• Adrian, R. J., , and Westerweel J. , 2010: Particle Image Velocimetry. Cambridge Aerospace Series, Vol. 30, Cambridge University Press, 586 pp.

  • Bishop, C. M., 1994: Neural networks and their applications. Rev. Sci. Instrum., 65, 18031832, doi:10.1063/1.1144830.

  • Calbo, J., , and Sabburg J. , 2008: Feature extraction from whole-sky ground-based images for cloud-type recognition. J. Atmos. Oceanic Technol., 25, 34, doi:10.1175/2007JTECHA959.1.

    • Search Google Scholar
    • Export Citation
  • Cazorla, A., , Olmo F. J. , , and Alados-Arboledas L. , 2008: Development of a sky imager for cloud cover assessment. J. Opt. Soc. Amer., 25A, 2939, doi:10.1364/JOSAA.25.000029.

    • Search Google Scholar
    • Export Citation
  • Chow, C. W., , Urquhart B. , , Lave M. , , Dominguez A. , , Kleissl J. , , Shields J. , , and Washom B. , 2011: Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed. Sol. Energy, 85, 28812893, doi:10.1016/j.solener.2011.08.025.

    • Search Google Scholar
    • Export Citation
  • Chu, Y., , Pedro H. T. C. , , and Coimbra C. F. M , 2013: Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning. Sol. Energy, 98, 592603, doi:10.1016/j.solener.2013.10.020.

    • Search Google Scholar
    • Export Citation
  • Efron, B., , and Gong G. , 1983: A leisurely look at the bootstrap, the jackknife, and cross-validation. Amer. Stat., 37, 3648.

  • Geisser, S., 1993: Predictive Inference. Monogr. on Statistics and Applied Probability, Vol. 55, Chapman & Hall/CRC, 31–42.

  • Ghonima, M. S., , Urquhart B. , , Chow C. W. , , Shields J. E. , , Cazorla A. , , and Kleissl J. , 2012: A method for cloud detection and opacity classification based on ground based sky imagery. Atmos. Meas. Tech. Discuss., 5, 45354569, doi:10.5194/amtd-5-4535-2012.

    • Search Google Scholar
    • Export Citation
  • Gueymard, C. A., , and Wilcox S. M. , 2011: Assessment of spatial and temporal variability in the US solar resource from radiometric measurements and predictions from models using ground-based or satellite data. Sol. Energy, 85, 10681084, doi:10.1016/j.solener.2011.02.030.

    • Search Google Scholar
    • Export Citation
  • Heinle, A., , Macke A. , , and Srivastav A. , 2010: Automatic cloud classification of whole sky images. Atmos. Meas. Tech., 3, 557567, doi:10.5194/amt-3-557-2010.

    • Search Google Scholar
    • Export Citation
  • Holland, J. H., 1992: Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence. MIT Press, 211, pp.

  • Ineichen, P., 2006: Comparison of eight clear sky broadband models against 16 independent data banks. Sol. Energy, 80, 468478, doi:10.1016/j.solener.2005.04.018.

    • Search Google Scholar
    • Export Citation
  • Ineichen, P., 2008: A broadband simplified version of the Solis clear sky model. Sol. Energy, 82, 758762, doi:10.1016/j.solener.2008.02.009.

    • Search Google Scholar
    • Export Citation
  • Ineichen, P., , and Perez R. , 2002: A new airmass independent formulation for the Linke turbidity coefficient. Sol. Energy, 73, 151157, doi:10.1016/S0038-092X(02)00045-2.

    • Search Google Scholar
    • Export Citation
  • Inman, R., , Pedro H. , , and Coimbra C. , 2013: Solar forecasting methods for renewable energy integration. Prog. Energy Combust. Sci., 39, 535576, doi:10.1016/j.pecs.2013.06.002.

    • Search Google Scholar
    • Export Citation
  • Jain, A. K., , Duin R. P. W. , , and Mao J. , 2000: Statistical pattern recognition: A review. IEEE Trans. Pattern Anal. Mach. Intell., 22, 437, doi:10.1109/34.824819.

    • Search Google Scholar
    • Export Citation
  • Kohavi, R., 1995: A study of cross-validation and bootstrap for accuracy estimation and model selection. IJCAI ’95: Proceedings of the 14th International Joint Conference on Artificial Intelligence, Vol. 2, C. S. Mellish, Ed., Morgan Kaufmann, 1137–1145.

  • Lendasse, A., , Wertz V. , , and Verleysen M. , 2003: Model selection with cross-validations and bootstraps—Application to time series prediction with RBFN models. Artificial Neural Networks and Neural Information Processing—ICANN/ICONIP 2003, O. Kaynak et al., Eds., Lecture Notes in Computer Science, Vol. 2714, Springer, 573–580.

  • Li, C. H., , and Lee C. K. , 1993: Minimum cross entropy thresholding. Pattern Recognit., 26, 617625, doi:10.1016/0031-3203(93)90115-D.

  • Li, C. H., , and Tam P. K. S. , 1998: An iterative algorithm for minimum cross entropy thresholding. Pattern Recognit. Lett., 19, 771776, doi:10.1016/S0167-8655(98)00057-9.

    • Search Google Scholar
    • Export Citation
  • Li, Q., , Lu W. , , and Yang J. , 2011: A hybrid thresholding algorithm for cloud detection on ground-based color images. J. Atmos. Oceanic Technol., 28, 12861296, doi:10.1175/JTECH-D-11-00009.1.

    • Search Google Scholar
    • Export Citation
  • Long, C. N., , and Ackerman T. P. , 2000: Identification of clear skies from broadband pyranometer measurements and calculation of downwelling shortwave cloud effects. J. Geophys. Res., 105, 15 60915 626, doi:10.1029/2000JD900077.

    • Search Google Scholar
    • Export Citation
  • Long, C. N., , Sabburg J. M. , , Calbó J. , , and Pages D. , 2006: Retrieving cloud characteristics from ground-based daytime color all-sky images. J. Atmos. Oceanic Technol., 23, 633652, doi:10.1175/JTECH1875.1.

    • Search Google Scholar
    • Export Citation
  • López, G., , Batlles F. J. , , and Tovar-Pescador J. , 2005: Selection of input parameters to model direct solar irradiance by using artificial neural networks. Energy, 30, 16751684, doi:10.1016/j.energy.2004.04.035.

    • Search Google Scholar
    • Export Citation
  • Mantelli Neto, S. L. M., , von Wangenheim A. , , Pereira E. B. , , and Comunello E. , 2010: The use of Euclidean geometric distance on RGB color space for the classification of sky and cloud patterns. J. Atmos. Oceanic Technol., 27, 15041517, doi:10.1175/2010JTECHA1353.1.

    • Search Google Scholar
    • Export Citation
  • Marquez, R., , and Coimbra C. F. M. , 2011: Forecasting of global and direct solar irradiance using stochastic learning methods, ground experiments and the NWS database. Sol. Energy, 85, 746756, doi:10.1016/j.solener.2011.01.007.

    • Search Google Scholar
    • Export Citation
  • Marquez, R., , and Coimbra C. F. M. , 2013: Intra-hour DNI forecasting methodology based on cloud tracking image analysis. Sol. Energy, 91, 327336, doi:10.1016/j.solener.2012.09.018.

    • Search Google Scholar
    • Export Citation
  • Marquez, R., , Gueorguiev V. G. , , and Coimbra C. F. M. , 2013a: Forecasting of global horizontal irradiance using sky cover indices. J. Sol. Energy Eng.,135, 011017, doi:10.1115/1.4007497.

