Implementation of an intelligent sensor formeasurement and prediction of solar radiation and

atmospheric temperatureJoao M. Gomes∗,Pedro M. Ferreira†, Antonio E. Ruano‡

∗,‡Centre for Intelligent Systems - IDMEC†Algarve STP, Algarve Science & Technology Park

‡University of Algarve8005-139 Faro, Portugal

Email: ∗ joaomealhagomes@gmail.com, † pfrazao@ualg.pt, ‡ aruano@ualg.pt

Abstract— The aim of this study was to develop an intelligentsensor for acquiring temperature, solar radiation data andestimate cloudiness indexes, and use these measured values topredict temperature and solar radiation in a close future. Theprototype produced can ultimately be used in systems related tothermal comfort in buildings and to the efficient and intelligentuse of solar energy.

To incorporate these functionalities, a small and portableprototype was developed, which consisted in: a CCTV camerawith a fish-eye lens, for sky images acquisition; a computerof format mini-itx with a Linux operative system, for dataacquisition and processing; a GPS, to enable automatic use, inde-pendent of the system’s geographical position; a pyranometer, forregular measurements of solar radiation; a temperature probe,for regular measurements of outdoor temperature; a shadowband, to eliminate the sun’s flare effect on sky images; Arduino,an open source electronics prototyping platform that acquiresdata from the temperature and solar radiation sensors, as wellas processing the data provided by the GPS and controlling theshadow band; neural networks of the type NARX, which use theacquired data to forecast the cloudiness index, solar radiationand temperature, in the next four hours period.

The system was programmed to acquire data, both from thesensors and the camera, every five minutes.

Keywords: temperature, solar radiation, cloudiness indexes,neural networks.

I. INTRODUCTION AND MOTIVATION

Nowadays there is a global need for decreasing our energyconsumption, to ensure that what we get from our naturalresources is efficiently used, and that we live in balance withthe environment. Increasingly so, efforts in R&D are directedtowards this goal. A major field consists in developing intel-ligent systems capable of integrating environmental data for amore efficient use of resources and sustainable functioning ofman-made utilities. In this study, the main focus is on globalsolar radiation, as it influences the majority of living beingsin many different ways. Thus an accurate prediction of itsevolution in time is important for several different areas ofapplication, such as the field of renewable energy (see [1]for an extensive review), people’s comfort in buildings, where

possible applications are luminescence control and thermalcomfort as [2], among others, proved to be relevant. In the fieldof meteorological data acquisition/estimation, more specif-ically cloudiness indexes, solar radiation and atmospherictemperature, there is room for further developments. First,there are not a lot of different systems which incorporate,as a package, sky pictures and sensor data. No commercialexisting solutions that take photographs for cloudiness studiesare portable in a way that they do not allow an easy assembleon the spot. Furthermore, and more importantly, no portablecommercial solution incorporates the possibility of forecastingthe estimated of measured variables.

This study follows a previous work done [2], while theestimation is an evolution of [3] and [4]. The starting pointwas the TSI440 from Yankee Environmental Systems, Inc.[5],as it only takes photographs of the sky but does not incor-porate any type of sensor, thereby lacking in environmentalinformation. This study arises from the need to have a devicethat centralizes a set of tasks - estimates the cloudinessindex using photographs of the sky, measures the globalsolar radiation, measures the atmospheric temperature andpredicts the evolution of these values within an horizon offour hours. The previous work [2] focuses in the prediction ofsolar radiation for 30 minutes in future, using multi-objectivegenetic algorithms [6] for the off-line design of radial basisfunction neural networks.

The existing research in cloudiness indexes differs in theirgoal, in how data is obtained and how it is processed. Someare concerned with the type of clouds being observed, othersonly with the amount of radiation filtered. As for data acqui-sition, the most used technologies are: satellites, ground all-sky imaging systems and spectroradiometers. Thus for clouddetection, satellites tend to be more used, as shown by [7],[8],[9] and [10]. Nevertheless, ground to sky imagers tend to bea cheap alternative allowing images with greater resolutions,the frequency between each new image being higher than thesatellites and also providing more focused results, as shownby [11], [12], [13], [14], [15] and [16].

