Solar Power Prediction with an Hour-Based Ensemble Machine Learning Method

Hittite Journal of Science and Engineering, 2020, 7 (1) 35–40 ISSN NUMBER: 2148–4171 DOI: 10.17350/HJSE19030000169

Solar Power Prediction with an Hour-based Ensemble Machine Learning Method

Seyda Ertekin Middle East Technical University, Department of Computer Engineering, Ankara, Turkey

Article History: ABSTRACT Received: 2019/11/23 Accepted: 2020/03/17 Online: 2020/03/26 n recent years, the share of solar power in total energy production has gained a rapid Iincrease. Therefore, prediction of solar power production has become increasingly im- Correspondence to: Şeyda Ertekin, portant for energy regulations. In this study we proposed an ensemble method which gives Department of Computer Engineering, competitive prediction performance for solar power production. This method firstly de- Middle East Technical University, 06680, Ankara, Turkey composes the nonlinear power production data into components with a multi-scale decom- E-mail: [email protected], position technique such as Empirical Mode Decomposition (EMD). These components Phone: +90 312 210 5509, Fax: +90 312 210 5544. are then enriched with the explanatory exogenous feature set. Finally, each component is separately modeled by nonlinear machine learning methods and their results are aggregated as final prediction. We use two different training approaches such as Hour-based and Day-based, for predicting the power production at each hour in a day. Experimental results show that our ensemble method with Hour-based approach outperform the examined machine learning methods. Keywords: Solar power, Time series forecasting, Machine learning, Ensemble methods, Empirical mode decomposition

