• No results found

Forecasting Solar Power Generated by Grid Connected PV Systems Using Ensembles of Neural Networks


Academic year: 2021

Share "Forecasting Solar Power Generated by Grid Connected PV Systems Using Ensembles of Neural Networks"


Loading.... (view fulltext now)

Full text


Forecasting Solar Power Generated by Grid

Connected PV Systems Using Ensembles of Neural


Mashud Rana

Australian Energy Research Institute University of New South Wales

NSW, Australia md.rana@unsw.edu.au

Irena Koprinska

School of Information Technologies The University of Sydney

NSW, Australia irena.koprinska@sydney.edu.au Vassilios G Agelidis

Australian Energy Research Institute University of New South Wales

NSW, Australia vassilios.agelidis@unsw.edu.au Abstract—Forecasting solar power generated from

photovoltaic systems at different time intervals is necessary for ensuring reliable and economic operation of the electricity grid. In this paper, we study the application of neural networks for predicting the next day photovoltaic power outputs in 30 minutes intervals from the previous values, without using any exogenous data. We propose three different approaches based on ensembles of neural networks - two non-iterative and one iterative. We evaluate the performance of these approaches using four Australian solar datasets for one year. This includes assessing predictive accuracy, evaluating the benefit of using an ensemble, and comparing performance with two persistence models used as baselines and a prediction model based on support vector regression. The results show that among the three proposed approaches, the iterative approach was the most accurate and it also outperformed all other methods used for comparison.

Keywords—solar power forecasting; renewable energy; neural networks; ensemble of classifiers; clustering; iterative prediction


Solar power generated from photovoltaic (PV) systems is one of the fastest growing and most promising sources of renewable energy [1]. PV systems produce electricity directly by utilizing the absorbed solar irradiance. Many countries have built large-scale PV plants and connected them to the electricity grid [2, 3]. In Australia, the solar irradiance per square meter is the highest compared to any other continent in the world [4], which has led to a rapid increase of grid-connected and standalone PV systems.

However, the power produced by a PV system is highly variable and intermittent, which makes the integration of the produced solar energy into the electricity transmission and distribution networks a challenging task. The power produced by a PV system depends on the solar irradiance, cloud cover and other weather and environmental conditions, and also on

the parameters of the PV systems. To ensure reliable and economic operation of the power grid, it is essential to forecast the power output from PV systems accurately at short time intervals (from minutes to a day ahead). This is needed not only for maintaining the stability of the grid, but also for supporting the transactions of electricity suppliers, providers and traders at competitive electricity markets.

Solar power forecasting has received an increased attention recently due to new legislations encouraging the deployment of solar power plants. The most prominent approaches are based on statistical methods such as auto regression, moving average and their combinations [5, 6], computational intelligence methods such as Neural Networks (NNs) [2, 3, 5, 7, 8], nearest neighbor [5] and Support Vector Regression (SVR) [9-11], and also fuzzy inference

Most of the existing approaches predict the PV power output indirectly by firstly predicting the solar irradiance (or using predictions provided by meteorological centers) and then converting the predicted solar irradiance into PV power output by considering the characteristics of the PV plant such as area and efficiency. However, obtaining local weather forecasts of solar irradiance, cloud coverage and other meteorological variables for the site of the PV plant is not always possible. In this paper, we consider the task of directly predicting the PV power output from previous PV output data only, without using any exogenous environmental and meteorological variables.

While the majority of previous work focuses on one step ahead prediction, i.e. at time t, the task is to predict the power output at time t+1, we consider multiple step ahead prediction – our goal is to predict all half-hourly PV power outputs for the next day. More formally, given a time sequence of PV power outputs up to the day d, , , … , , , where Pi is a vector of half-hourly PV power outputs for day i, i.e.


, … , , our goal is to forecast , … , ,

all half-hourly power outputs for day d+1.

As prediction models we developed and applied ensembles of NNs. Although feed-forward NNs have been successful for forecasting time series in various applications, their performance is sensitive to many parameters, including the network architecture and random initialization of weights. Combining several NNs in an ensemble can reduce this sensitivity. Ensembles of NNs have been shown to be successful for multiple steps ahead prediction of electricity load data [12, 13].

