic objects at the local mass density of dark matter near Earth over the past decade.

A neural network to retrieve cloud cover from all‐sky cameras: A case of study over Antarctica

AbstractWe present a new model based on a convolutional neural network (CNN) to predict daytime cloud cover (CC) from sky images captured by all‐sky cameras, which is called CNN‐CC. A total of 49,016 daytime sky images, recorded at different Spanish locations (Valladolid, La Palma, and Izaña) from two different all‐sky camera types, are manually classified into different CC (oktas) values by trained researchers. Subsequently, the images are randomly split into a training set and a test set to validate the model. The CC values predicted by the CNN‐CC model are compared with the observations made by trained people on the test set, which serve as reference. The predicted CC values closely match the reference values within 1 oktas in 99% of the cloud‐free and overcast cases. Moreover, this percentage is above 93% for the rest of partially cloudy cases. The mean bias error (MBE) and standard deviation (SD) of the differences between the predicted and reference CC values are calculated, resulting in oktas and oktas. The MBE and SD are also represented for different intervals of measured aerosol optical depth and Ångström exponent values, revealing that the performance of the CNN‐CC model does not depend on aerosol load or size. Once the model is validated, the CC obtained from a set of images captured every 5 min, from January 2018 to March 2022, at the Antarctic station of Marambio (Argentina) is compared against direct field observations of CC (not from images) taken at this location, which is not used in the training process. As a result, the model slightly underestimates the observations with an MBE of 0.3 oktas. The retrieved data are analyzed in detail. The monthly and annual CC values are calculated. Overcast conditions are the most frequent, accounting for 46.5% of all observations throughout the year, rising to 64.5% in January. The annual mean CC value at this location is 5.5 oktas, with a standard deviation of approximately 3.1 oktas. A similar analysis is conducted, separating data by hours, but no significant diurnal cycles are observed except for some isolated months.

Особенности функционирования аллегории в произведениях русской монашеской словесности XX века

The aim of the research is to present the main features of the functioning of allegory in the works of 20th-century Russian monastic writers. The paper examines Christian allegory as a specific tropeic phenomenon. The specificity of the use of Christian allegory lies in its orientation toward the culture-forming texts of the Bible; in its fixation in the didactic parts of the text dedicated to the salvation of the soul, spiritual perfection, the need to strive for virtue; in the authors’ appeal to allegorical images with stable interpretations embedded in the culture; in its concentration on allegories that are part of the established allegorical imagery corpus, which formed as Christian literature were developing. The central place in the work is occupied by the study of the representation of ideas important for church writers through allegory, the actualization of Christian meanings. The scientific novelty of the research lies in the fact that it uses theolinguistic analysis to study the works of Russian monastic literature, indicates the stylistic functions of allegory in the ecclesiastical-religious style, and identifies the main features of the use of allegory. As a result, the following has been proved: when affirming thoughts related to the perniciousness of sins, authors resort to Christian allegories, using images of a swamp, abyss, net, thorns, tree branches, spiderwebs; when describing the life of a person and the path of a Christian, images of a ship, boat, sea, storm, and house are used; in reasoning about virtues, images of water, sun, and sky are chosen. There are several stylistic functions of allegory identified in the texts of monastic literature, such as pictorial, syndicating, emotive, prescriptive, meaning-forming, didactic, axiological. The features of the functional orientation of allegory in the works of monastic literature are as follows: the use with didactic religious-Christian goals, i.e., aversion from sin and encouragement to acquire virtues, strive for the salvation of the soul, moral perfection, spiritual alertness, and the cultivation of the inner person.