INTRODUCTION

nergy sustainability is crucial for economic Solar power forecasting models are categorized growth and financial development. Therefore, according to model designs and inputs used by models. Esupplying energy through sustainable resources plays The Numerical Weather Forecast (NWP) model has an important role in national economic strategies. It been widely applied for solar radiation estimation [4,5]. is well known that renewable energy sources such as This approach uses mathematical models of weather solar power are sustainable in long-term timeline as conditions for prediction. This method is often used for compared to fossil resource [1]. With the developing short-term (such as 6-hour) forecasts and model perfor- technology, the cost of solar panels such as Photo- mance varies depending on weather data such as cloud Voltaic, which directly produce electricity from solar cover and temperature [6,7]. energy, is decreasing. Thus, in developing countries such as Turkey, solar energy production comes to the Time series models aim to predict future data from increasingly important level. past observed data using statistical and machine learning techniques. In literature, there are various time The total installed capacity of solar power plants in series methods [8, 9, 10, 11], some of them are directed Turkey increased by 57% over the last two years and has towards for forecasting solar power [12, 13, 14, 15, 16, reached to 5 GW [2]. With the planned capacity increa- 17, 18]. Solar time series data includes high volatility se, the installed capacity is expected to exceed 30 GW in and nonlinearity. Thus, nonlinear time series predic- the next 10 years [3]. In short, the share of total production methods are more preferable to achieve high pre- tion of solar energy production in Turkey is increasing diction accuracies in solar power forecasting [8, 10, 14]. rapidly. Therefore, forecasting solar power production Nonlinear methods, such as Artificial Neural Networks for making decision in power system scheduling and (ANNs), Gradient Boosting Machines (GBM), Support grid regulation plays an important role in Turkey’s elect- Vector Regressor (SVR) can arbitrarily fit complex non- ric industry. stationary functions with high accuracy. S. Ertekin / Hittite J Sci Eng, 2020, 7(1) 35–40 The developed method includes three main parts: Mul parts: main three includes method developed The The architecture of the method is shown in Fig. 1. Fig. in shown is method of the architecture The METHODS Since these components are highly correlated in themselves, themselves, in correlated highly are components these Since EMD are proposed [22, 23, 24]. [22, 23, proposed are EMD developed ensemble method takes various irradiation-rela various takes method ensemble developed amo the with proportional directly is production energy fea explanatory enriched but also data, historical use only different time periods. In order to achieve high accuracy accuracy high achieve order to In periods. time different other hand, in Day-based approach, each hour prediction in in prediction hour each approach, Day-based in hand, other methods in the prediction stage. We performed day-ahead We day-ahead stage. performed prediction the in methods more accurate forecasting results can be achieved. Solar Solar achieved. be can results forecasting more accurate with the corresponding hours in the historical data. On the the On data. historical the in hours corresponding the with rate predictions. In this ensemble method, the predictions predictions the method, ensemble this In predictions. rate results in solar power forecasting, several methods using using methods several forecasting, power solar in results unt of time that the solar power panels are exposed to the the to exposed are panels power solar the that of time unt radiance. In quarter. and month as such features chronological uses gether with prediction methods to increase prediction per prediction increase to methods prediction with gether prediction where we predict next day (24-hours) production (24-hours) production day next we predict where prediction play an important role in energy production. Therefore, the the Therefore, production. energy role in important an play properties and has different volatility and fluctuations in in fluctuations and volatility different has and properties non-stationary and nonlinear shows usually data power for every hour. We followed two approaches: Hour-based Hour-based hour. We approaches: two for every followed formance. Empirical Mode Decomposition (EMD) is one of one of is (EMD) Decomposition Mode Empirical formance. and Day-based. In Hour-based approach, each hour predic hour each approach, Hour-based In Day-based. and level at ground irradiance and sun of the angle the addition, and separately components EMD for each performed are any nonlinear and nonstationary data into components [21]. components into data nonstationary and nonlinear any a day is performed by training the model with the entire his entire the with model the bytraining performed is a day sun. Therefore, during the summer season at noon hours, at hours, noon season summer the during Therefore, sun. advantageous in solar power predictions, since it captures it captures since predictions, power solar in advantageous tiscale decomposition, feature enrichment and modeling. modeling. and enrichment feature decomposition, tiscale ted features into account. into features ted these variations, the seasonal developed ensemble method capture order to In appear. figures production highest the Solar result. prediction final as aggregated are results their more to accu approximations better yields which set ture not does which EMD using method ensemble an with ted time series problems [9, 19, 20]. EMD is able to decompose [9, problems series decompose to 19, able is time 20]. EMD different in technique decomposition used widely most the tion in a day is performed separately by training the model model the bytraining separately performed is aday in tion the high correlation between the production and solar ir solar and production the between correlation high the more is approach We Hour-based data. that showed torical The developed method is tested with ANN, SVR, GBM GBM SVR, ANN, with tested is method developed The Multi-scale decomposition techniques are also used to used also are techniques decomposition Multi-scale In this study, Turkey’s solar power production is predic is study, Turkey’s production power this solar In ------3 6 Multiscale Decomposition Packet Decomposition (WPD), Fourier Transform (FT) (FT) Transform (WPD), Fourier Decomposition Packet Solar power data shows nonlinear and non-stationary non-stationary and nonlinear shows data power Solar different time periods. Using multiscale decomposition decomposition multiscale Using periods. time different decompose the data accordingly. The decomposed compo decomposed The accordingly. data the decompose nents, are called Intrinsic Mode Function (IMF) and have have and (IMF) Function Mode Intrinsic called are nents, nonstationarity in solar power data. Thanks to this, forecas this, to Thanks data. power solar in nonstationarity given time series throughout the decomposition, thus it thus decomposition, the throughout series time given properties and has different volatility and fluctuations in in fluctuations and volatility different has and properties and etc. [26]. EMD is able to preserve time scale of the of the scale time [26]. preserve to able etc. is and EMD such as Empirical Mode Decomposition (EMD), Wavelet Decomposition Mode Empirical as such there are several multiscale decomposition techniques techniques decomposition multiscale several are there [25]. series literature, In time these in nonlinearities this analyze to method effective and efficient are techniques two important properties: properties: important two ting methods are able to produce more accurate results. more produce accurate to able are methods ting be described follows: as identify the intrinsic oscillatory modes in the data and then then and data the in modes oscillatory intrinsic the identify [21]. others the than technique amore preferable is Figure 1. 1. Each IMF has its own local characteristic time scale. scale. time characteristic local own its has IMF 1. Each These properties allow to resolve nonlinearity and and nonlinearity resolve to allow properties These The essence of the EMD technique is to empirically empirically to is technique EMD of the essence The 2. IMFs are relatively stationary series. stationary relatively are IMFs 2. Let y(t) be a solar power data and EMD technique can can technique EMD and data power y(t)Let asolar be The architecture of proposed ensemble method ensemble proposed of architecture The - - m = + y() t∑ imf () tkm r () t (1) k =1 where imf(t) represents the intrinsic mode functions, and r(t) is the residue. The detailed EMD procedure is as follows [27]:

1. All local extrema maxima and minima points of time series y(t) are identified.

Figure 2. 2. Upper and lower envelopes are constructed for iden- Monthly distribution of the solar power production of Turkey between Sep.2017-Sep.2019 tified extrema points by using cubic spline interpolation. Extraterrestrial, solar-oriented beam irradiance events 3. Mean value m(t) of the envelopes are computed. occurring in the Earth's atmosphere are examined in three different components [1]: Direct Normal Irradiance (DNI), 4. Take difference of mean and given series and obtain Diffuse Horizontal Irradiance (DHI) and Global Horizontal the first components: h(t)=y(t)-m(t) Irradiance (GHI).

5. Check properties of being IMF for h(t):

a. If h(t) meets two above conditions, it is con- sidered as an IMF and it is denoted as imf(t) in Equation 1. Then, let residual r(t)=y(t)-h(t) and replace y(t) with the residual r(t) return step 1. b. If h(t) does not satisfy the conditions, let y(t)=r(t) and return step 1.

6. Steps 1-5 are repeated until the obtained residual r(t) is a monotonic function. Figure 3. Quarterly distribution of the solar power production of Turkey between Sep.2017-Sep.2019 Feature Enrichment The production of solar energy highly depends on vario- DNI is the amount of solar radiation that is perpendi- us factors such as the presence of the sun, the angle of the cular to the Sun, received per unit area at a given location. sun and irradiance [1, 28]. Therefore, rather than using This measurement is particularly important to solar ther- only historical data, those features can also be included mal installations that track the position of the sun. DFI is into the feature set to increase forecasting accuracies. the amount of solar radiation at the Earth's surface from These features can be divided into two groups: Chrono- light scattered by the atmosphere and comes equally from logical and irradiance-related features. all directions. Unlike DNI, DFI does not take radiation co- ming directly from the sun into account. GHI is the total Seasonal variations of solar energy production can amount of irradiance from the sun on Earth. This value is be captured with chronological features. For this purpose, the sum of direct irradiance (DNI) and diffuse irradiance month and quarter knowledge is extracted from the time (DHI) which can be shown as follows: information and added into the feature set. GHI=+∗ DHI DNIcos q (2) The monthly distribution of solar energy production is where q is the solar zenith angle. Since GHI indicates given in Fig. 2. Production values reaches the highest levels the summation of DNI and DFI, solar power plants particu- in August, but it is very low in December due to insufficient larly follow this amount. sunny days. These irradiance values, which explain the angle of The quarterly distribution of solar energy production is the sun and quantity of the sun, are important explanatory given in Fig 3. It is evident that the production in the fourth features for solar energy production. Therefore, these are quarter of the year, which includes sunless months, is consi- also added to the attribute set of the developed ensemble S. ErtekinS. / Hittite J Sci Eng, 2020, 7(1) 35–40 derably less than that of the third quarter of the year, which method. includes sunny months.