  • Marquez, R., , Pedro H. T. C. , , and Coimbra C. F. M. , 2013b: Hybrid solar forecasting method uses satellite imaging and ground telemetry as inputs to ANNs. Sol. Energy, 92, 176188, doi:10.1016/j.solener.2013.02.023.

    • Search Google Scholar
    • Export Citation
  • McLachlan, G. J., , Do K.-A. , , and Ambroise C. , 2004: Analyzing Microarray Gene Expression Data. Wiley Series in Probability and Statistics, Vol. 422, Wiley-Interscience, 270–285 pp.

  • Mellit, A., , and Kalogirou S. A. , 2008: Artificial intelligence techniques for photovoltaic applications: A review. Prog. Energy Combust. Sci., 34, 574632, doi:10.1016/j.pecs.2008.01.001.

    • Search Google Scholar
    • Export Citation
  • Mellit, A., , and Pavan A. M. , 2010: A 24-h forecast of solar irradiance using artificial neural network: Application for performance prediction of a grid-connected PV plant at Trieste, Italy. Sol. Energy, 84, 807821, doi:10.1016/j.solener.2010.02.006.

    • Search Google Scholar
    • Export Citation
  • Otsu, N., 1979: A threshold selection method from gray-level histograms. IEEE Trans. Syst. Man Cybern., 9, 6266, doi:10.1109/TSMC.1979.4310076.

    • Search Google Scholar
    • Export Citation
  • Pai, P. F., , and Hong W. C. , 2005: Forecasting regional electricity load based on recurrent support vector machines with genetic algorithms. Electr. Power Syst. Res., 74, 417425, doi:10.1016/j.epsr.2005.01.006.

    • Search Google Scholar
    • Export Citation
  • Pedro, H. T. C., , and Coimbra C. F. M. , 2012: Assessment of forecasting techniques for solar power production with no exogenous inputs. Sol. Energy, 86, 20172028, doi:10.1016/j.solener.2012.04.004.

    • Search Google Scholar
    • Export Citation
  • Reno, M. J., , Hansen C. W. , , and Stein J. S. , 2012: Global horizontal irradiance clear sky models: Implementation and analysis. Sandia National Laboratories Rep. SAND2012-2389, 67 pp. [Available online at http://energy.sandia.gov/wp/wp-content/gallery/uploads/SAND2012-2389_ClearSky_final.pdf.]

  • Rossow, W. B., , and Schiffer R. A. , 1991: ISCCP cloud data products. Bull. Amer. Meteor. Soc., 72, 220, doi:10.1175/1520-0477(1991)072<0002:ICDP>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Seiz, G., , Shields J. , , Feister U. , , Baltsavias E. P. , , and Gruen A. , 2007: Cloud mapping with ground-based photogrammetric cameras. Int. J. Remote Sens., 28, 20012032, doi:10.1080/01431160600641822.

    • Search Google Scholar
    • Export Citation
  • Souza-Echer, M. P., , Pereira E. B. , , Bins L. S. , , and Andrade M. A. R. , 2006: A simple method for the assessment of the cloud cover state in high-latitude regions by a ground-based digital camera. J. Atmos. Oceanic Technol., 23, 437447, doi:10.1175/JTECH1833.1.

    • Search Google Scholar
    • Export Citation
  • Voss, K., , and Zibordi G. , 1989: Radiometric and geometric calibration of a visible spectral electro-optic “fisheye” camera radiance distribution system. J. Atmos. Oceanic Technol., 6, 652662, doi:10.1175/1520-0426(1989)006<0652:RAGCOA>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Yang, J., , W. , , Ma Y. , , Yao W. , , and Li Q. , 2009: An automatic ground based cloud detection method based on adaptive threshold. J. Appl. Meteor. Sci., 20, 713721.

    • Search Google Scholar
    • Export Citation
  • Younes, S., , and Muneer T. , 2007: Clear-sky classification procedures and models using a world-wide data-base. Appl. Energy, 84, 623645, doi:10.1016/j.apenergy.2007.01.016.

    • Search Google Scholar
    • Export Citation
  • Zhao, G., , and Di Girolamo L. , 2006: Cloud fraction errors for trade wind cumuli from EOS-Terra instruments. Geophys. Res. Lett., 33, L20802, doi:10.1029/2006GL027088.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Schematic of SACI. Dashed-line square encloses the SIC.

  • View in gallery

    Examples of (top) some original images and (bottom) their NRBR distribution histograms for the (left) clear, (middle) overcast, and (right) partly cloudy periods.

  • View in gallery

    Examples of (top) original images, (middle) RBR images, and (bottom) CSL method–subtracted RBR images for the (left) clear, (middle) overcast, and (right) partly cloudy periods. Color bars indicate RBR magnitudes for the pixels in each image.

  • View in gallery

    Main steps of the grid cloud fractions method. (a),(b) Original consecutive image pair used in the analysis. (c),(d) Image pair is projected onto a rectangular grid, and (e) velocity vectors computed by the PIV algorithm. (f) Cloud classified image, where white pixels represent cloud; gray pixels represent clear sky; and black pixels are obstacles, which are excluded from the analysis.

  • View in gallery

    Example time series of the PIV and the manually identified cloud moving directions on 1 Apr 2013. Error range (marked as yellow) of the manually identified directions is 20°. PIV-identified directions that locate in the yellow region are considered accurate tested instances.

  • View in gallery

    Examples of SACI detection results for (a) clear; (b) overcast and partly cloudy with (c) optically thick clouds; and (d) optically thin clouds sky images, where the top row is the original image and the bottom row is the classified binary cloud maps. White pixels represent cloud, gray pixels represent sky, and black pixels represent the masked region.

  • View in gallery

    Plots of error distributions for (a) 5-, (b) 10-, and (c) 15-min forecasts. The y axis of the plots is in logarithmic scale.

  • View in gallery

    Sample time series of forecasted GHI and absolute errors for 10-min horizon forecast: (a) clear period on 9 Mar 2013, (b) overcast period on 2 Mar 2013, and (c) partly cloudy period on 1 Apr 2013.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 198 198 15
PDF Downloads 197 197 19

A Smart Image-Based Cloud Detection System for Intrahour Solar Irradiance Forecasts

View More View Less
  • 1 Department of Mechanical and Aerospace Engineering, Jacobs School of Engineering, and Center for Renewable Resource Integration, and Center for Energy Research, University of California, San Diego, La Jolla, California
© Get Permissions
Full access

Abstract

This study proposes an automatic smart adaptive cloud identification (SACI) system for sky imagery and solar irradiance forecast. The system is deployed using off-the-shelf fish-eye cameras that offer substantial advantages in terms of cost when compared to industry-standard sky imagers. SACI uses a smart image categorization (SIC) algorithm that combines the sky images and solar irradiance measurements to classify sky conditions into three categories: clear, overcast, and partly cloudy. A cloud detection scheme, optimized for each image category, is used to quantify cloud cover from the sky images. SACI is optimized and validated against manually annotated images. Results show that SACI achieves overall classification accuracy higher than 90% and outperforms reference cloud detection methods. Cloud cover retrieved by SACI is used as an input for an artificial neural network (ANN) model that predicts 1-min average global horizontal irradiance (GHI), 5-, 10-, and 15-min ahead of time. The performance of the ANN forecasting model is assessed in terms of common error statistics (mean bias and root-mean-square error) and in terms of forecasting skill over persistence. The model proposed in this work achieves forecasting skills above 14%, 18%, and 19% over the persistence forecast for 5-, 10-, and 15-min forecasts, respectively.

