‘The Sixth Intemational Conference on Innovative Computing Technology (INTECH 2016) Predicting Daily Mean Solar Power Using Machine Learning Regression Techniques Faizan Jawaid Karachi Institute of Economies and Technology, Pakistan [email protected] Absrace—Daily mean sola irradiance i the most cial persueie io sting the intl of saat power gronation {nis The average solar radiation ona speieoction can hep predict the amount of electricity that wil he generated through Solar panel and an accurate forecast can hep in caulting the Size of the system, return on investment (ROM) and system hd Imeasurenens. To predict he mean solar radiation © various ‘egrescionalgoriths have been used in conjunction wih various tarametrstclated to solar irradance In this paper we present S"comperative ansgss of frccasting thrtgh arial nearal fetworks (ANN) aganat the standard regression algorithms. Furthermore, we show that incorporation of szinwath and renith parameters in the mode sigan proves the performance, 1. Intropuction With the dwindling fossil fuel resources, research in re newable energy has gained significant impetus. The leading. source for renewable energy is solar power generation mainly through photovoltaic cells. Advantages of using solar energy include its immunity to imitative circumstances like the oil prices, a clean source of energy, and reduction of imports and dependability on external resources. Though photovol cells are considered as a major souree for future energy generation, their return on investment and upfront cost is, hindering their deployments. One of the reasons for this is lack. of predictable supply because of changing weather conditions. Since photovoltaic cells generate electricity by converting solar energy to electric current, the amount of solar energy beit provided in 8 day is very important to size the photovoltaic system. Therefore, the amount of electricity produced depends ‘upon solar irradiance in a particular day, which itself depends. fon various parameters such as location, time, and weather patterns, Solar irradiance is defined as the power per unit area received from the Sun in the form of electromagnetic radiation in the wavelength range of the solar cell being used. Untavourable weather reduces the output of the solar plant to a large extent. Therefore, in order to fulfil the energy requirements, a power supply company needs (0 supplement the remaining amount by purchasing power trom different power generation companies running on costlier fossil fuels. ‘The power rates charged by these companies not only depend ‘upon the amount of the power required but also on the timeliness of the order. A timely order placement with these ‘companies not only helps to meet the promised power supply ‘goals, but also helps reduce the cost. Hence prior knowledge ff the power produced plays a crucial role in maintaining ‘978-1-5090-2000-3/16'831.00 ©2016 IEEE Khurum NazirJunejo ‘Singapore University of Technology and Design, Singapore Karachi Institute of Economies and Technology, Pakistan khurram@ pafkict.edu.pk quality of service and reducing cost. With solar power, its possible to predict the prodtion knowing current and the Past information about the weather andthe iradiance. Various researchers have proposed forecasting mechanisms with good results however room for improvement stil exists, There are to orthogonal avenues of improvement in this domain, on algorithm design for forecasting and second is identification and quantification of the effect of parameters on forecast In this paper we attempt to improve the state ofthe arin both the dimensions. Our fist contribution is evaluation of forecast of soar irradiance verity of machine Teaming repression algorithms. Artifical neural nwork (ANN) in particular achieve avery high degree of accuracy. Our second contribution isthe use of solar angles (Azimuth and Zenit) in conjunction with the weather data Ve show that these two angles help improve the prediction performance. Priclaly we try 10 pret the mem dally Solar energy that can be used by a solar plant at a given location based on the weather forecast. To accomplish this we consider various weather parameters which include, but not limited to, mean temperature, wind speod, visibility etc. We Dt together a 10 years of observed weather and irdiance data from Los Angeles, California Zip 722950 for ths study We evaluate the learned models cross-validation, and split validation In Section TI-we review the previous research work con- ducted on solr energy prediction, We discuss our methodol- ogy in Section IT. Section TV pertain t0 the results based on two different validations techniques, while Section -V concludes our work is more efcier UL, RELATED Work ‘The current eras’ accelerating advances in all fields rely ‘on the constant supply of the electric power. The fossil fuels are key sourve of the reliable and consistent energy provision, but they have high cost associated with them, emit dangerous gases, and are subject to uncertainty of the international oil prices. On the other hand, solar power is cheap, clean source of energy that ean be produced by every country, but its depen- dence on weather conditions make it less reliable, To cope with