978-1-4577-1402-3/11/$26.00 ©2011 IEEE

Solar radiation estimation is not a trivial subject, speciallyif zones with a complex topology are considered, such asmountain cases, where it is difficult to support economicallyand maintain the equipment. Another problem in this type ofterrains is the number of variables involved; these require adenser neural network. There are a few studies in this area,such as [17], who performed a study in which the performanceof active and passive energy systems are analysed and [18],who developed a model of daily solar irradiance estimatedfrom air temperature and precipitation data. Other studiesalso explore the benefit of using neural network instead ofan empirical model, such as the study performed by [19],which considers also two classical approaches to simulatethe propagation of solar radiation on earth surface developedby [20]. Besides the classical methods, the solar radiationcould be estimated by radiation measurements interpolation,as studied by [21], or by satellite images analysis, as studiedby [22]. In this study, the solar radiation forecast followsprecisely the used methodology in the study of [23]. Similarlyto the previous mentioned studies [2], Multi-Objective GeneticAlgorithm (MOGA)[6] was employed in order to identify theANN’s input-output structure.

Similarly to solar radiation, neural networks are often usedto estimate meteorological data values in areas with sparsemeteorological stations, using also atmospheric temperaturedata, as shown by [24], [25] and [26].

II. NEURAL NETWORKS

Though no neural network model was developed in thepresent study, the ones developed by [23] were used. Theneural network set was identified using MOGA to optimizetheir structure in order to develop a non-linear autoregres-sive with exogenous input model type of neural network(NARX). The networks were trained using a retro-propagationalgorithm based on Levenberg-Marquardt’s optimisation algo-rithm ([27];[28]) with early stopping as a stopping criteria.The images are processed by the ANN Cloudiness Estimatorto estimate the amount of cloudiness of each image. Thecloudiness along with the temperature and solar radiationmeasurements, are fed into the respective neural networkmodels in order to make a prediction up to 48 steps ahead.Since each step is five minutes, 48 steps correspond to a fourhours horizon. The complete cascade of models consist of: oneNAR cloudiness predictive model; one NARX solar radiationpredictive model, using cloudiness as the exogenous input;and a NARX temperature model, using solar radiation as theexogenous input.

III. PROTOTYPE

This project took as a reference an industrial total skyimager used in the Centro de Sistemas Inteligentes (CSI) lab-oratory in UALg, the TSI440A from Yankee EnvironmentalSystems, Inc. The TSI structure has two parts, a chrome-platedsteel mirror on which the whole sky is reflected on, and an armthat holds an Axis200+ ip camera that captures the reflection

of the sky in the mirror. The acquisition of the images is doneby accessing the ip camera’s web server.

A prototype was developed (see Figure 1) to performmeasurements of global solar radiation and atmospheric tem-perature and to take photographs of the sky, and then feed thedata collected to a set of neural networks [23], in order topredict the atmospheric temperature and solar radiation in afour hours horizon.

A logic module was used to make the connection betweenthe solar radiation and temperature sensors and the computerwhich implemented the ANNs, as well as to control theshadow band and process the data retrieved from the GPS.The device used for these tasks, is an open-source electronicsprototyping platform called Arduino R©.

To measure cloudiness indexes, it was necessary to use acamera that can photograph the largest possible area of thesky. In order to be compact, the prototype was designed totake the photographs by aiming the camera at the zenith, sincethis option would require a smaller amount of space and doesnot require a special structure to fix it. These measurementstook two implementations. The first consisted on the cameraaimed directly at the zenith without any protection from thesun. This was tested mainly on sunny days at noon, to testits behaviour when exposed to bright light conditions. Theseconditions were considered as most relevant, since these canlead to cloud misclassification. Despite the fact that the auto-iris of the lens made a difference by reducing the sun’s glow,the lens caused a reflection that lead to a misclassification ofthat part of the image (i.e. clear sky to a cloud). To solve thisproblem, a shadow band was considered and implemented. Thesecond implementation, consists on the camera aimed directlyat the zenith, with a dynamic shadow band to cover the sunduring the day time.

The shadow band is positioned according to a sun pathalgorithm [29], which takes as input: time, date, latitude andlongitude; and outputs the azimuth, solar declination, solaraltitude (in degrees), sunrise time and sunset time.