Corresponding author address: Carlos F. M. Coimbra, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA 92093. E-mail: ccoimbra@ucsd.edu

Abstract

This study proposes an automatic smart adaptive cloud identification (SACI) system for sky imagery and solar irradiance forecast. The system is deployed using off-the-shelf fish-eye cameras that offer substantial advantages in terms of cost when compared to industry-standard sky imagers. SACI uses a smart image categorization (SIC) algorithm that combines the sky images and solar irradiance measurements to classify sky conditions into three categories: clear, overcast, and partly cloudy. A cloud detection scheme, optimized for each image category, is used to quantify cloud cover from the sky images. SACI is optimized and validated against manually annotated images. Results show that SACI achieves overall classification accuracy higher than 90% and outperforms reference cloud detection methods. Cloud cover retrieved by SACI is used as an input for an artificial neural network (ANN) model that predicts 1-min average global horizontal irradiance (GHI), 5-, 10-, and 15-min ahead of time. The performance of the ANN forecasting model is assessed in terms of common error statistics (mean bias and root-mean-square error) and in terms of forecasting skill over persistence. The model proposed in this work achieves forecasting skills above 14%, 18%, and 19% over the persistence forecast for 5-, 10-, and 15-min forecasts, respectively.

Corresponding author address: Carlos F. M. Coimbra, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA 92093. E-mail: ccoimbra@ucsd.edu

1. Introduction

The rapid increase in electric power production from solar irradiation in recent times dictated a growing need for accurate tools to forecast solar irradiance and the power output of solar farms (Inman et al. 2013). Given that cloud cover is the most important factor that determines the amount of solar irradiance at ground level, forecasting tools for solar and power output depend on good characterizations of cloud cover (cloud opacity, cloud motion, etc.). Forecasting models for horizons ranging from 30 min to several hours often use cloud information retrieved from satellite images (Rossow and Schiffer 1991; Zhao and Di Girolamo 2006; Marquez et al. 2013a). However, satellite images currently available [such as images from National Oceanic and Atmospheric Administration (NOAA)’s Geostationary Operational Environmental Satellite] have limited spatial and temporal resolutions that prevent their use for intrahour forecast (Ghonima et al. 2012; Marquez and Coimbra 2013). To overcome this issue, ground-based sky imagers have been developed and deployed to capture local high-resolution sky images at high temporal resolutions, thus enabling intrahour irradiance and power output forecasts (Voss and Zibordi 1989; Souza-Echer et al. 2006; Long et al. 2006; Seiz et al. 2007; Calbo and Sabburg 2008; Cazorla et al. 2008; Chow et al. 2011; Marquez et al. 2013a,b; Marquez and Coimbra 2013; Chu et al. 2013).

In general, images captured by these systems are recorded as red–green–blue (RGB) images. On these images, cloud pixels have different RGB intensities. Based on this fact, many automatic cloud identification algorithms (Souza-Echer et al. 2006; Long et al. 2006; Yang et al. 2009; Heinle et al. 2010; Mantelli Neto et al. 2010; Li et al. 2011; Ghonima et al. 2012) have been developed to distinguish cloud from sky pixels. The accuracy of these algorithms can exceed 90%, depending on cloud cover type. However, most commercial sky imagers are costly and incorporate shading systems to block direct sunlight from the sensors. This increases the sensor lifetime but can reduce greatly the ability to track clouds near the sun (depending on the shading).

This study uses images obtained with a generic off-the-shelf, high-resolution fish-eye dome network camera. Based on availability, price, specifications, and quality, we chose a Vivotek camera (model FE8171V). The advantages of this camera include high-resolution, easy installation, low cost, and absence of moving parts. On the other hand, given that direct sunlight is not blocked, the circumsolar region is affected by glare caused by forward Mie scattering and also from light scattered from the dome. The image glare and scattered light affect negatively the accuracy of the cloud segmentation, as their RGB values can be very similar to the ones from clouds.

Several cloud identification methods [fixed threshold method (FTM), minimum cross entropy (MCE) method, and clear-sky library (CSL) method, discussed in section 3] were used to detect clouds based on the proposed new camera. However, their performance is very sensitive to the type of cloud cover. For instance, neither FTM nor MCE method is capable of addressing the image glare, while the CSL method is not appropriate for cloud detection during overcast periods. To overcome these issues, we propose a smart adaptive cloud identification (SACI) system that combines solar irradiance measurements synchronized with sky images.

The main idea of SACI is that by combining global horizontal irradiance (GHI) measurements and sky images, it will be possible to distinguish between image glare (mostly seen during clear time periods) and clouds. Based on GHI measurements obtained from a pyranometer, each image was classified into three categories: clear, overcast, and partly cloudy. Once the category of the sky image was determined, SACI used the most appropriate cloud detection scheme to classify the image pixels into either cloud or sky. The classified binary cloud map was then translated into numerical cloud indices by a grid-cloud fraction method (Marquez et al. 2013a) in order to forecast GHI. The GHI forecast was produced using artificial neural networks (ANNs) that use the cloud indices and GHI lagged data as inputs. The longest forecast horizon considered was 15 min. This limitation arises from the fact that for longer forecast horizons the most relevant information, in terms of cloud cover, is outside of the field of view of the camera.

This paper is organized as follows: Section 2 covers the deployment of cameras and pyranometers, the collection of images and irradiance data, and the separation of data for training and validation; section 3 covers the methods used to detect clouds and forecast GHI; section 4 covers the results and discussion for both cloud detection and GHI forecast; and, finally, section 5 presents the conclusions.

2. Data and instruments

a. Sky images

In this work we use sky images obtained from two fish-eye cameras featuring a 3.1-megapixel complementary metal oxide semiconductor (CMOS) sensor and a 360° panoramic view lens that were deployed in two locations in California (Merced and Folsom). The fish-eye cameras capture images every minute during daylight hours and transfer them, via FTP, to a web server where they are in a MySQL database. This process is fully automated, requiring no on-site staff; however, occasional cleaning of the fish-eye lens is necessary.

To develop the cloud detection algorithm, we used 80 images captured in Merced and 30 images captured in Folsom between 5 March and 5 April 2013. These images were selected manually and represent atmospheric conditions that can be categorized into three groups: clear, overcast, and partly cloudy. Each pixel in these images was manually classified into either clear or cloudy. These images were used as the ground truth in the assessment of the cloud detection algorithm accuracy.

The images from Merced were separated into a training set (50 images) and a validation set (30 images). All the images from Folsom were used as an independent validation set to study the generality of our models. The partition of the images by cloud cover type is listed in Table 1. The training set is used to optimize the various parameters (introduced below) for all investigated cloud identification methods. The optimized parameters were then used to evaluate both Merced and Folsom validation images.

Table 1.

Partitions of the selected images.