this unreliability the prediction of the solar power contributes towards a consistent supply. The prediction of solar power is multidisciplinary research that needs contribution from 355‘meteorology, solar cell engineering, electrical engineering, and ‘machine learning computation. Fig. 1. Solr aint a eth angen [1] used the radial basis funetion (RB) to model the solar radiation based on sunshine duration and air teraperatre data. ‘They use Multi-Layer Perceptton (MLP) to model the hourly forecast of the solar radiation using present values of sir temperature and mean daily solar irradiance. [2] compare different models including Fuzzy Logic, Neural Networks and Autoregressive 1o predict the half daily values of soar radiation. In [3], 3-day forecast of the solar irradiance has been modelled using the forecast dala provided by European Centre for Medium-Range Weather Forecasts (ECMWF). In {4}, different weather parameters such as solar hours, latitude, longitude, elevation, maximum and minimum air temperature, humidity, and rainfall has been used to predict the solar power in Indonesia. It also used ANN to generate a model that could prot the solar power ofa particular region, A recent ation is the use of computer algorithm based on fuzzy logic contol (FLC) to estimate the wind and solar energies in a hybrid renewable energy system from natural factors [5]. The solar power was estimated using the temperature and the lighting 4 input parameters. Ak detail introduction to the current research on forecasting solar iradiance is presented in [6]. Its a in-depth review to facilitate selection of the appropriate Forecast method accond- ing to needs. They also comment on the statistical approaches and techniques based on images from satelite imagery. They also discuss numerical weather prediction (NWP) and hybrid models. Online forecasting of power production from PV systems is discussed in [7], to predict hourly values of solar power for horizons of up t0 36 h. A tworstage method of using ‘udapive linear time series movels for clear sky model and autoregressive (AR) models NWPs is elaborated. [8] compare ‘multiple regression techniques for generating prediction mod- els for solar power, including linear least squares, and support vector machines using multiple Kernel functions. IIL, MernopoLocy A. Dataset Reliable data availability and the choice of right attributes from this data are crucial to the accurate prediction in general and for solar power in specific. In this work we relied on the dataset provided by National Solar Radiation Database (NSRD)! and National Climate Data Center (NCDC)*. The former is provided by the National Renewable Energy Lab- oratory (NREL) which collects data for various resources including solar related energy for all cities of USA. The later is now known to be managed by National Oceanic and Atmospheric Administration(NOAA). ‘We chose 10 year weather information dataset over Los An- geles, California ZIP 722950 from above stated two different sources, The dataset available is based on hourly values of weather parameters. Average of the 24-hour data has been taken to convert it to the mean values per day. Different ‘weather parameters have been collected from 1991 to 2000, to explore the relationship between the mean solar iradiance and weather data for the accurate prediction of mean solar inradiance. The collected data comprises of the average daily values of air temperature, min, air temperature, max. air temperature, wind speed, dew-point, visibility, solar azimuth angle, and solar zenith angle, The solar azimuth angle gives the direction of the sun measured in degrees. Whereas, solar zenith tells how high the sun is. I is also measured in degrees. ‘The figure | shows the solar azimuth and zenith angles. To increase the prediction accuracy, month and week of the year is also made part of the dataset. Figure 2 shows the pattern of, different weather parameters measured on daily basis for Year 2000. These are the mean values of hourly data measured on daily basis. B. Regression Modelling for Prediction We have tested four commonly used machine learning techniques in this work with the selected dataset to evaluate individual performances with the choice of weather attributes. ‘The first one is K-Nearest Neighbor(K-NN) approach [9]. K- NN is a lazy Jeaming approach i.e. it does not explicitly build a model over the taining data, instead, given a unseen record from the test sample, it finds its closest K matches in the training data. Since our prediction variable is continuous valued, mean of these K closest matches is predicted as the output of the unseen test sample. We experimented with different values of K but report results for only when K=3 and Ke5. RMS error increases for values of K>3, Linear Regression (LR) is the most basic and widely used technique for regression [10], It models the relationship be- ‘oveen the input and output variables using linear predictor functions whose unknown model parameters are estimated from the data using a least squares approach, The parameter values can be estimated either by solving a set of linear equa- tions or using an iterative method such as gradient descent Support Vector Machines (SVM) are non-probabilistic bi- nary linear classifiers. They project the data in a higher dimensional feature space by means of a kernel function that separates the points. A hyperplane is leamed in this feature space to discriminate between the points of the 10) classes, The hyperplane is such that it maximizes the margin "nptrodcare govksolarod_datasedb/1991-2010houy st by. state nt Pt nde nova govpabidalesod! 