And finally, for the whole system to work, a computerEbox 4300 from ICop R©, was used to collect the data fromthe sensors, and to process the ANNs. The computer runsLinux R©, which was chosen due to its portability, stability andperformance.

Fig. 1. Developed prototype

IV. RESULTS AND DISCUSSION

The system validation was done in three different ways.The first validation consisted on comparing the results of theprototype system with those obtained on a CSI Laboratorycomputer, using the same data. This aims to validate theprototype implementation regarding the models. On the secondvalidation the results obtained were compared to those ob-tained by the reference system. The cloudiness was comparedwith a reference system (TSI), and its validation results arepresented in Section IV-A. Temperature and solar radiationresults were compared with those acquired by a weatherstation, incorporating a pyranometer (BF3 from ∆ − T R©)and a temperature sensor (RHT2 from ∆ − T R©). The solarradiation validation results are presented in Section IV-B, andthe validation results regarding temperature are presented inSection IV-C. This aims to validate the chosen sensors andthe acquisition hardware. The third validation is a comparisonbetween the predictions obtained by means of images acquiredby the prototype, and the predictions obtained using imagesacquired by the TSI.

The data used in the validation was acquired between the28th and the 30th of January of 2011. The first day wasused to initialize both NAR and NARX models, while the tworemaining days where used for comparison of predictions.

In the following sections the developed prototype will bereferred to as CloudSpotter.

A. Cloudiness Indexes

The first validation was performed by analysing the set ofimages acquired by the CloudSpotter in a computer of theCSI laboratory, and by measuring the maximum absolute errorbetween the the estimation of cloudiness computed in thelaboratory computer and the CloudSpotter. As for the result,a maximum absolute error of approximately 5.0e−14 wasobtained. This validation was also applied to the NAR modelused for prediction of cloudiness, using data acquired bythe CloudSpotter. The validation consisted on the comparisonof the maximum absolute errors obtained when comparingthe predictions computed by the CloudSpotter model imple-mentation, to those computed by the laboratory computer,considering the 1 step ahead prediction. The maximum errorobtained was 7.1e−15.

The second validation consisted on the comparison of thecloudiness estimated by means of CloudSpotter images to thatestimated by images acquired using the TSI (see Figure 2).The mean absolute error between the two systems is 12.20%with a variance of 15.39%. The difference observed betweenthe two systems, is due to the fact that the neural networkemployed was trained to provide a threshold using imagesfrom the TSI. When this threshold is applied to images witha different tonality, as those acquired by the CloudSpotter, itinduces an error in the cloudiness estimation (see Figure 3).

In Figure 3, it is possible to observe the difference inapplying the same threshold to images acquired by bothsystems. As already mentioned, the CloudSpotter images have

0

20

40

60

80

100

%

2011/01/28 2011/01/29 2011/01/30

Cloudiness Index Mean Absolute Error = 12.20%

Cloud Spotter

Total Sky Imager

Fig. 2. Estimated cloudiness indexes

Fig. 3. CloudpSpotter with threshold (top left), original image (top right);TSI with threshold (bottom left), original image (bottom right)

a different tonality than those of the TSI, which translated intoa cloud misclassification when the image is processed.

The results of the third validation are shown in Figure 4,which presents a comparison of the 1 step-ahead (5 minutes)predictions obtained by the cloudiness NAR predictive model,when using data from the TSI and from the CloudSpotter.The mean absolute error between the two systems is, for 1step ahead 12.062% with a variance of 14.156% and for 48steps ahead 9.937% with a variance of 10.65%, which arein accordance with the error obtained between the cloudinessestimations(see Figure 2).