The contributions of this paper are:

1. We propose three different NN ensemble based approaches for forecasting the half-hourly PV power outputs for the next day, using only previous power data. The first two approaches predict all power outputs for the next day simultaneously and the third one does this iteratively by using the previously predicted values.

2. We comprehensively evaluate the performance of the proposed approaches using four datasets of Australian PV power half-hourly data, for one year. We compare their performance with two persistence models used as baselines, and a SVR based prediction model.

This paper is organized as follows. Section II provides an overview of the related work. Section III describes the data used in our study. Section IV presents our proposed approaches for forecasting the solar power output. Section V presents and discusses the results, and Section VI concludes the paper.


Forecasting solar power output from PV systems is a relatively new topic that is receiving significant attention due to the growing production and use of solar energy. In this section we briefly review the previous work on PV power prediction. Most of the existing approaches predict the solar irradiance and use it to estimate the power output (indirect prediction) but there are also some recent approaches that directly predict the PV power output.

Inman et al. [14] reviewed methods for solar power forecasting and classified them into five main groups: statistical (regressive) methods (e.g. auto regressive, moving average, and combinations of them such as ARIMA), methods based on artificial intelligence techniques (e.g. NNs, nearest neighbor), numerical weather prediction methods, remote sensing methods (e.g. satellite and statistical satellite) and local sensing methods (e.g. sky-imager).

Pedro and Coimbra [5] predicted the solar power 1 and 2 hours ahead from a time series of previous solar power values only, without using any exogenous variables. They compared the performance of four methods: ARIMA, k nearest neighbor, NN trained with the backpropagation algorithm and NN trained with a genetic algorithm. They conducted an evaluation using data for two full years and found that the two NN based methods outperformed the other methods, and that the NN trained with the genetic algorithm was the most accurate

prediction model. The two NN approaches obtained Mean Absolute Error (MAE) in the range of 42.96 - 61.92 kW for 1 hour ahead prediction and 62.53 - 87.76 kW for 2 hours ahead prediction for a 1 MW PV power plant.

Chen et al. [2] introduced a new approach for 1 to 24 hours ahead solar power prediction based on Radial Basis Function NN (RBFNN). At first, they categorized the days into sunny, cloudy and rainy using self-organizing map NNs and based on the weather predictions of solar irradiance and cloudiness. Then, a separate RBFNN prediction model for each group was trained to predict the 24 hourly PV power outputs for the next day.

Shi et al. [9] proposed a similar approach – the days were clustered into four groups (clear-sky, cloudy, foggy and rainy) and a separate SVR prediction model was built for each group. The obtained Mean Relative Error (MRE) was between 4.85% (for sunny day) and 12.42% (for cloudy day).

Chow et al. [8] applied NNs for predicting the PV power output 10 and 20 minutes ahead. As inputs to the NNs they used solar irradiation, temperature, solar elevation angle and solar azimuth angle. They developed multi-layer perceptron with one hidden layer, trained with the backpropagation algorithm, with early stopping criterion based on validation set to avoid overtraining. The results were promising and showed that NNs can successfully model the nonlinear relationship between the meteorological parameters and the PV solar power output.

Mandal et al. [7] used wavelet transform in conjunction with RBFNNs. They firstly decomposed the highly fluctuating PV power time series data into multiple time-frequency components. The one hour ahead decomposed PV power output was then predicted using the decomposed components, as well as previous solar irradiation and temperature data. The final prediction was generated by applying the inversed wavelet transform. The results showed good accuracy, with the combination of wavelet transform and RBFNN outperforming RBFNN without wavelets. Mellit et al. [15] presented a different wavelet based approach, called wavelet network. Instead of decomposing the data and applying NNs to predict each component, they used wavelets as activation functions in the NNs. The approach was effective, achieving Mean Absolute Percentage Error (MAPE) of about 6%.

Zeng and Qjao [10] studied the application of SVR for solar power forecasting. They applied SVR to predict the atmospheric transmissivity using historical transmissivity and other meteorological data. The predicted transmissivity was then converted back to solar power according to the latitude of the PV site and time of the day. The evaluation showed that SVR was more accurate compared to ARIMA and RBFNN.