Table 1.

b. Irradiance data

To measure GHI, pyranometers have been deployed next to the cameras. Because of availability, a LI-COR LI-200 silicon photodiode pyranometer was installed in Folsom and an Eppley Laboratory precision spectral pyranometer (PSP) with a thermopile sensor was installed in Merced. The Eppley PSP is a secondary standard radiometer that meets the highest standards of the World Meteorological Organization radiometer characterizations. The LI-200 is less accurate but still meets the requirements for our purposes. For both instruments, data was logged as 1-min averages with a Campbell Scientific CR1000 datalogger. Data were accessed in real time and stored in a MySQL database.

c. Data for GHI forecast

For the GHI forecast, irradiance data and images were paired and collected in Folsom at 19 920 time points during the winter and spring of 2013 (13 January–2 April 2013). During these seasons there are sufficient data for clear, overcast, and partly cloudy periods. Data used for this study considered only periods when the solar elevation angle was larger than 30°. These data were randomly separated in two subsets: 70% as the learning set (14 000 time points) and 30% as the testing set (5920 time points). The learning set is used to optimize the ANN models, and the testing set is used to assess the performance of the ANN models.

3. Methods

a. SACI procedure

SACI applies separate cloud identification schemes from the literature according to the image categories to maximize the accuracy of cloud identification. With a proposed smart image categorization (SIC), SACI first classifies the sky image into three categories: clear, overcast, and partly cloudy. A clear period is defined as a period of time when clouds do not obscure the sun and the total sky cloud coverage is less than 5%. Overcast is defined as a period of time when the sun is obscured by clouds and the total sky cloud coverage is higher than 90%. The remaining data points are defined as partly cloudy. Based on the discussion above, SACI uses FTM for overcast images, the CSL method and FTM for clear images (CSL plus FTM), and the CSL method with the MCE method for partly cloudy images (CSL plus MCE). Figure 1 illustrates the workflow for SACI.

Fig. 1.
Fig. 1.

Schematic of SACI. Dashed-line square encloses the SIC.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

b. SIC

To apply the correct method, we developed an algorithm—SIC— that receives the images and the most recent GHI data and that decides which cloud identification method is applied. SIC integrates the hybrid thresholding algorithm (HYTA), which classifies images based on their normalized red–blue ratio (NRBR) histograms, with the clear-sky irradiance thresholding (CSIT) method, which uses the GHI time series to detect clear-sky periods. SIC classifies each input image into three atmospheric categories. First, CSIT identifies whether the image corresponds to a clear or cloudy period based on the irradiance. If the clear-sky irradiance thresholding returns true, then the image is categorized as clear. Otherwise, HYTA is used to analyze the NRBR histogram and classifies the images as overcast or partly cloudy.

CSIT is a clear-sky detection method (Long and Ackerman 2000; Younes and Muneer 2007; Reno et al. 2012) that uses five criteria computed from the irradiance time series:

  1. mean GHI (G),
  2. maximum GHI (M),
  3. length of the GHI time series (L)
    e1
  4. variance of GHI changes (σ)
    e2
    where k is the clear-sky indices, and
  5. maximum deviation from clear-sky GHI slope (ΔS)
    e3
    where the superscript clr represents the predicted GHI from clear-sky model.

The first and second criteria are the simplest way to differentiate a cloudy period from a clear-sky period. The third and fourth criteria identify the clear-sky periods based on the irradiance time series shape, which identify the increased irradiance variability caused by different cloud genres. The fifth criterion ensures that successive periods with cloud shadows occurring on the period boundaries are classified as nonclear (Reno et al. 2012).

In this study, CSIT calculates the five criteria for the portion of the GHI time series corresponding to the past 10 min. Then the criteria are compared to their clear-sky counterparts for the same period, which are computed with Ineichen’s clear-sky irradiance model (Ineichen 2008). The period is classified as clear when all the five criteria meet preset thresholds. The thresholds of CSIT applied in this study, listed in Table 2, are similar to those proposed by Reno et al. (2012).

Table 2.

Clear-sky irradiance thresholds (applied to a 10-min window). Symbol Δ represents the difference between the criteria of measured GHI and the criteria of clear-sky model. The criteria values selected by CSIT follow Reno et al. (2012). The period is classified as clear when all the five criteria meet the listed thresholds.

Table 2.

The clear-sky model used for this study is described in Ineichen and Perez (2002), Ineichen (2006), Gueymard and Wilcox (2011), and Inman et al. (2013), which is based on the following equation:
e4
where DNI stands for direct normal irradiance, I0 = 1360 W m−2 is the extraterrestrial irradiance, AM is the air mass derived based on solar geometry, δ is the total optical thickness for clear and dry atmosphere (cda) (depends on AM), and TL is an airmass independent Linke turbidity coefficient derived from broadband beam radiation measurements. Therefore, to calculate the clear-sky values, time, longitude, latitude, and average Linke turbidity values have to be used as an input. Linke turbidity values are available as monthly averages derived from global beam radiation measurements and are publicly available. For a detailed description of the calculation of Linke turbidity values, please see Ineichen and Perez (2002).

HYTA, proposed by Li et al. (2011), categorizes images into two groups based on their NRBR distribution histogram: unimodal or bimodal. Unimodal histograms, characterized by a single and sharp peak, correspond to uniform cloud cover (either clear or overcast); bimodal histograms, characterized by more than one peak and a large variance, correspond to partly cloudy sky. In general, the standard deviation of a unimodal image is significantly lower than that of a bimodal image. Thus, unimodal images can be differentiated from bimodal images by using a standard deviation threshold (SDT). Examples of the NRBR histograms for different cloud cover can be seen in Fig. 2. The overcast image yields a sharp peak in its NRBR histogram with a low standard deviation. The clear-sky image is characterized by an NRBR histogram with a high deviation because of the glare in the circumsolar region whose pixels have a high red–blue ratio (RBR) similar to cloud pixels. As a result, the standard deviation thresholding is unable to distinguish between clear and cloudy skies. This observation validates the introduction of the CSIT to discriminate between clear images and partly cloudy images.

Fig. 2.
Fig. 2.

Examples of (top) some original images and (bottom) their NRBR distribution histograms for the (left) clear, (middle) overcast, and (right) partly cloudy periods.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

The SIC thresholds employed in this work, which are based on suggestions in the literature (Long and Ackerman 2000; Younes and Muneer 2007; Ineichen 2008; Li et al. 2011; Reno et al. 2012; Marquez and Coimbra 2013), correctly classified the sky condition for each of the 110 sky images listed in Table 1. However, the SIC thresholds may be sensitive to different local microclimates and not universally applicable. Therefore, the SIC needs calibration with locally collected image and irradiance data when applied to a new location.

c. Cloud identification methods

FTM is based on the fact that cloud pixels (in an RGB image) have higher red (R) intensity values than sky pixels. Generally, the ratio (RBR = R/B) or difference (RBD = R − B) of red intensity to blue intensity is calculated for every pixel in the image and compared with a fixed threshold to determine whether the pixel corresponds to cloud or clear sky. An alternative to these two quantities is the NRBR [=(R − B)/(R + B)]. NRBR shows improved robustness to noise because it avoids extremely large RBRs when pixels have very low blue intensities (Li et al. 2011). For this reason, the FTM used in this study is based on the NRBR parameter. The threshold of FTM is estimated by maximizing the FTM cloud identification accuracy (defined below in section 3d) using the training set of images.

FTM performs very well for clear or overcast images, but it is incapable of detecting thin clouds such as cirriform (Long et al. 2006; Li et al. 2011). Yang et al. (2009) and Li et al. (2011) show that the MCE method achieved better performance than the FTM for cumuliform and cirriform cloud identification.