356co g 100 2 E10 3 80 = E 2 60 = 60 z 2 4 59 100 200 300 * 100 200 300 2 80 8 2 0 [ranybeye 3 10 & 2 £40 zs s = = a 100 200 300 i 100 200 300 6 100 z : 50 8 0 100 200 300 ° 100 200 300 90 so 7 60 100 200 300 ° 100 200 300 Fie. 2 Mean dell values of weather parameters ofthe Yeur 2000, X-Lahol=Duy of Yur 2000 between the closest points of the two classes, thus achieving. a better generalization over the unseen data, This has led 'SVM to exhibit high performance on most of the real world classification problems [11]. {12}, 13] propose a method © use SVM for regression. Afler projecting the data into a higher dimensional feature space a linear regression is performed. Artificial Neural Network (ANN) are inspired by the fune- tionality of neurons in animals. A neuron is a processing unit hhaving output and inputs and an activation function. There fare many variations of the ANN, for this work we have used. the simple feed forward neural network with back-propagation [14], The weather parameters are given as inputs to the ANN and it gives the predicted output asthe solae power. The neuron functions are learned iteratively using training dataset, For the ANN model, learning cycles are 1000 while learning rate is, (0.2 and momentum is 0.1. It is observed that as the number of leaming cycles were increased, the RMS error decreased. Increasing the learning rate and momentum effected the RMS error us well The discussed techniques are designed and tested with RapidMiner [15]. The dataset is split in testing and training dataset, the techniques used to split along with the results are discussed in ILC. ©. Validation To better understand and evaluate the (rained models, (60 different validation techniques are used namely, K-fold Cross- Validation (CV) and Split Validations (SV). At first, the CV technique is used which orderly divides the data into K seg- ‘ments. The training sequence is performed on K-1 segments and testing is done on the remaining segment. This whole cycle is repeated K times. Each validation cycle yields a root means squared (RMS) error value. The final result is an average of, these K RMS values, RMS error [3] is defined as follows 3871S (mode RMS observed)? (I) ‘where n=10 years * 365 = 3650 instances. Due to the chrono- logical nature of the data, we also gather results using split validation (SV) approach. Therefore we divide the first nine years of data (years from 1991 to 1999) for training while {esting only on the last one year of data (year 2000) (Figure 3). This scenario is more realistic as well as more challenging as the distribution of the data in the training data might be different than that of the test data because of the concept shift that oceurs over time. 1991 199 Yen) 199162000 Testing Data 2100 ea) Fig. 3. Split validation ofthe dataset IV, RESULTS AND Discussion PDH GE ppp Mee ie as Fig 4. RMS Eror of cifernt moss As disused previously, 10 validation methods, K-fold cross-validation and split-vaidaton were used, and we present their els separately, In order to ascertain how much taining dita is sufficient for achieving a good prediction we vary the size of the data used, We used three sizes of the da, i. two, tive, and nine years. Figre 4 compares the performance othe four modes sing the K-fold cross validation approach on these three dataset sizes. It can be observed that ANN, utperorms the rest of the three models bya sufficient margin Secondly, increasing the size ofthe data decreases the RMS error. It might be noted that RMS error reported here for daily mean solar power Wf, is an overestimate of the average difference ofthe predicted value from the actual value, This is because RMS error metric is sensitive to outliers because it squares the difference of the predicted and the observed value. ‘Therefore, the actual average difference between the observed and predicted value is lesser. RMS error is used because of its nice convergence properties. The result for LR. and SVM 9-year are missing because of the scalability issues. RMSERROR Fig 5. RMS Emor of dierent modes using SV Figure 5 shows the results for the split validation approach. ‘The figure shows the result on the one year data used in the fest set, The results for SV tell a similar story. There is n0, significant difference in the errors from CV approach, This indicates that there were no significant change in weather pattern for the Los Angeles area over the ten years. A. RMS Error Comparison q- ene q4 en i teal i in Fig. 6, Predicted Vs Observed (Actual) Daily Mean Solar Power We plot out the predicted values against the actually ob- served values using the SV approach in Figure 7. KN ‘overestimates the daily mean solar power deviation forthe first half of the year 2000, and underestimates it for the second half of the year. Whereas the prediction of SVM remains ffuctuating