37 S. Ertekin / Hittite J Sci Eng, 2020, 7(1) 35–40 While solar power production in the morning and evening evening and morning the in production power solar While The solar energy production data data production energy solar The RESULTS AND DISCUSSIONRESULTS AND Modelling Sep. 2017 to Sep. 2019. In this dataset, the solar producti solar the dataset, 2017Sep. 2019. Sep. to this In GBM, SVR methods that we used in this study. Finally, fore study. Finally, this in we used that methods SVR GBM, For Turkey solar power forecasting, hourly production production hourly For Turkey forecasting, power solar on values vary between 0 and 5039 Megawatt (MW), whi (MW), Megawatt 5039 0and between vary on values data is publicly available [29]. This hourly dataset con dataset [29]. hourly This available publicly is data This Fig.5. in shown is which used is years two of last data casts from each component are summed up. summed are component each from casts equation: n number of components ( of components n number rithms, ANN, GBM, and SVR, are adopted into the deve the into adopted are SVR, and GBM, ANN, rithms, power dataset is split into three sets such as train, validation validation train, as such sets three into split is dataset power adjust the hyperparameters of the methods. The last 90 days days 90 last The methods. of the hyperparameters the adjust to set validation the as used is set train 25% of the test. and and enriched feature set ( set feature enriched and approach. In this approach, each hour is trained and mode and trained is hour each approach, this In approach. Hour-based an with forecast, ahead day produce to aims sists of 17520 solar production values in Megawatt from from Megawatt in values of 17520 production solar sists loped ensemble method. In the modelling phase, the solar solar the phase, modelling the In method. ensemble loped MW. 979 is value production hourly average le the shows the hourly distribution of solar energy production. production. energy of solar distribution hourly the shows ber of the observed the data (imf(t-1),imf(t-2),…,imf(t-a)) data the observed of the ber anum with modelled is component each and by EMD, less variance show higher predictive performance. Fig.4 Fig.4 performance. predictive show higher variance less is hour each within production because is This itself. in led by machine learning models, as shown in the following following the in shown as models, learning by machine hours is low, it increases in noon hours. noon in low, is hours it increases with data in trained models Thus, itself. in correlated highly Figure 4. between Sep.2017-Sep.2019 m m m f f a t imf t imf t imf g t y ) ( ) ( ) ( ) ( =−−− − − += We propose to model the developed method, which which method, developed We the model to propose In the experiments, 3 different machine learning algo learning machine 3different experiments, the In where where 1 Hourly distribution of the solar power production of Turkey of production power solar the of distribution Hourly ( g is the machine learning models such as ANN, ANN, as such models learning machine the is .. , , ,..., 2 , 1 f_chronological,f_irradiance imf_1,imf_2,imf_3,…,imf_n y(t) is decomposed into into decomposed is hirr ch ) (3) ) ) ------38 We used those best hyperparameters in order to report our our report order to in hyperparameters best We those used Table in 1 algorithm of each space We search the provided (t) Mean Absolute Error (NMAE) and Normalized Mean Squ Mean Normalized and (NMAE) Error Absolute Mean of this time series, which has hourly 720 days of solar po of solar days 720 hourly has which series, time of this error results in Table in 2. results error hyperparameters. the tune to We search grid apply dataset. the present to used set test The set. training the are days 472 while set, validation the are days 630 remaining of the NMSE specifies the normalized average of the squared error. squared of the average normalized the specifies NMSE Table 1. wer production data, is used as a test set. The last 158 days days 158 last The set. atest as used is data, wer production roach, the same experimental setup is implemented for Day- implemented is setup experimental same the roach, rage of the absolute errors over a given prediction whereas prediction over agiven errors absolute of the rage performance of the methods is the last three months of the of the months three last the is methods of the performance ared Error (NMSE) which are formulated as follows: as formulated are which (NMSE) Error ared and we indicated the ones that give the best results in bold. bold. in results best the give that ones the we indicated and solar power capacity. NMAE measures the normalized ave normalized the measures NMAE capacity. power solar between methods, two error metrics are used: Normalized Normalized used: are metrics error two methods, between results performance comparing While approach. based Figure 5. between Sep.2017-Sep.2019 is the forecasted value at given time t. t. time at given value forecasted the is GBM ANN SVR In order to show the effectiveness of Hour-based app of Hour-based effectiveness show order to the In where where MEetN t e N t e NMSE NMAE The hyper-parameters of the applied machine learning methods learning machine applied the of hyper-parameters The Hourly distribution of the solar power production of Turkey of production power solar the of distribution Hourly e(t)=y(t)-y^(t) • • • • • • • • • =