IoT and machine learning models for multivariate very short‐term time series solar power forecasting

AbstractIn solar energy generation, the inherent variability caused by cloud cover and weather events often leads to sudden fluctuations in power outputs. Addressing this challenge, the authors’ study focuses on very short‐term solar irradiance (SI) prediction. Leveraging multivariate time series datasets, the authors improve very short‐term SI predictions. To achieve accurate very short‐term SI predictions, the authors employ machine learning techniques throughout the forecasting process. Additionally, the authors’ work pioneers the integration of the Internet of Things (IoT) into solar power systems, a novel approach in the field. The authors leverage LoRa (long range) technology for low‐cost, low‐power, and long‐range wireless control networks. The authors’ study focuses on SI forecasting using long short‐term memory and bi‐directional long short‐term memory (Bi‐LSTM) models, achieving high accuracy. The SI forecasts are evaluated in terms of root‐mean‐square error (RMSE) and mean absolute error in relation to meteorological data and sky image data. The improvement in performance can be attributed to the Bi‐LSTM's bidirectional nature, allowing it to incorporate future information during training, thereby enhancing its predictive capabilities. Overall, the results suggest that the Bi‐LSTM model is more suitable for accurately forecasting SI, particularly in scenarios requiring short‐term predictions based on rapidly changing environmental factors.

A radiant shift: Attention-embedded CNNs for accurate solar irradiance forecasting and prediction from sky images

The continuous increase in solar power integration with energy systems can be attributed to the push for cleaner energy use globally, highlighting the importance of accurate solar forecasts. Conventional prediction methods, although valuable, often fall short of delivering the precision required for dynamic energy management systems due to their inability to effectively capture the intricate relationships inherent in solar irradiance variations. This research addresses this limitation by introducing a novel approach that harnesses the potential of Attention-embedded Convolutional Neural Networks (ATT_CNN) to forecast Global Horizontal Irradiance (GHI), Direct Normal Irradiance (DNI), and Diffuse Horizontal Irradiance (DHI) for intra-hour solar forecasting utilizing sequences of sky images. The main contribution of this study is the integration of attention mechanisms into the CNN architecture, strategically designed to enhance their efficacy in predicting GHI, DNI, and DHI by fostering an adaptive focus on sky image features. First, we use SRRL dataset with GHI, DNI, and DHI features as predicting labels, then the Mean Bias Error (MBE), Root Mean Square Error (RMSE), and Forecasting RMSE skill score (FSS) are used to evaluate the model's performance. This study utilized six lead times and four sequence lengths, based on this the best combination is: Sequence length: 4 with Minutes: 20. This combination provides a balanced and optimal performance with low RMSE (62.75 W/m2), low MBE (2.71 W/m2), and a high FSS (38.81), indicating good accuracy, minimal bias, and high skill score. Furthermore, when compared with state-of-the-art models, the proposed model yielded superior results. This advancement holds profound implications for the optimization of solar power utilization within mainstream energy systems, further underscoring the significance of cutting-edge deep learning (DL) techniques in advancing sustainable energy technologies. The findings of this study indicate that integrating attention mechanisms is essential to enhance the accuracy and reliability of forecasts, and using longer sequences of images can further improve forecasting performance.

Results from the InSight atmospheric imaging campaign

NASA's InSight lander monitored the Martian atmosphere while conducting its primarily geophysical investigation. Atmospheric imaging was used to study dust and ice at the site for over two Mars years in 2018–2022. An optical depth record, including dust and ice, was derived from systematic sky imaging in the mornings (for the first part of the mission) and evenings. Optical depths ranged from 0.5 to 1.9 but were typically under 1. Dust storms were seen at expected times in late northern autumn and early winter, including one shortly after landing, along with one late summer storm in January 2022. The optical depth record closely matched that of Curiosity, 600 km to the south, except for the expected additional water ice content during the aphelion cloud belt (ACB, spring to early summer). In addition to ice hazes, the ACB included discrete clouds, whose motion was tracked to show northeasterly to southeasterly daytime winds. While InSight recorded many meteorological vortices, no dust devils were seen, requiring that dust-devil occurrence was <10−3 times as common as during Spirit rover dust devil seasons.