in the frst half but predicts better in the second half LR, overall, performs better than K-NN, and SVM. It also exhibits a lesser deviation than the two. ANN on the other hand fits the observed curve almost perfectly Finally, we try to determine the most important tribute in prediction of the daily mean solar power. For this purpose 358Mean Daly Soran (Win2) 8 “BRSSSRSRSSEARRARRARARG SESRSSRSRRARERERRE converse pavers Mean Day Soar leone (Whi) Fig 7 ‘we excluded the attributes one by one from the dataset and releamed the whole model. The maximum drop in the performance of the model were because of the exclusion of solar azimuth and zenith. Figure 6 shows the difference in the performance of ANN using the SV approach with and without these two attributes, V. ConcLusion Over the past years solar power has gained a significant importance as a renewable, clean, and alternative source of energy. However, the output of a solar power plant heavily. depends on the weather patterns and time of the year, An accurate prediction of daily solar power by modelling these varying patterns can help solar power companies complement their production shortfalls in a timely manner, thus saving high cost of purchasing power from the market at the last moment. We therefore, demonstrate the ability of machine learning regression techniques, especially artificial neural networks t0 accurately predict the daily mean solar power up 0 a very high degree of accuracy, In addition to using weather and time of year specific attributes, we also propose and establish the effectiveness of using solar angles (azimuth and zenith) (o increase the accuracy of such predictions, We validated ‘our models on 10 years of solar irradiance and weather data, collected for the Los Angeles city, USA. This study is a first ‘Mean Daily Star leraiace (Win 2) Priced vs Observed Daly Meun Solar Power using SV. step in suggesting sites that are most suitable for solar power generation. REFERENCES, A. Melt M. Menghancm, and M. Bendekhis, “Arificial neural net ‘ork mle! for prediction sla ailation dt: appleton for sizing sandstone photovolie power system: in Pover Engineering Society General Meeting. 200. IEE, Tne 2008p. 40-4 Vol. 1 [o'Man, TF. Zaralj, J. Polo, Av Navarro, R. Macha, sid M: Cony, “Prediction of ploal sola igadance Rise on tine feviee analysis Application «0 solar thermal power plants energy proton planning. Solar Energy, vol 84, m0. 10, pp. 1772 — 178, 010. Online. Avnishe: hip scenceiect comical ps0ss092x10002579, [51 E-Lovene J Hock, D. Heinemann, and H. Beyer, “radiance fre Ing forthe power prefiston of ri-comctt photovoltaic sytem” Selected Ties in Applied Earth Obnervaton and Remote Sensing, TREE dural of, ol. 2,9 1p 2-10, Mach 208, [4 A. Prstiwn and R Dalim, "New approach om renewable eneray tele pres preicton it bomsia boc oe einen neta tectniqu: Southern roion of sues tan. sudy case” ia QR (Quality ix Research), 2013 International Conference on, Sune 2013, pp 166-109. S' Figur A. Yabyaouy, HL Tih and J. Sabor, “Implementiag fzzy logic based sgoritn 10 peodct solr ant wind cares Ina hybrid renewable energy systems" dn. Pacsy Sy App ol 4. a0, 3 pp. 10-24 Tul. 2015, [Online Avalble: tp is dori 4O18IBFSA 2018070102 [6] ML Diagn, M. David. P Lawet, J. Boland, and N. Schmutz, “Review of solar iadlance forecasting methods and proposition fo smallscale instar gods” Renewable and Sttainable Energy BI is 359eviews, sol. 27, pp. 65 ~ 76, 2013. (Online) Available: tp: ‘ew sciencediec.comisciencefariclpi/S 1364032113003 Bacher. H. Madsen, and H. A. Niclen "Online sbetterm solr power forciting” Solar Emerge, vol 88, n0. 10, pp. 1772 ~ 178%, 5009. [Onn Ausable: pw scence om/cienclaicl PRISIOSMODEXOMOOTSES Is] No Sharma, Pe Sharma, Do. Irwin, and PJ. Shenoy, “Predicting slit generation from weather forecasts using. machine leaning” in SmansGridComom. IEEE, 2011, pp. S2H-533, [Online], Avalble hp :flpun-tiedldhconlsmangidconay Ssmargrcomm201 [hnl#SharmeSIS11 NUS" Altman, "An Inaction #0. Kem and. Neaes-Neightor Nonpuramote Regression” The American Statsicin, vol 46, 0.3, poe t7S-188, 1992. [Online| Avalble: hp on/1023077 easton (10) D.C. Montgomery. B.A. Peck, and GG. Vining. Iduction to Hier gresion ants. on Wiley Sons, 2018 [1] LAhmad, Mt Hussain, A. Alshamdl, and A. Alelawi, “Eahancing ‘im performance in Insion detsion sing opal fete ht Selection tusod on genetic principal components” Neural Computing land Applications, 24,0, 7-8 gp. 671-1689, 2014 ‘A. Smla and V. Vapaik. "Suppor vector regression maces” Ad anes in new ifrmation processing systems, vo. 8. pp. 185-16]. [13] D. Basa, $. Pa, apd D.C. Parana, “Support vector regression” [Neural Ijormation Processing Leer and Reviews, Vl 11. 8 10 20204 2007 (4) DE Ramet, G. E, Hit, and RJ, Wiliams, “Neurocomputing: Foundations of reeuch’” "A. Anderson and E. Roseoeld. Ed Cambridge, MA, USA: MEF Press 19, ch, Leaoing Representations by. Backpropizating Eros, pp, 636-609. [Online| Avalable plage 65669 105851 1S] MeHotmann and: Klnkenber, Repidiner: Dra Mining Use Cases ‘and Business Aralies Applications. Chapman HIVCRC, 2013 mI pl‘Authorized icons use ied to: University of East London, Downloaded on February 202023 at 15:00:40 UTC fom IEE Xplore, Restctions app