Approaches based on fuzzy logic were also proposed. Jararzadeh et al. in [16] investigated the application of interval type-2 Takagi-Sugeno-Kang fuzzy systems. Using temperature and solar irradiance as inputs, they predicted the output of PV plants under different operating conditions, and showed better results than ARIMA. Yona et al. [17] proposed a hybrid approach by combing NNs and fuzzy theory. They first applied a fuzzy model to estimate the hourly insolation using different


weather variables such as clouds, humidity and temperature. The output of the fuzzy model was then fed to a recurrent NN, to predict the hourly power output of the PV plant.

Yang et al. [11] integrated SOM, SVR, and fuzzy inference to develop a hybrid approach for one day ahead solar power prediction. SOM and SVR were applied to classify the historical input data and to develop the prediction model, respectively. The fuzzy inference was used to select the best model from a group of trained SVRs, depending on the available weather predictions. An evaluation using one year of solar data showed that the hybrid method outperformed NN and SVR.


A. Case Studies

We use data collected from the 1.22 MW PV system installed at the St. Lucia campus of the University of Queensland in Brisbane, Australia, and available from [18]. This is the largest PV system in Australia and consists of more than 5000 polycrystalline silicon solar panels across four different sites. We use data from all four sites and consider the data from each site as a separate case study.

For each case study, we use the data from 1st January to 31st December, 2013. The data represents the power output of the PV arrays at the given location, in 30 minutes intervals. We only consider the data between 7:00 am to 5:00 pm as most of the data values outside this time window are either zero or not available. Thus, there are 7,300 data points (365 days × 20 points per day) available for each case study. The data was normalized between -1 and 1.

B. Data Characteristics

Fig. 1 shows the PV power output for the four case studies for one week – from Monday 4th February to Sunday 10th February, 2013. We can see that the range of solar power varies for the different case studies – it is highest for case study 1 (0.09-430 kW) and lowest for case study 2 (0.03-90 kW). This is due to the different capacity of the solar panel arrays installed at the four sites (e.g. the number of modules installed at site 1 is about five times higher compared to the number of modules at site 2).

We can also observe that the four graphs show relatively similar patterns for the same day. This is as expected since the four sites are located close to each other and receive similar amount of solar irradiation. In general, for a typical sunny day (such as the first day in Fig. 1), the solar power starts to increase at the beginning of the day, reaches its peak around midday when the solar irradiance is highest and then gradually declines until the end of the day.

By comparing the solar power profiles of different days for a given case study, we can see that the solar power is highly volatile – it changes rapidly following the variations of the factors affecting its production, such as solar irradiance, cloud coverage, rainfall, temperature, humidity etc. The variable nature of the data makes the prediction task very challenging.

Fig. 1. PV power output for each case study for one week (from Monday 4th February to Sunday 10th February, 2013)

C. Training, Validation, and Testing Sets

We divided the data for each case study into three non-overlapping subsets: training (Dtrain), validation (Dvalid), and testing (Dtest). The data split was 50%-25%-25%, following the recommendation in [19], resulting in 183 instances in Dtrain, 91 in Dvalid and 91 in Dtest. One instance is 20-dimensional vector, where each value corresponds to the half-hourly PV power output for a single day from 7:00 am to 5:00 pm.

0 100 200 300 400 500 1 20 39 58 77 96 115 134 Solar   Power   [k W]

Time Lag (1 lag = 30 min) a) Case study 1 0 20 40 60 80 100 1 20 39 58 77 96 115 134 Solar   Power   [k W]

Time Lag (1 lag = 30 min) b) Case study 2 0 50 100 150 200 250 300 350 400 1 20 39 58 77 96 115 134 Solar   Power   [k W]

Time Lag (1 lag = 30 min) c) Case study 3 0 50 100 150 200 250 300 350 1 20 39 58 77 96 115 134 Solar   Power   [k W]

Time Lag (1 lag = 30 min) d) Case study 4


T pred best used T appr outp A. A T the p NNs to arch show NN ahea T half-next Thus the h with pred F prop W the n each of n the num weig sepa (LM grad faste grad betw The training diction models architecture f d to evaluate th IV This section roaches for for puts.