The MCE method is an adaptive thresholding method based on the Otsu algorithm (Otsu 1979; Li and Lee 1993; Li and Tam 1998; Li et al. 2011). Marquez and Coimbra (2013) also used the MCE method and concluded that the value of the MCE method threshold must be confined within an interval to maintain a satisfactory performance. The interval limits are estimated from the training set. Once the threshold is determined, pixels with higher RBR than that value are classified as cloudy.

When glare is present (see Fig. 3), the RBR of circumsolar pixels increases, leading FTM and the MCE method to misclassify those pixels as cloud. Glare is highly dependent upon solar geometry and can be mostly circumvented by utilizing the CSL method (Ghonima et al. 2012).

Fig. 3.
Fig. 3.

Examples of (top) original images, (middle) RBR images, and (bottom) CSL method–subtracted RBR images for the (left) clear, (middle) overcast, and (right) partly cloudy periods. Color bars indicate RBR magnitudes for the pixels in each image.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

The CSL method is a database of previously captured clear-sky images for different solar zenith angles. It is used to remove the geometric variation of clear-sky RBRs that depend on the sun-pixel angle and the solar zenith angle. This is accomplished by subtracting or offsetting the RBR of the input image by the reference CSL RBR that corresponds to the same zenith angle, generating the Diff image: Diff = RBR minus CSL. The cloud detection methods can then be applied to the DIFF image. For the overcast images, the circumsolar region cloud pixels are likely to be misidentified as clear pixels because of the subtraction process of the CSL method, as shown in Fig. 3 (middle column).

A second-order effect on the solar irradiance is caused by aerosol, quantified by the aerosol optical depth. This effect is not considered in the CSL method, but it also affects the clear-sky RBRs and the accuracy of cloud identification. To account for this effect, a haze correction factor (HCF) can be used to correct the images, DiffHCF = RBR − (CSL)(HCF), before applying the cloud identification. The HCF is computed iteratively (Seiz et al. 2007; Ghonima et al. 2012):

  1. Preliminary cloud identification is performed with the CSL method, and HCF is set to 1.
  2. The ratio of sky mean RBR to the CSL method mean RBR (Prt) for each pixel is obtained as the temporary HCF.
  3. The CSL method is multiplied by the Prt and another round of cloud identification is initiated.
  4. Steps 2 and 3 are repeated until Prt converges below a threshold (which is set to 0.01 in this study).
  5. The converged Prt is returned as the HCF.
Examples of DiffHCF are shown in Fig. 3 (bottom row).

d. Cloud detection performance

Each pixel from the cloud classification image is compared to its counterpart in the manually annotated cloud map. Using the confusion matrix shown in Table 3 for sky and cloud pixels, we quantify the hits and misses in all the pixels in the image. Based on these values, the accuracy is defined as the percentage of pixels that are correctly classified:
e5
Table 3.

Confusion matrix of automatic cloud detection for each pixel.

Table 3.

e. Grid-cloud fraction method

Clouds contribute differently to the near-future behavior of GHI. For instance, clouds moving toward the sun are much more relevant to the forecasting models than clouds moving away from the sun. To extract information of clouds that may shade the sun in the near future, we use the grid-cloud fraction method developed by Marquez and Coimbra (2013).

This method can be summarized in four main steps. First, all images are projected from the convex mirror space onto a flat rectangular space (shown in Fig. 4b) to remove the geometric distortion of the images. Second, pairs of consecutive images are processed using particle image velocimetry (PIV) to compute the flow direction of the clouds. The PIV algorithm (Adrian and Westerweel 2010) partitions the image into interrogation windows (32 pixels × 32 pixels in this case). The correlation between two consecutive interrogation windows is analyzed through the minimum quadratic difference method to determine the “particles” (the clouds) displacement. The velocity field (shown in Fig. 4b) is obtained by dividing the displacements by the separation time of the images and clustering using the k-means clustering method. The velocity magnitudes usually cluster around two means: one near or equal to zero and the other with a larger magnitude. The first result is from windows where there are no clouds or where they are stationary and is therefore disregarded. The second result is selected as the representative velocity for the cloud motion. Currently, our algorithm provides binary cloud maps and a single cloud motion vector. Therefore, it is unable to consider multiple layer cases of clouds. Noise created by the PIV algorithm is addressed by clustering all nonzero grid velocities to obtain a representative velocity. Moreover, based on the assumption that the cloud flow does not significantly change in a short time, we use as the representative velocity the median speed (for both x and y directions) of the past 10 min. More details about the method for the calculation of the cloud motion vector can be found in Adrian and Westerweel (2010), Marquez and Coimbra (2013), and Chu et al. (2013).

Fig. 4.
Fig. 4.

Main steps of the grid cloud fractions method. (a),(b) Original consecutive image pair used in the analysis. (c),(d) Image pair is projected onto a rectangular grid, and (e) velocity vectors computed by the PIV algorithm. (f) Cloud classified image, where white pixels represent cloud; gray pixels represent clear sky; and black pixels are obstacles, which are excluded from the analysis.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

To qualitatively study the accuracy and robustness of PIV in computing the cloud moving direction, we compared the cloud moving directions from PIV with manually identified cloud moving directions on seven days. An example of the comparison is illustrated in Fig. 5. We found that, in about 95% of the tested instances, the PIV-identified cloud moving directions are in the error range of the manually identified directions.

Fig. 5.
Fig. 5.

Example time series of the PIV and the manually identified cloud moving directions on 1 Apr 2013. Error range (marked as yellow) of the manually identified directions is 20°. PIV-identified directions that locate in the yellow region are considered accurate tested instances.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

Third, the SACI is used to classify the sky images into binary cloud images. A set of grid elements (X1, X2, …, X6) originating from the sun’s position and oriented in the reverse direction of the representative velocity for the clouds is placed over the binary cloud images (shown in Fig. 4c) In this work we set the area of grid elements to 100 pixels × 100 pixels. Finally, the cloud indices CIi are computed as the percentage of pixels classified as cloud for element i. This process is applied to all the images under study, generating CIi time series that are used as potential inputs for the ANN-based forecasting models.

f. Artificial neural network

ANNs are widely used nonlinear regression tools for irradiance forecasts that require no prior assumptions concerning data relationships (Bishop 1994; López et al. 2005). In this study, we use a multilayer feed-forward perceptron that is widely used in forecasting and modeling of solar irradiance (Mellit and Kalogirou 2008; Mellit and Pavan 2010; Marquez and Coimbra 2011; Pedro and Coimbra 2012; Marquez et al. 2013b). ANNs use interconnected signal processing elements, which are called neurons. Neurons are placed in layers, and the layers between the input layer (the first) and the output layer (the last) are called hidden layers. Neurons receive weighted inputs incoming from the previous layers and add a bias or threshold to the sums. After that, the sums are processed by the activation function of the neurons (sigmoidal function in this study) to generate outputs. The outputs Yi are then used as the inputs for neurons on the following layer:
e6
where X and Y are the input and output, respectively; w is the weight; β is the bias; and f is the activation function. Once the ANN structure (the number of layers and neurons) is established, several free parameters corresponding to the weight and bias values in each neuron are found in the training process. The training adjusts these parameters seeking to minimize the error between the ANN’s outputs and the targets. This is an iterative process that stops once discrepancy is lower than a preset value. In this work, the Bayesian regularization process that uses the Levenberg–Marquardt optimization is used to train the ANNs. This operation is represented by
e7
where ANNu represents the untrained feed-forward neural network with L hidden layers and N neurons per layer. The I represents the training inputs and T represents the training targets in the training process.