Fisheye Night Sky Imager: A Calibrated Tool to Measure Night Sky Brightness

The Night Skies Team of the U.S. National Park Service (NPS) has developed a new camera system equipped with a fisheye lens to measure night sky brightness. This NPS Fisheye Night Sky Imager comprises a back-illuminated Sony IMX455 CMOS sensor housed in the ZWO ASI6200MM camera, a Johnson V filter, and a Sigma 8 mm F3.5 fisheye lens. All of the components are commercially available. The fisheye lens allows us to capture the entire sky in a single frame. The wide-angle view also introduces calibration challenges. To address this, we have customized tools and devised innovative methods for calibrating the system, including flat-field correction, pixel scale determination, and field of view characterization. Additionally, we use standard stars for precise positional and photometric calibration. Furthermore, we create an open-source Python pipeline to process these fisheye images and leverage Git and GitHub for source code version control and distribution. The resulting images provide a positionally and brightness-calibrated fisheye view of the night sky, with a photometric calibration uncertainty of 0.12 mag. When comparing our Fisheye Night Sky Imager measurements to images taken with other calibrated systems, we reliably achieve consistent results under both dark and bright skies. Our Fisheye Night Sky Imager not only enables rapid assessment of sky quality but also supports the feasibility of establishing permanent monitoring locations. Recognizing that the natural dark sky is a critical element for natural, cultural, educational, and visitor experiences, this new camera system enhances our ability to assess resource conditions and fulfill the NPS mission of conserving resources unimpaired for the enjoyment of this and future generations.

The height of the diurnal atmosphere: The twilight

An equation has been derived through which the height of the diurnal atmospheric layer (lower atmosphere) can be calculated using ground base twilight image analysis. This equation relies on the angular altitude of the twilight with the dip at that moment. Based on the images, it was found that true twilight of dawn occurs at a dip of 14.90° ± 0.17° as detected by the camera or ∼15° by considering the elevation above sea level (while it may be delayed to ∼14.7–14° with the naked eye), which is accompanied by an average twilight angular altitude of 2.66° ± 0.23°. From these values, the twilight height is estimated to be 303.80 ± 5.69 km above sea level. It is shown that the twilight region is the peak zone of the thermosphere. Its height also varies with time according to changes in various atmospheric parameters. Lower wind speeds, higher relative humidity, higher dew point, and lower temperatures resulted in a higher altitudinal angle of the atmosphere and consequently a greater diurnal atmospheric layer’s thickness. Image analysis shows that the thermosphere always has a maximum value of light brightness during the night even before the appearance of twilight. Additionally, It defined the light source, shapes, and stages of twilights that detected by images. This implies that sky image analysis is an inexpensive method for understanding the properties and various parameters of the atmosphere.

Computer Vision Techniques Demonstrate Robust Orientation Measurement of the Milky Way Despite Image Motion.

Many species rely on celestial cues as a reliable guide for maintaining heading while navigating. In this paper, we propose a method that extracts the Milky Way (MW) shape as an orientation cue in low-light scenarios. We also tested the method on both real and synthetic images and demonstrate that the performance of the method appears to be accurate and reliable to motion blur that might be caused by rotational vibration and stabilisation artefacts. The technique presented achieves an angular accuracy between a minimum of 0.00° and a maximum 0.08° for real night sky images, and between a minimum of 0.22° and a maximum 1.61° for synthetic images. The imaging of the MW is largely unaffected by blur. We speculate that the use of the MW as an orientation cue has evolved because, unlike individual stars, it is resilient to motion blur caused by locomotion.

Intelligent constellation detection using pattern recognition

The study of constellations and the mechanism for its identification has always been an captivating area in astronomy, from ancient methods like star charts, to modern techniques of computer vision. Along with continuous discoveries of new celestial objects on a regular basis, the study of constellations are getting equal attention in astronomical data analysis. In this paper, we present an intelligent constellation detection algorithm that identifies the constellation using star patterns in the given image and produces meta-data for it to the end user. That is, for a given a deep sky image, the system effectively detects the stars, produces the orientation and labels the stars and constellations patterns, maintains the database of the images and the detected objects. The system provides a comprehensive user interface for data input and to display the results obtained. The system demonstrated 90% success rate in correctly identifying sources and mapping the constellations. The system also serves as a valuable platform for astronomers studying deep sky images enabling the identification of anomalies in the images such as unidentified objects not cataloged or previously unidentified in constellation images.