B. Solar radiation

The first validation was performed by comparing the max-imum absolute errors for the NAR and NARX (cloudiness isthe exogenous variable) prediction models for solar radiation.The comparison was made between the the 1 step-aheadpredictions computed in the CloudSpotter and correspondingpredictions computed in a laboratory computer. This validationwas made with data acquired by the SPLite2 from Kipp &Zonnen R©, that was employed in the CloudSpotter. For theNAR prediction the maximum absolute error was 4.5e−13,and for the NARX prediction the maximum absolute error

20

0

20

40

60

80

100

120

RM

SE (

%)

2011/01/29 2011/01/30

NAR Cloudiness: prediction 1 step ahead Mean Absolute Error: (1 step) 12.062 %(48 steps): 9.937 %

@CloudSpotter: CloudSpotter data

@Lab. Computer: TSI data

Fig. 4. NAR cloudiness indexes prediction

was 8.9e−16.The second validation was performed by comparing the

measurement from the two sensors, the SPLite2 and the BF-3 from ∆ − T R©. The results of this validation are shown inFigure 5. The mean absolute error between the two sensorsis 16.525W/m2 with a variance of 49.182W/m2. Anotherimportant observation, is the fact that the dynamics of the twosensors are very alike, as expected.

0

100

200

300

400

500

600

700

800

900

W/m

2

2011/01/28 2011/01/29 2011/01/30

Solar Radiation

Mean Absolute Error = 16.525 W/m2

KippZonnen SPLite 2

∆−T BF3

Fig. 5. Measured solar radiation

The results of the third validation are presented in Figures6 and 7. The first shows the results of the NAR predictivemodel of the solar radiation NAR predictive model. Thiswas performed using the datum of solar radiation acquiredand computed by the CloudSpotter, in comparison to thedatum acquired by the BF-3 and computed in a laboratorycomputer. The mean absolute error between the two systemsis not significant and decreases when the number of steps isincreased. For 1 step-ahead it is 15.874W/m2 with a varianceof 35.515W/m2, and for 48 steps-ahead it is 8.816W/m2 witha variance of 6.111W/m2.

Figure 7 shows the predictions made by solar radiationNARX prediction model. As before, the validation was madeby the prediction computed using the datum of solar radiationacquired and processed by the CloudSpotter in comparison tothe prediction computed using the datum acquired by the BF-3 and processed in a laboratory computer. The mean absolute

100

0

100

200

300

400

500

600

700

800

RM

SE (W/m

2)

2011/01/29 2011/01/30

NAR Solar Radiation: prediction 1 step ahead

Mean Absolute Error: (1 step): 15.874 W/m2 (48 steps): 8.816 W/m2

@CloudSpotter: CloudSpotter data

@Lab. Computer: TSI data

Fig. 6. NAR solar radiation prediction

0

200

400

600

800

RM

SE (W/m

2)

2011/01/29 2011/01/30

NARX Solar Radiation: prediction 1 step ahead

Mean Absolute Error: (1 step) 22.40 W/m2 (48 steps) 15.28 W/m2

@CloudSpotter: CloudSpotter data

@Lab. Computer: TSI data

Fig. 7. NARX solar radiation prediction

error between the two systems increased when compared withthe NAR predictions. For 1 step-ahead and 48 steps-ahead,22.40W/m2 with a variance of 42.48W/m2 and 15.28W/m2

with a variance of 19.40W/m2, were achieved respectively.Most certainly, these errors are partially induced by the errorof the exogenous variable predictions, that are employed inthe NARX approach.

C. Atmospheric Temperature

The first validation was performed by comparing the maxi-mum absolute errors for the two NARX temperature predictivemodels, one where the radiation was provided by a NARmodel and the other, where the radiation was provided by aNARX model, cloudiness being the exogenous variable. Thecomparison was between the model predictions computed inthe CloudSpotter, and the corresponding predictions computedin the laboratory computer. This validation as performed usingthe the data acquired by the LM35DZ temperature sensorfrom National Semiconductor R©, that was employed in theCloudSpotter, considering the 1 step-ahead prediction. Forthe case where the radiation was provided by a NAR modelthe maximum absolute error was 3.2e−15. For the the othercase, where the cloudiness was the exogenous variable, themaximum absolute error was 1.8e−15.

The second validation was made by comparing the mea-surements of two sensors employed. The results are presented

in Figure 8.