In this study, two ANN models (ANNc and ANNnc) were developed. ANNc considers as potential inputs the time-lagged measured GHI values, ranging from 0 to 20 min in 5-min steps, the total sky cloudiness, and the CIi extracted by the grid fraction method. ANNnc, which works as a control group, considers only time-lagged measured GHI values.

g. ANN validation

Overfitting is a possible problem of ANN training. Overfitting manifests when the optimal solution found for the training data performs poorly when applied to new data. In this work, the cross-validation method (CVM) (Efron and Gong 1983; Geisser 1993; Kohavi 1995; Jain et al. 2000; Lendasse et al. 2003), a popular method in model estimation, is used to determine whether a hypothesis ANN model fits the data of interest. CVM divides the learning data into K disjoint subsets Si with i = 1, …, K [K is set to 10 as suggested in previous studies (Kohavi 1995; McLachlan et al. 2004)]. After the data partition, one subset is reserved for validation and the remaining subsets are used to train the ANN model. Root-mean-square error (RMSE) is taken as the performance of the model on the validation set. This process is repeated K times, each time taking a different subset as the validation set and returning an RMSE. The mean and variance of the K RMSEs are taken as the CVM validation performance. Models with high mean and variance of the validation RMSEs are very dependent on the training data and therefore prone to be overfitted.

h. ANN optimization

The performance of the forecasting model depends greatly on its input variables and the ANN topology (e.g., number of layers, number of neurons per layer). To obtain the best possible forecast, we use a genetic algorithm (GA) to determine these free parameters. GA is an efficient tool (Holland 1992; Pai and Hong 2005; Marquez and Coimbra 2011; Pedro and Coimbra 2012) that iteratively scans potential solutions (individuals) of the search space to find the optimal solution. GA optimization is initiated with randomly selected individuals (the population) whose fitness is evaluated according to an objective function. In this study, the GA optimization objective function is the average validation RMSE. The individuals with the lowest validation RMSE are selected as parents and generate the new generation through crossover and mutation. This iterative process is finished when the average fitness of the population converges.

i. Evaluation of the GHI forecast

The quality of the forecasting is benchmarked against the persistence model. The persistence model is the simplest forecasting model and is very accurate for low-variability periods. To remove the deterministic diurnal solar variation from the persistence forecast, we used the clear-sky index persistence model as a baseline. This persistence model assumes that the clear-sky index of GHI remains constant between t and t + Δt, resulting in the forecasting
e8
where the subscript p denotes persistence and Gclr is provided by the clear-sky model used above in the CSIT.
The forecast performance of the model is assessed using three statistical metrics: mean biased error (MBE),
e9
RMSE,
e10
and forecast skill (s), which measures the improvement of the investigated forecast over the reference persistence model:
e11

4. Results and discussion

a. Cloud detection training

The training set, which contains 50 manually annotated images obtained at Merced, is used to determine the optimal thresholds for the SACI algorithm. These are 1) the threshold of FTM, 2) the threshold of CSL plus FTM, 3) the upper and lower bounds of the MCE method threshold, 4) the upper and lower bounds of the CSL-plus-MCE threshold, and 5) the SDT for the hybrid thresholding. Because identifying cirriform/cumuliform clouds is of primary interests in many studies, particularly for power generation forecasts, these parameters are optimized to maximize the overall accuracy for partly cloudy images. The optimal thresholds determined on the training set are shown in Table 4.

Table 4.

Optimal thresholds that maximize the cloud identification accuracy on the learning dataset. The third column represents the threshold values used by the methods listed in the second column. SACI integrates the three reference methods, and the SDT represents the standard deviation threshold for the HYTA method. All thresholds are dimensionless.

Table 4.

b. Cloud detection validation

Images from the validation set were used to evaluate all cloud identification models with the optimized thresholds listed in Table 4. The accuracy for the different models applied to the validation images from Merced and Folsom is listed in Table 5.

Table 5.

Average accuracy (%) for the validation images using Eq. (5). The values inside the parenthesis are the standard deviation. Boldface values indicate best performance in each category.

Table 5.

The validation results show that the SACI categorizes all validation images accurately and achieves the highest overall accuracy for different weather conditions. FTM and the MCE method cannot distinguish glare from clouds during the clear period, while the CSL-based methods misclassified nearly half of the pixels in the overcast images. Given that SACI was optimized using only images from Merced and that it achieves similar accuracy when applied to Folsom (see Table 5), we can conclude that this method performs consistently, independently of the location.

Cloud misclassification is usually due to three reasons: 1) saturated pixels in the circumsolar region because of glare that are not completely offset by the CSL method, 2) diffraction of sunlight on the glass dome, and 3) dirt specks that diffract the sunlight intensively during the clear time. An example of clear-sky SACI cloud detection is shown in Fig. 6a. The average accuracy of SACI for overcast images is 94%. Most misclassified pixels are located near the edge of the images (close to the horizon), where R intensities are not as high as those in the center of the images during overcast periods, which can be seen in Fig. 6b.

Fig. 6.
Fig. 6.

Examples of SACI detection results for (a) clear; (b) overcast and partly cloudy with (c) optically thick clouds; and (d) optically thin clouds sky images, where the top row is the original image and the bottom row is the classified binary cloud maps. White pixels represent cloud, gray pixels represent sky, and black pixels represent the masked region.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

Partly cloudy images usually contain both thick and thin clouds. Thick cloud pixels generally have an RBR significantly higher than clear pixels and therefore are detected easier by cloud detection algorithms. When the majority of clouds in an image are optically thick, the algorithm is, on average, over 95% accurate. An example can be seen in Fig. 6c. On the other hand, thin clouds, whose RBRs are close to or overlap clear RBRs, are generally more difficult to be identified correctly. Therefore, thin cloud pixels generally have low identification accuracy, and the SACI algorithm is, on average, approximately 85% accurate. Most of the misclassified pixels are false negative cases, which indicate that the algorithm is more likely to classify thin cloud pixels as sky pixel. An example of thin cloud detection is shown in Fig. 6d.

c. GHI forecast results

The GA algorithm takes 50 generations until the average fitness converges. The optimal set of input variables, hidden layers, and the number of neurons per layer selected are shown in the Table 6. These results show that the total sky cloud coverage and cloud indices are useful for GHI forecast. Table 6 also shows that the last measured values of GHI are selected as forecast inputs for all horizons. This indicated that the latest GHI values are highly informative for short-term irradiance forecast.

Table 6.

ANN inputs and parameters selected by the GA for the GHI forecast model. Indices represent selected ANN inputs: 1–5 are the time-lagged GHI values, 6 is the total sky cloudiness, and 7–11 are the CIi extracted by the grid fraction method. The number of hidden layers is represented by L, and N represents the number of neurons per hidden layer.

Table 6.

The GA optimized ANN models are evaluated on the testing set that is independent of the training process and compared to the persistence model. The forecasting performance metrics are presented in Table 7. In terms of bias, all models exhibit small values. In terms of RMSE and forecast skill, both ANN models significantly outperform the persistence. The results show that ANNc benefits from integrating the cloud cover information, achieving significantly higher forecast skill than the ANNnc. Improvements over the ANNnc range from 2.9% to 8.1% depending on the forecast horizon.

Table 7.

Statistical error metrics for testing results of the different forecast models for 5-, 10-, and 15-min horizons. Bolded numbers identify the best-performing method for a given metric. MBE and RMSE are in W m−2 and s is in %

Table 7.