=       1 n 1 n ∑ ∑ ) ( ) ( and and 2 learning_rate :[0.01,0.05,0.10,0.2 Parameter searchspaces n_est activ optimizer :[A / / y(t) imator :[300,900,3900,4900] ation :[ReLU, Sigmoid, Tanh ] architecture =(nx2n1) k max_depth :[2,3,4,5] ernel =[linear, poly, rbf] is the actual data value, value, data actual the is C =[0.1,0.5,1,10] epsilon =0.1 dam, Nadam, Sgd] N is the installed the is 0] (4) y^ y^ - - - - The experiment results are shown in Table 2. The CONCLUSION machine learning methods used in the proposed ensemble architecture are additionally executed individually and the- Solar power production forecasting plays an important ir results are also compared in terms of both error metrics. role in Turkey’s electricity industry for making decisi- Furthermore, all these experiments are performed for both ons in power system scheduling and grid regulation. In Day-based and Hour-based approaches, as seen in Table 2. this study, we developed an ensemble method which can work with different machine learning algorithms on Table 2. Experimental Results (All NMAE results are multiplied by 100.) EMD components of the time series data with explana- Approach Method NMAE NMSE tory exogenous features. We use this method to forecast GBM 2.80 14.93 solar power production of Turkey. This method firstly SVR 3.12 16.89 decomposes the solar power data into the components by ANN 2.72 13.00 Day-based the EMD decomposition technique. These components Ensemble – GBM 2.63 12.02 are then enriched with explanatory exogenous features. Ensemble – SVR 2.97 14.67 Each of these components are modeled separately by Ensemble – ANN 2.71 12.88 machine learning methods and their results are aggrega- GBM 2.55 11.41 ted to obtain final prediction result. We trained our en- SVR 3.01 15.91 semble method with two different approaches, Day-based ANN 2.64 11.83 and Hour-based. We performed the experiments on the Hour-based last two years of solar energy production data of Turkey. Ensemble – GBM 1.96 8.07 Our experiments show that when (i) the time series data Ensemble – SVR 2.13 10.98 is decomposed before feeding into the machine learning Ensemble – ANN 2.01 8.91 algorithms, (ii) explanatory exogenous features are used in addition to historical data, (iii) each hour in a day is When individual execution of machine learning met- predicted with the model trained with the corresponding hods are compared, it is seen that GBM and ANN give hours in the historical data. As a result, the developed better results than SVR. For example, in the Hour-based ensemble method has a high potential in prediction of approach, ANN and GBM give 2.55 and 2.64 NAME error solar power production in energy industry. results, respectively, while NAME result of SVR is slightly more than 3. As a result, ANN and GBM are able to yield better generalization in this nonlinear dataset. When the GBM and ANN results are compared in themselves, ANN has better results in individual use in execution, whereas GBM is able to have more accurate results in the proposed References ensemble approach. 1. Inman RH, Hugo TC, Carlos FM. Solar forecasting methods for renewable energy integration. Volume 39, Issue When the results of the proposed ensemble methods 6, December 2013, Pages 535-576, doi:10.1016/j. are examined, it is clear that it has improved the results pecs.2013.06.002. of the individually executed machine learning methods. 2. Enerji Atlasi, 2019, Gunes Enerji Santralleri, retrieved In both Hour-based and Day-based approaches, ensemb- September 12, 2019, from: https://www.enerjiatlasi.com/ gunes/ le methods are able to produce more accurate results for 3. GUYAD, 2019, retrieved October 22, 2019, from: http:// both error metrics. For example, in the Hour-based appro- www.guyad.org/TR,360/teias-10-yillik-kapasite- ach, GBM, SVR and ANN machine learning methods are planlamasini-yayinladi.html improved by the proposed ensemble method by 23%, 28%, 4. Heinemann D, Lorenz E, Girodo M. Forecasting of solar and 24%, respectively. This indicates that when machine le- radiation in: solar energy resource management for arning methods are fed with decomposed time series data electricity generation from local level to global scale. Nova with descriptive exogenous features, the movement in data Sciences Pulishers; 2006. Perez R, Beauharnois M, Lorenz E, Pelland S, Schlemmer J. can be captured more effectively. 5. Evaluation of Numerical Weather Prediction Solar Irradiance Forecasts in the US Proc. ASES Annual Conference. Raleigh, When Day-based and Hour-based approaches are com- NC, USA-1721, May; 2011. pared, it is clear that the Hour-based approach gives more 6. Marquez R, Carlos FM. Intra-hour dni forecast based on accurate results than the Day-based approach. As shown in cloud tracking image analysis.Solar Energy,91, 327-336, Fig. 4, solar power production values have different distri- May 2013, doi.org/10.1016/j.solener.2012.09.018. butions within each hour. Hour-based approach, which re- 7. Bosh J, Zheng Y, Kleissl J. Deriving cloud velocity from a area ErtekinS. / Hittite J Sci Eng, 2020, 7(1) 35–40 of solar radiation measurements. Solar Energy, 87, 196- duces the variance of these different distributions, outper- 203, Jan. 2013, doi.org/10.1016/j.solener.2012.10.020. forms Day-based results between 4% and 25%.