Sky Image Classification Based on Transfer Learning Approaches.

Cloudy conditions at a local scale pose a significant challenge for forecasting renewable energy generation through photovoltaic panels. Consequently, having real-time knowledge of sky conditions becomes highly valuable. This information could inform decision-making processes in system operations, such as determining whether conditions are favorable for activating a standalone system requiring a minimum level of radiation or whether sky conditions might lead to higher energy consumption than generation during adverse cloudy conditions. This research leveraged convolutional neural networks (CNNs) and transfer learning (TL) classification techniques, testing various architectures from the EfficientNet family and two ResNet models for classifying sky images. Cross-validation methods were applied across different experiments, where the most favorable outcome was achieved with the EfficientNetV2-B1 and EfficientNetV2-B2 models boasting a mean Accuracy of 98.09%. This study underscores the efficacy of the architectures employed for sky image classification, while also highlighting the models yielding the best results.

Non-thermal emission from mildly relativistic dynamical ejecta of neutron star mergers: spectrum and sky image

Abstract Binary neutron star mergers are expected to produce fast dynamical ejecta, with mildly relativistic velocities extending to β = v/c &amp;gt; 0.6. In a preceding paper, we derived an analytic description of the time-dependent radio to X-ray synchrotron flux produced by collisionless shocks driven by such fast ejecta into the interstellar medium, for spherical ejecta with broken power-law mass (or energy) distributions, M( &amp;gt; γβ)∝(γβ)−s with s = sKN at γβ &amp;lt; γ0β0 and s = sft at γβ &amp;gt; γ0β0 (where γ is the Lorentz factor). Here, we extend our analysis and provide analytic expressions for the self-absorption frequency, the cooling frequency, and the observed angular size of the emitting region (which appears as a ring in the sky). For parameter values characteristic of merger calculation results - a ”shallow” mass distribution, 1 &amp;lt; sKN &amp;lt; 3, for the bulk of the ejecta (at γβ ≈ 0.2), and a steep, sft &amp;gt; 5, ”fast tail” mass distribution – the analytic results reproduce well (to tens of percent accuracy) the results of detailed numeric calculations, a significant improvement over earlier order-of-magnitude estimates (based on extrapolations of results valid for γβ ≪ 1).

Classification of Rainfall Intensity and Cloud Type from Dash Cam Images Using Feature Removal by Masking

Weather Report is an initiative from Weathernews Inc. to obtain sky images and current weather conditions from the users of its weather app. This approach can provide supplementary weather information to radar observations and can potentially improve the accuracy of forecasts However, since the time and location of the contributed images are limited, gathering data from different sources is also necessary. This study proposes a system that automatically submits weather reports using a dash cam with communication capabilities and image recognition technology. This system aims to provide detailed weather information by classifying rainfall intensities and cloud formations from images captured via dash cams. In models for fine-grained image classification tasks, there are very subtle differences between some classes and only a few samples per class. Therefore, they tend to include irrelevant details, such as the background, during training, leading to bias. One solution is to remove useless features from images by masking them using semantic segmentation, and then train each masked dataset using EfficientNet, evaluating the resulting accuracy. In the classification of rainfall intensity, the model utilizing the features of the entire image achieved up to 92.61% accuracy, which is 2.84% higher compared to the model trained specifically on road features. This outcome suggests the significance of considering information from the whole image to determine rainfall intensity. Furthermore, analysis using the Grad-CAM visualization technique revealed that classifiers trained on masked dash cam images particularly focused on car headlights when classifying the rainfall intensity. For cloud type classification, the model focusing solely on the sky region attained an accuracy of 68.61%, which is 3.16% higher than that of the model trained on the entire image. This indicates that concentrating on the features of clouds and the sky enables more accurate classification and that eliminating irrelevant areas reduces misclassifications.