To further understand the ANNc model’s performances, its error distributions for the 5-, 10- and 15-min horizons are plotted in Fig. 7. These figures show that in the low-error region (between −0.1 and 0.1 kW m−2), both persistence and ANNc have sharp peaks near the 0. However, the persistence forecast has a higher frequency of errors with high absolute values (>0.2 kW m−2), resulting in a longer and heavier error distribution tail. Compared to the persistence error distribution, the ANNc reduces the occurrence of high absolute value errors, particularly for negative errors, and produces an error distribution with shorter tails and rounder shoulders. Consequently, the distributions of ANNc forecast errors are narrower, resulting in lower RMSEs. Sample time series of forecasted GHI and absolute errors for 10-min horizon forecast are plotted in Fig. 8.

Fig. 7.
Fig. 7.

Plots of error distributions for (a) 5-, (b) 10-, and (c) 15-min forecasts. The y axis of the plots is in logarithmic scale.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

Fig. 8.
Fig. 8.

Sample time series of forecasted GHI and absolute errors for 10-min horizon forecast: (a) clear period on 9 Mar 2013, (b) overcast period on 2 Mar 2013, and (c) partly cloudy period on 1 Apr 2013.

Citation: Journal of Atmospheric and Oceanic Technology 31, 9; 10.1175/JTECH-D-13-00209.1

5. Conclusions

A high-resolution fish-eye dome network camera is used as a low-cost alternative to sky imagers for cloud and sky condition identification. The main advantages of the sky camera over sky imagers are resolution, portability, absence of moving parts, and substantially lower cost than sky imagers. The main drawbacks of using this type of camera for sky imaging relate to excessive glare. Nonetheless, we presented a number of methods in this work that help circumvent this deficiency.

A smart adaptive cloud identification (SACI) system was proposed and deployed. This system integrates the fixed thresholding method, minimum cross entropy thresholding method, and the use of a clear-sky library. The overall methods addresses glaring caused by forward Mie scattering and lens flaring solely through image processing algorithms. SACI uses SIC, which analyzes global horizontal irradiance data and normalized red-to-blue ratio distribution to categorize the input images. The method then employs an optimal cloud detection scheme for each categorized image. Optimization is performed on manually annotated images, and the validation tests for different locations show that SACI achieves robust, location independent, and the most accurate cloud classification among all the reference models analyzed. Mean accuracy of this system is over 92%, 94%, and 89% for clear, overcast, and partly cloudy images, respectively.

By using the numerical cloud indices extracted from the binary cloud classified images, an ANN model is developed for 1-min-average GHI forecasts for 5-, 10-, and 15-min horizons. Application of the SACI-ANN model to an independent testing set shows that the proposed forecasting methodology achieves forecasting skill of 14.4%, 18.4%, and 19.7% for 5-, 10-, and 15-min horizons, respectively.

Acknowledgments

Partial support from the National Science Foundation (NSF) EECS-EPAS Award 1201986 (managed by Dr. Paul Werbos) is gratefully acknowledged.

REFERENCES

  • Adrian, R. J., , and Westerweel J. , 2010: Particle Image Velocimetry. Cambridge Aerospace Series, Vol. 30, Cambridge University Press, 586 pp.

  • Bishop, C. M., 1994: Neural networks and their applications. Rev. Sci. Instrum., 65, 18031832, doi:10.1063/1.1144830.

  • Calbo, J., , and Sabburg J. , 2008: Feature extraction from whole-sky ground-based images for cloud-type recognition. J. Atmos. Oceanic Technol., 25, 34, doi:10.1175/2007JTECHA959.1.

    • Search Google Scholar
    • Export Citation
  • Cazorla, A., , Olmo F. J. , , and Alados-Arboledas L. , 2008: Development of a sky imager for cloud cover assessment. J. Opt. Soc. Amer., 25A, 2939, doi:10.1364/JOSAA.25.000029.

    • Search Google Scholar
    • Export Citation
  • Chow, C. W., , Urquhart B. , , Lave M. , , Dominguez A. , , Kleissl J. , , Shields J. , , and Washom B. , 2011: Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed. Sol. Energy, 85, 28812893, doi:10.1016/j.solener.2011.08.025.

    • Search Google Scholar
    • Export Citation
  • Chu, Y., , Pedro H. T. C. , , and Coimbra C. F. M , 2013: Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning. Sol. Energy, 98, 592603, doi:10.1016/j.solener.2013.10.020.

    • Search Google Scholar
    • Export Citation
  • Efron, B., , and Gong G. , 1983: A leisurely look at the bootstrap, the jackknife, and cross-validation. Amer. Stat., 37, 3648.

  • Geisser, S., 1993: Predictive Inference. Monogr. on Statistics and Applied Probability, Vol. 55, Chapman & Hall/CRC, 31–42.

  • Ghonima, M. S., , Urquhart B. , , Chow C. W. , , Shields J. E. , , Cazorla A. , , and Kleissl J. , 2012: A method for cloud detection and opacity classification based on ground based sky imagery. Atmos. Meas. Tech. Discuss., 5, 45354569, doi:10.5194/amtd-5-4535-2012.

    • Search Google Scholar
    • Export Citation
  • Gueymard, C. A., , and Wilcox S. M. , 2011: Assessment of spatial and temporal variability in the US solar resource from radiometric measurements and predictions from models using ground-based or satellite data. Sol. Energy, 85, 10681084, doi:10.1016/j.solener.2011.02.030.

    • Search Google Scholar
    • Export Citation
  • Heinle, A., , Macke A. , , and Srivastav A. , 2010: Automatic cloud classification of whole sky images. Atmos. Meas. Tech., 3, 557567, doi:10.5194/amt-3-557-2010.

    • Search Google Scholar
    • Export Citation
  • Holland, J. H., 1992: Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence. MIT Press, 211, pp.

  • Ineichen, P., 2006: Comparison of eight clear sky broadband models against 16 independent data banks. Sol. Energy, 80, 468478, doi:10.1016/j.solener.2005.04.018.

    • Search Google Scholar
    • Export Citation
  • Ineichen, P., 2008: A broadband simplified version of the Solis clear sky model. Sol. Energy, 82, 758762, doi:10.1016/j.solener.2008.02.009.

    • Search Google Scholar
    • Export Citation
  • Ineichen, P., , and Perez R. , 2002: A new airmass independent formulation for the Linke turbidity coefficient. Sol. Energy, 73, 151157, doi:10.1016/S0038-092X(02)00045-2.

    • Search Google Scholar
    • Export Citation
  • Inman, R., , Pedro H. , , and Coimbra C. , 2013: Solar forecasting methods for renewable energy integration. Prog. Energy Combust. Sci., 39, 535576, doi:10.1016/j.pecs.2013.06.002.

    • Search Google Scholar
    • Export Citation
  • Jain, A. K., , Duin R. P. W. , , and Mao J. , 2000: Statistical pattern recognition: A review. IEEE Trans. Pattern Anal. Mach. Intell., 22, 437, doi:10.1109/34.824819.

    • Search Google Scholar
    • Export Citation
  • Kohavi, R., 1995: A study of cross-validation and bootstrap for accuracy estimation and model selection. IJCAI ’95: Proceedings of the 14th International Joint Conference on Artificial Intelligence, Vol. 2, C. S. Mellish, Ed., Morgan Kaufmann, 1137–1145.