39 S. Ertekin / Hittite J Sci Eng, 2020, 7(1) 35–40 19. 18. 17. 16. 15. 14. 13. 12. 11. 10. 9. 8. 175, Neurocomputing, 50, 159–175, doi:10.1016/ 50, 2002, 175, Neurocomputing, 151-163, 2019, doi.org/10.1016/j.neucom.2019.05.099. Zhang P. Time Series Forecasting Using a Hybrid ARIMA aHybrid Using Forecasting P. Series Zhang Time Artificial Neural Networks. In 2015 North American Power American 2015 In North Networks. Neural Artificial Journal forecasting. series time univariate for model ANN 2011, 11, 2664–2675, doi:10.1016/j. Comput., Soft Appl. Xu X, Qi Y, Hua Z. Forecasting demand of commodities of demand Y, Forecasting Z. Qi Xu X, Hua 343-355, 2007. 2010, doi.org/10.1016/j.solener.2010.07.002. 2010, 2017. https://doi.org/10.1155/2017/2437387. gazimmfd.508394. Gazi of Architecture and Engineering of Faculty the of method hybrid ARIMA-ANN a new using data series time of after natural Disasters. Expert Systems with Applications, with Systems Expert Disasters. natural after and NeuralNetwork Model. Neurocomputing 50, 159- 50, Neurocomputing Model. NeuralNetwork and asoc.2010.10.015. autoregressive and dynamical system (cards) model. Solar model. (cards) system dynamical and autoregressive and empirical mode decomposition, Neurocomputing 361, solar radiation on an hourly time scale using a coupled using scale time hourly an on radiation solar solener.2012.10.012. Glasbey CA, Allcroft DJ. A spatiotemporal auto-regressive Aspatiotemporal DJ. Allcroft CA, Glasbey One Day Ahead of Global Solar Irradiance. SSRN Scholarly SSRN Irradiance. Solar Global of Ahead Day One Time on Based Irradiance Solar Global of Prediction M. Cony S0925-2312(01)00702-0. Series Analysis: Application to Solar Thermal Power Plants Power Thermal to Solar Application Analysis: Series Symposium (NAPS), 1–5, 2015. moving average model for solar radiation. Appl Stat, 57, Stat, Appl radiation. solar for model average moving Intell. Soft Comput., 1, 52–76, 2008. Comput., Soft Intell. Energy Production. Solar Energy, 84, 1772-1781, 84, Energy, Solar Oct. Production. Energy Prediction. Research Article. Journal of Renewable Energy, Renewable of Journal Article. Research Prediction. Power Solar ANN and Selection J.Feature O, Kubby Daniel 29, 2018.May Network, Research Science NY: Social Rochester, Paper. Forecasting for Network Neural Artificial K. Himdi E, Hamid Using Forecasting Power Solar C. Chowdhury A, Mohamed and R, Marchante, A, LF, Navarro J, Polo Zarzalejo L, Martin J.Artif. Int. AReview. Data: Radiation Solar of Forecasting and Modelling for Technique Intelligence Artificial A. Mellit, J. Forecasting Boland M, Agrawal M, Korolkiewicz J, Huang ARIMA- hybrid Afeature-based S. Ertekin UC, Buyuksahin Networksand ARIMA Models for Time Series Forecasting. Neural Artificial of Hybridization ANovel M. Bijari M, Khashei accuracy forecasting Improving S. Ertekin UC, Buyuksahin Energy; 87, 136-149,Energy; 2013, doi.org/10.1016/j. Jan. University 35:1 (2020) 467-478 35:1 doi.org/10.17341/ (2020) University 4 0 24. 23. 22. 21. 20. 29. 28. 27. 26. 