0

5

10

15

20

25

oC

2011/01/28 2011/01/29 2011/01/30

Temperature Mean Absolute Error = 2.154 oC

NS LM35

∆−T RHT2

Fig. 8. Measured temperature

As it can be observed, there is an offset between themeasurements, which translates into a mean absolute errorof 2.154oC and a variance of 1.371oC. The observed offsetbetween the sensors, might be explained by inappropriatehousing of the LM35DZ sensor in the prototype, which wastoo small and probably providing insufficient air flow. Partof this offset, would probably be removed if the two sensorswould have been calibrated at the same time using the samereferences. Nevertheless, the observed patterns are very alikesuggesting that the LM35DZ can be used in future designsprovided that proper calibration is performed and adequatehousing is built.

6

8

10

12

14

16

18

20

22

RM

SE (oC

)

2011/01/29 2011/01/30

NARX Temperature (NAR model radiation): prediction 1 step ahead Mean Absolute Error: (1 step) 0.148 oC (48 steps) 1.780 oC

@CloudSpotter: CloudSpotter data

@Lab. Computer: TSI data

Fig. 9. NAR temperature prediction

The third validation is presented in Figures 9 and 10. Thefirst shows the prediction computed by the NARX modelof temperature where the radiation was provided by a NARmodel. The plot shows the model prediction computed usingthe datum of temperature acquired and processed by theCloudSpotter, in comparison to the prediction computed bymeans of the datum acquired by the RHT2 and processed in alaboratory computer. The mean absolute error between the twosystems, considering 1 step-ahead is 0.148oC with a varianceof 0.109oC and for 48 steps-ahead is 1.780oC with a varianceof 1.468oC.

Figure 10 shows the prediction computed by the temperature

6

8

10

12

14

16

18

20

22

RM

SE (oC

)

2011/01/29 2011/01/30

NARX Temperature (NARX model radiation): prediction 1 step ahead Mean Absolute Error: (1 step) 1.570 oC (48 steps) 1.164 oC

@CloudSpotter: CloudSpotter data

@Lab. Computer: TSI data

Fig. 10. NARX temperature prediction

model where solar radiation is provided by means of an exoge-nous variable, the cloudiness. As before, the prediction com-puted using the datum of temperature acquired and processedin the CloudSpotter, is compared to the prediction computedusing the datum acquired by the RHT2 and processed in alaboratory computer. The mean absolute error between the twosystems, considering 1 step-ahead is 1.570oC with a varianceof 1.536oC and for 48 steps-ahead is 1.164oC with a varianceof 0.935oC. These larger error values, as for the solar radiationcase, are probably due to the inclusion of the NAR cloudinesspredictive model as an exogenous input to the solar radiationmodel, which in turn was employed as an exogenous input inone of the temperature predictive models.

V. CONCLUSIONS

Apart from the estimation of cloudiness, which still needsimprovements, the other functionalities of the prototype de-veloped were achieved, as it may be concluded by analysingthe results obtained. The cloudiness estimation neural networkwas trained with images from the TSI, and since the camerasof the TSI and the prototype have different characteristics,the obtained difference of results were expected. This resultinfluences the prediction of the radiation (NARX), becausethe cloudiness indexes are the external variable for this pre-diction. The images may be improved by changing the cameraprotective dome which causes some unwanted reflections. Themeasurement of temperature and its prediction was very satis-factory, despite the fact that the measurement of atmospherictemperature was influenced by its housing. The measurementof solar radiation had excellent results, when comparing thetwo used sensors. The mean absolute error between the twois minimal and expected, since the sensors have differentoperation characteristics.

As a final remark, the developed system is portable, in away that it is easy to assemble and behaves nicely under allweather conditions.

VI. FUTURE WORK

The greater improvement should be done in the acquisitionof sky images. The dome must be replaced by one that does notreflect the inside part of the camera housing, and to overcome

the cloud mismatch two options must be considered. The firstrelies on the tuning of the camera, to change the colour setupfor this to be similar to the one of the TSI, in other words,to turn the images of the prototype more blue instead of thepresent setup where the images are more grey. The secondoption considers the training of the neural network with imagescaptured by the prototype. Both options should return reliableresults.

Regarding the measurement of atmospheric temperature, animproved housing and calibration of the sensor should beconsidered, to achieve more reliable measurements.

The presented system might need an improved coolingsystem, if it is going to be used in extremely hot environments.