A CNN-LSTM model using elliptical constraints for temporally consistent sun position estimation

More accurate sun position estimation could transform the design and operation of solar power systems, weather forecasting services, and outdoor augmented reality systems. Although several image-based approaches to sun position estimation have been proposed, their performance is significantly affected by momentary disruptions in cloud cover because they use only a single image as input. This study proposes a deep learning-based sun position estimation system that leverages spatial, temporal, and geometric features to accurately regress sun positions even when the sun is partially or entirely occluded. In the proposed approach, spatial features are extracted from an input image sequence by applying a separate Resnet-based convolution network to each frame. To ensure that the temporal changes in the brightness distribution across frames are considered, the spatial features are concatenated and passed on to a stack of LSTM layers prior to regressing the final sun position. The proposed network is also trained with elliptical (geometric) constraints to ensure that predicted sun positions are consistent with the natural elliptical path of the sun in the sky. The proposed approach's performance was evaluated on the Sirta and Laval datasets along with a custom dataset, and an R2 Score of 0.98 was achieved, which is at least 0.1 higher than that of previous approaches. The proposed approach is capable of identifying the position of the sun even when occluded and was employed in a novel sky imaging system consisting of only a camera and fisheye lens in place of a complex array of sensors.

Improving cross-site generalisability of vision-based solar forecasting models with physics-informed transfer learning

Forecasting solar energy from cloud cover observations is crucial to truly anticipate future changes in power supply. On an intra-hour timescale, ground-level sky cameras located near a solar site offer the most valuable source of information on incoming clouds. In the literature, the analysis of these hyperlocal cloud cover observations for solar modelling is increasingly performed by deep learning algorithms trained and tested on years’ worth of local data. However, this approach is not suitable for industrial applications since solar energy producers cannot wait for years of local data collection to start generating reliable solar forecasts. However, they might own relevant multi-location data collected from other solar sites over time. This study thus explores the capability of such algorithms to generalise beyond their training location in two data scarce conditions: zero-shot learning (i.e. direct application of a trained model to a new location without local fine-tuning) and few-shot learning (i.e. calibration of a pre-trained model based on very limited local data such as a day of observations). Zero-shot learning results show that using local clear-sky models to normalise output variables (e.g. solar irradiance or solar energy production values) facilitates cross-dataset transfer learning. Compared to previous methods, the resulting forecast skill increases by close to 25% in cloudy conditions and by more than 700% in clear-sky conditions. An additional gain is observed when local data collected in overcast weather conditions are used for model calibration via few-shot learning. The corresponding neural networks trained in data scarce conditions achieve comparable performance to expert local models based on years of training data. These promising results shed light on the potential of large-scale and multi-location sky image datasets to improve the generalisation skills of solar forecasting algorithms.

SkyGPT: Probabilistic ultra-short-term solar forecasting using synthetic sky images from physics-constrained VideoGPT

The variability of solar photovoltaic (PV) power output, driven by rapidly changing cloud dynamics, hinders the transition to reliable renewable energy systems. Information on future sky conditions, especially cloud coverage, holds the promise for improving PV output forecasting. Leveraging recent advances in generative artificial intelligence (AI), we introduce SkyGPT, a physics-constrained stochastic video prediction model, which predicts plausible future images of the sky using historical sky images. We show that SkyGPT can accurately capture cloud dynamics, producing highly realistic and diverse future sky images. We further demonstrate its efficacy in 15-minute-ahead probabilistic PV output forecasting using real-world power generation data from a 30-kW rooftop PV system. By coupling SkyGPT with a U-Net-based PV power prediction model, we observe superior prediction reliability and sharpness compared with several benchmark methods. The propose approach achieves a continuous ranked probability score (CRPS) of 2.81 kW, outperforming a classic convolutional neural network (CNN) baseline by 13% and the smart persistence model by 23%. The findings of this research could aid efficient and resilient management of solar electricity generation, particularly as we transition to renewable-heavy grids. The study also provides valuable insights into stochastic cloud modeling for a broad research community, encompassing fields such as solar energy meteorology and atmospheric sciences.