25. January 2017, doi.org/10.1016/j. 288-298, January Pages Yu W, Mu-Chen C. Forecasting the short-term metro short-term the Forecasting C. Yu W, Mu-Chen 454 (1998) 903-995. doi:10.1098/rspa.1998.0193. Angela Z, Faltermeier R, Keck I, Tomé A, Puntonet C, Puntonet Tomé A, I, Keck R, Faltermeier Z, Angela Volume 37, Issue 6, June 2010, Pages 4313-4317, 2010, Pages June 37, 6, Volume doi. Issue energy.2016.11.061. org/10.1016/j.eswa.2009.11.069. org/10.1016/j.trc.2011.06.009. 148-162, 2012, https://doi. Pages 1, April 21, Issue of global solar radiation based on multiscale decomposition doi:10.1109/ICCPCT.2017.8074179. 2017, April 20-21 (ICCPCT), Technologies Computing and Power Circuit, on 2018, from: https://seffaflik.epias.com.tr/transparency/ time nancial to non-stationary transform Huang Hilbert of applied to solar irradiance, global temperature,sun spot temperature,sun global irradiance, to solar applied and the Hilbert spectrum for nonlinear and non-stationary study of high frequancy global solar raiation sequences raiation solar global frequancy high of study series analysis. Applied Stochastic Models in Business and Business in Models Stochastic Applied analysis. series Calif R, Schimtt FG, Huang Y, Soubdhan T. Intermittency Y, Soubdhan Huang FG, Schimtt R, Calif under a tropical climate. Solar Energy, 98, 349-365, 2013. 349-365, 98, Energy, Solar climate. atropical under uretim/gerceklesen-uretim/gercek-zamanli-uretim.xhtml Science & Engineering, 6(1), (2010), 75-81. &Engineering, Science of Journal Hittite Model. ATransient Community: Small a for Demand Electricity to Provide Storage Heat Sensible neural networks, Transportation Research Part C, Volume C, Part Research Transportation networks, neural passenger flow with empirical mode decomposition and methods: A hybrid approach. Energy, Volume 119, Volume Energy, 15 approach. A hybrid methods: Terr Solar JAtmos data. concentration co2 and number Industry, 19(3), (2003), 245-268. 19(3), (2003), Industry, time series analysis. Proceedings of the Royal Society of Society Royal the of Proceedings analysis. series time Networks, 1–8, 2010. Networks, Neural on Conference Joint International the of Proceedings Introduction. -an Decomposition Mode Empirical E. Lang Monjoly S, Andre M, Calif R, Soubdhan T. Hourly forecasting forecasting T. Hourly Soubdhan R, Calif M, Andre S, Monjoly Forecasting Power Solar N. Nayak MK, Behera I, Majumder 2011;Phys 73:1771. decomposition mode Empirical WE. Eichinger BL, Barnhart Sciences Engineering and Physical Mathematical, A: London decomposition mode empirical The HH. Liu Tung CC, NC, Q, Yen Zheng HH, Shih MC, Wu SR, Long Z, Shen NE, Huang Epias, 2018, Gerçek Zamanlı Üretim, retrieved October 26, October retrieved Üretim, Zamanlı 2018, Gerçek Epias, with Plant Cogeneration aSolar of Simulation Performance Annual S. Y, Su J, Li Y, C, Riffat G, Wang Pei Kutlu, SP. Shen Applications SR, W, Long Qu ML, Wu NE, Huang Using a Hybrid EMD-ELM Method. International Conference Conference International Method. EMD-ELM a Hybrid Using