  • Lendasse, A., , Wertz V. , , and Verleysen M. , 2003: Model selection with cross-validations and bootstraps—Application to time series prediction with RBFN models. Artificial Neural Networks and Neural Information Processing—ICANN/ICONIP 2003, O. Kaynak et al., Eds., Lecture Notes in Computer Science, Vol. 2714, Springer, 573–580.

  • Li, C. H., , and Lee C. K. , 1993: Minimum cross entropy thresholding. Pattern Recognit., 26, 617625, doi:10.1016/0031-3203(93)90115-D.

  • Li, C. H., , and Tam P. K. S. , 1998: An iterative algorithm for minimum cross entropy thresholding. Pattern Recognit. Lett., 19, 771776, doi:10.1016/S0167-8655(98)00057-9.

    • Search Google Scholar
    • Export Citation
  • Li, Q., , Lu W. , , and Yang J. , 2011: A hybrid thresholding algorithm for cloud detection on ground-based color images. J. Atmos. Oceanic Technol., 28, 12861296, doi:10.1175/JTECH-D-11-00009.1.

    • Search Google Scholar
    • Export Citation
  • Long, C. N., , and Ackerman T. P. , 2000: Identification of clear skies from broadband pyranometer measurements and calculation of downwelling shortwave cloud effects. J. Geophys. Res., 105, 15 60915 626, doi:10.1029/2000JD900077.

    • Search Google Scholar
    • Export Citation
  • Long, C. N., , Sabburg J. M. , , Calbó J. , , and Pages D. , 2006: Retrieving cloud characteristics from ground-based daytime color all-sky images. J. Atmos. Oceanic Technol., 23, 633652, doi:10.1175/JTECH1875.1.

    • Search Google Scholar
    • Export Citation
  • López, G., , Batlles F. J. , , and Tovar-Pescador J. , 2005: Selection of input parameters to model direct solar irradiance by using artificial neural networks. Energy, 30, 16751684, doi:10.1016/j.energy.2004.04.035.

    • Search Google Scholar
    • Export Citation
  • Mantelli Neto, S. L. M., , von Wangenheim A. , , Pereira E. B. , , and Comunello E. , 2010: The use of Euclidean geometric distance on RGB color space for the classification of sky and cloud patterns. J. Atmos. Oceanic Technol., 27, 15041517, doi:10.1175/2010JTECHA1353.1.

    • Search Google Scholar
    • Export Citation
  • Marquez, R., , and Coimbra C. F. M. , 2011: Forecasting of global and direct solar irradiance using stochastic learning methods, ground experiments and the NWS database. Sol. Energy, 85, 746756, doi:10.1016/j.solener.2011.01.007.

    • Search Google Scholar
    • Export Citation
  • Marquez, R., , and Coimbra C. F. M. , 2013: Intra-hour DNI forecasting methodology based on cloud tracking image analysis. Sol. Energy, 91, 327336, doi:10.1016/j.solener.2012.09.018.

    • Search Google Scholar
    • Export Citation
  • Marquez, R., , Gueorguiev V. G. , , and Coimbra C. F. M. , 2013a: Forecasting of global horizontal irradiance using sky cover indices. J. Sol. Energy Eng.,135, 011017, doi:10.1115/1.4007497.

  • Marquez, R., , Pedro H. T. C. , , and Coimbra C. F. M. , 2013b: Hybrid solar forecasting method uses satellite imaging and ground telemetry as inputs to ANNs. Sol. Energy, 92, 176188, doi:10.1016/j.solener.2013.02.023.

    • Search Google Scholar
    • Export Citation
  • McLachlan, G. J., , Do K.-A. , , and Ambroise C. , 2004: Analyzing Microarray Gene Expression Data. Wiley Series in Probability and Statistics, Vol. 422, Wiley-Interscience, 270–285 pp.

  • Mellit, A., , and Kalogirou S. A. , 2008: Artificial intelligence techniques for photovoltaic applications: A review. Prog. Energy Combust. Sci., 34, 574632, doi:10.1016/j.pecs.2008.01.001.

    • Search Google Scholar
    • Export Citation
  • Mellit, A., , and Pavan A. M. , 2010: A 24-h forecast of solar irradiance using artificial neural network: Application for performance prediction of a grid-connected PV plant at Trieste, Italy. Sol. Energy, 84, 807821, doi:10.1016/j.solener.2010.02.006.

    • Search Google Scholar
    • Export Citation
  • Otsu, N., 1979: A threshold selection method from gray-level histograms. IEEE Trans. Syst. Man Cybern., 9, 6266, doi:10.1109/TSMC.1979.4310076.

    • Search Google Scholar
    • Export Citation
  • Pai, P. F., , and Hong W. C. , 2005: Forecasting regional electricity load based on recurrent support vector machines with genetic algorithms. Electr. Power Syst. Res., 74, 417425, doi:10.1016/j.epsr.2005.01.006.

    • Search Google Scholar
    • Export Citation
  • Pedro, H. T. C., , and Coimbra C. F. M. , 2012: Assessment of forecasting techniques for solar power production with no exogenous inputs. Sol. Energy, 86, 20172028, doi:10.1016/j.solener.2012.04.004.

    • Search Google Scholar
    • Export Citation
  • Reno, M. J., , Hansen C. W. , , and Stein J. S. , 2012: Global horizontal irradiance clear sky models: Implementation and analysis. Sandia National Laboratories Rep. SAND2012-2389, 67 pp. [Available online at http://energy.sandia.gov/wp/wp-content/gallery/uploads/SAND2012-2389_ClearSky_final.pdf.]

  • Rossow, W. B., , and Schiffer R. A. , 1991: ISCCP cloud data products. Bull. Amer. Meteor. Soc., 72, 220, doi:10.1175/1520-0477(1991)072<0002:ICDP>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Seiz, G., , Shields J. , , Feister U. , , Baltsavias E. P. , , and Gruen A. , 2007: Cloud mapping with ground-based photogrammetric cameras. Int. J. Remote Sens., 28, 20012032, doi:10.1080/01431160600641822.

    • Search Google Scholar
    • Export Citation
  • Souza-Echer, M. P., , Pereira E. B. , , Bins L. S. , , and Andrade M. A. R. , 2006: A simple method for the assessment of the cloud cover state in high-latitude regions by a ground-based digital camera. J. Atmos. Oceanic Technol., 23, 437447, doi:10.1175/JTECH1833.1.

    • Search Google Scholar
    • Export Citation
  • Voss, K., , and Zibordi G. , 1989: Radiometric and geometric calibration of a visible spectral electro-optic “fisheye” camera radiance distribution system. J. Atmos. Oceanic Technol., 6, 652662, doi:10.1175/1520-0426(1989)006<0652:RAGCOA>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Yang, J., , W. , , Ma Y. , , Yao W. , , and Li Q. , 2009: An automatic ground based cloud detection method based on adaptive threshold. J. Appl. Meteor. Sci., 20, 713721.

    • Search Google Scholar
    • Export Citation
  • Younes, S., , and Muneer T. , 2007: Clear-sky classification procedures and models using a world-wide data-base. Appl. Energy, 84, 623645, doi:10.1016/j.apenergy.2007.01.016.

    • Search Google Scholar
    • Export Citation
  • Zhao, G., , and Di Girolamo L. , 2006: Cloud fraction errors for trade wind cumuli from EOS-Terra instruments. Geophys. Res. Lett., 33, L20802, doi:10.1029/2006GL027088.

    • Search Google Scholar
    • Export Citation
Save