A great improvement in the cloudiness data acquisition,could be the use of a magnetic field that would control theposition of a magnetic dot, which would move on the top of astructure or a lens to shadow the sun. This improvement wouldallow the use of all the available area of the image, withoutbeing affected by the sun.

Another improvement would be the air pollution dispersioncalculations using the Delta-T method (see [30]), which couldbe used on upcoming studies.

Acknowledgments

Financial support for this work was provided by projectPTDC/ENR/73345/2006, funded by the Portuguese NationalScience Foundation. The second author thanks the EuropeanCommission for the grant PERG-GA-2008-239451.

REFERENCES

[1] Kalogirou S. A, “Artificial neural networks in renewable systems appli-cations: a review,” Renewable and Sustainable Energy Reviews, vol. 5,pp. 373–401, 2001.

[2] Eduardo M. Crispim, Pedro M. Ferreira, and Antonio. E. Ruano,“Prediction of The Solar Radiation Evolution using ComputationalIntelligence Techniques and Cloudiness Indices,” International Journalof Innovative Computing, Information and Control, vol. 4, 2008.

[3] A.E. Ruano, E.M. Crispim, E.Z.E. Conceicao, and M.M.J.R. Lucio,“Prediction Of Building’s Temperature Using Neural Networks Models,”Energy and Buildings, 2005.

[4] Pedro M. Ferreira and A. E. Ruano, “Application of ComputationalIntelligence Methods to Greenhouse Environmental Control. Part II:Model Predictive Control,” Computers and Electronics in Agriculture,2007.

[5] Yankee Environmental Systems, Inc., Total Sky Imager Installation andUser Guide Model 440A Rev.C, 2007.

[6] C. Fonseca and P. Fleming, “Multiobjective optimization and multipleconstraint handling with evolutionary algorithms i: A unified formula-tion,” IEEE Transactions on Systems, Man and Cybernetics - Part A:Systems and Humans, vol. 28, no. 1, pp. 26–37, 1998.

[7] Arnout Feijt, Paul De Valk, and Sibbo Van Der Veen, “Cloud DetectionUsing Meteosat Imagery and Numerical Weather Prediction ModelData,” American Meteorological Society, 2000.

[8] Xiaoning Song, Yingshi Zhao, and Zhenhua Liu, “Cloud Detectionand Analysis of MODIS Image,” Geoscience and Remote SensingSymposium, 2004.

[9] Patrick Minnis, Qing Z. Trepte, Szedung Sun-Mack, Yan Chen, DavidR. Doelling, David F. Young, Douglas A. Spangenberg, Walter F. Miller,Bruce A. Wielicki, Ricky R. Brown, Sharon C. Gibson, and ErikaB. Geier, “Cloud Detection in Nonpolar Regions for CERES usingTRMM VIRS and Terra and Aqua MODIS Data,” IEEE Transactionson Geoscience and Remote Sensing, vol. 46, no. 11, pp. 3857–3884,2008.

[10] B. H. Kahn, M. T. Chahine, G. L. Stephens, G. G. Mace, R. T. Marchand,Z. Wang, C. D. Barnet, A. Eldering, R. E. Holz, R. E. Kuehn, and D.G. Vane, “Cloud Type Comparisons of AIRS, CloudSat, and CALIPSOCloud Height and Amount,” Atmospheric Chemistry and Physics, 2008.

[11] Buch, K.A., C.H. Sun, and L. R. Thorne, “Cloud Classification UsingWhole Sky Imager Data,” Proc. Fifth Atmospheric Radiation Measure-ment Science Team Meeting, San Diego, CA, U.S. Dept. of Energy, 3539,, 1995.

[12] G. Pfister, R. L. Mckenzie, J. B. Liley, A. Thomas, B. W. Forgan, andC. N. Long, “Cloud Coverage Based on All-Sky Imaging and Its Impacton Surface Solar Irradiance,” American Meteorological Society, 2003.

[13] A. Cazorla, F.J. Olmo, and L. Alados-Arboledas, “Using a Sky Imagerfor Aerosol Characterization,” Atmospheric Environment, 2007.

[14] H.F. Assuncao, J.F. Escobedo, and A.P. Oliveira, “A New Algorithmto Estimate Sky Condition Based on 5 Minutes-averaged Values ofClearness Index and Relative Optical Air Mass,” Theor. Appl. Climatol.,2007.

[15] J. Calbo and J. Sabburg, “Feature extraction from whole-sky ground-based images for cloud-type recognition,” Journal of Atmospheric andOceanic Technology, vol. 25, 2008.

[16] Francisco J. Olmo, Alberto Cazorla, Lucas Alados-Arboledas, Miguel A.Lopez-Alvarez, Javier Hernandez-Andres, and Javier Romero, “Retrievalof the Optical Depth Using an All-sky Camera,” Applied Optics, vol. 47,no. 34, pp. H182–H189, 2008.

[17] J.L. Bosch, G. Lopez, and F.J. Batlles, “Daily Solar Irradiation Esti-mation Over A Mountainous Area Using Artificial Neural Networks,”Renewable Energy, vol. 33, pp. 1622–1628, July 2008.

[18] J.C. Winslow, E.R. Hunt, and S.C. Piper, “A Globally ApplicableModel Of Daily Solar Irradiance Estimated From Air Temperature AndPrecipitation Data,” Ecol Model, 2001.

[19] Jerome G. Fortin, Francois Anctil, Leon-Etienne Parent, and MartinA. Bolinder, “Comparison Of Empirical Daily Surface Incoming SolarRadiation Models,” Agriculture And Forest Metereology, 2008.

[20] Tymvios F.S., Jacovide C.P., Michaelides S.C., and Scouteli C., “Com-parative Study Of Angstroms And Artifcial Neural Networks Method-ologies In Estimating Global Solar Radiation,” Solar Energy, vol. 78,pp. 752–762, 2005.

[21] Y. Xia, M. Winterhalter, and P. Fabian, “Interpolation Of Daily GlobalSolar Radiation With Thin Plate Smoothing Splines,” Theor. Appl.Climatol, vol. 66, pp. 109–115, 2000.

[22] Lefevre M., L. Diabate, and Wald L., “Using Reduced Data Sets ISCCP-B2 From The Meteosat Satellites To Assess Surface Solar Irradiance,”Solar Energy, vol. 81, pp. 240–253, 2007.

[23] Igor Antonio Cale Martins, “Neural models design for solar radiationand atmospheric temperature forecast,” Master’s thesis, Universidade DoAlgarve - Faculdade De Ciencias E Tecnologia, 2010.

[24] Kostas I. Chronopoulos, Ioannis X. Tsiros, Ioannis F. Dimopoulos,and Nikolaos Alvertos, “An application of artificial neural networkmodels to estimate air temperature data in areas with sparse network ofmeteorological stations,” Journal of Environmental Science and Health,vol. A, no. 43, pp. 1752–1757, 2008.

[25] P.M. Ferreira, A.E. Ruano, and C.M. Fonseca, “Evolutionary mul-tiobjective design of radial basis function networks for greenhouseenvironmental control,” Proceedings of the 16th IFAC World Congress,vol. 16, no. 1, 2005.

[26] Irineo L. Lopez-Cruz, Abraham Rojano-Aguilar, Waldo Ojeda-Bustamante, and Raquel Salazar-Moreno, “Arx models for predictinggreenhouse air temperature: A methodology,” Agrociencia, vol. 41,p. 181 192, 2007.

[27] P. M. Ferreira and A. E. Ruano, “Exploiting the separability of linear andnon-linear parameters in radial basis function neural networks,” IEEESymposium 2000: Adaptive Systems for Signal Processing, Communica-tions and Control, pp. 321–326, 2000.

[28] P. M. Ferreira, E. Faria, and A. E. Ruano, “Neural network models ingreenhouse air temperature prediction,” Neurocomputing, vol. 43, no. 1-4, pp. 51–75, 2002.

[29] Carruthers D, Uloth C, and Roy G. G., “An Evaluation of Formulaefor Solar Declination and the Equation of Time,” Research Report NoRR17, 1990.

[30] B. M. Bowen, J. Dewart, and A. Chen, “Stability class determination: Acomparison for one site.,” Proceedings Sixth Symposium on Turbulenceand Diffusion, pp. 211–214, 1983.