Approach 1 (A

The first appro prediction mo s instead of a the random itecture. Adeo wed that ensem

for time series ad prediction. The key idea o -hourly PV po t day by using

s, a single NN half-hourly PV h V nodes in dictions for day

Fig. 2. Architectu posed approaches We first const number of ne h structure NNs (we use ensemble mber of hidden ghts. Each of arately on the M) algorithm [2 dient descent er convergenc dient descent a ween them bas

set was use s, the validatio for the NN en he accuracy of V. PROPOSED describes ou recasting the d A1) oach, A1, appl

odel. The moti single NN is t initialization odatoa et al. [ mbles of NNs s prediction, e of A1 is to sim ower outputs f g the power N from the en V power outp n it, and 20 y d+1, as show ture of a single NN s A1 and A2. A3 ha no tructed V diff eurons in the , we then bu ed n=20) as sh has the sam

neurons, but i the n membe training data 21]. We chose backpropaga ce. The LM a and the Gaus sed on the com

ed to develo on set was us nsemble, and t f the predictio APPROACHES ur three prop day ahead half

ies an ensemb ivation of usin to reduce the n of weigh 13] and Ferrei were more ac especially for m multaneously from 7:00 am outputs from nsemble has 2 put in day d, 0 output nod wn inFig. 2. N, part of the NN as the same archit ode. ferent NN stru hidden layer uilt an ensemb hown in Fig. 3 me structure is initialized to ers of ensemb

using the Lev e LM over the ation algorithm algorithm com ss-Newton alg mplexity of th op and train sed to identify the testing set n models. S posed NN b f-hourly PV p ble of NNs to b ng an ensemb sensitivity of hts and netw ira et al. [20] ccurate than s more than one predict the tw to 5:00 pm fo the previous 20 input node one hidden des providing N ensemble, for th tecture but one ou

uctures by var from 1 to V

ble that con 3. Each memb

, i.e. the s o different ran ble was tra venberg Marqu e standard stee m because o mbines the stee

gorithm, switc he error surfac n the fy the t was based power build ble of f NNs work have single e step wenty or the day. es for layer g the he utput rying V. For nsists ber of same ndom ained uardt epest of its epest ching ce. It pres conv of t learn W tran func was T the com pred h g j=1, T valid the l to pr B. A T clus each PV pow mete sign with F is pa the of th mea serves the a vergence of th the steepest ning rate. We used the sfer function ction for the o set to 3000 an To predict the individual p mbined by tak diction for tim

, , … enerated by ,...,n. The performan dation set. Th lowest predict redict the testi

Approach 2 (A

The main idea ters, and then h cluster. The output for di wer output ha eorological an nificant variati h clear sky [9] Fig. 4 presents artitioning the X-Means algo he popular k-M ans doesn’t re advantages o he Gauss-New gradient desc Fig. 3. An ense following NN for the hidd output neuron nd the regulari e half-hourly predictions of ing their med me (half an h

… , , , w an ensemble nce of each en he best perform

tion error (MA ing data. A2) a of A2 is to n build a sep use of cluster ifferent types as different nd environme ions during ra .

s the main step e available tra orithm [22]. X Means algorit equire the num

of both algo wton algorithm cent when u emble of NNs EV NN parameters den neurons a ns; the maxim ization param power output f the ensem dian value (se hour) h for da where , i e member

nsemble wa ming ensembl AE), was then

first group th parate NN pre ring is motivat of days. In characteristic ental conditio ainy days, clou

ps of approac aining data in X-Means is an thm. In contra mber of clust orithms: the m and the sta used with a


s: tangent sig and linear tra mum training e eter was set to ts for a given mble members ee Fig. 3), i.e ay d+1 is: s the predictio , h=1,...,20 as evaluated o e, i.e. the one n selected and

he data into se ediction mode ted by the diff particular, the cs under diff ons [2] and s udy days and h A2. The firs

nto k clusters u n extended ve ast to k-Mean ters and the i

fast ability small gmoid ansfer epoch o 0.9. n day, s are e. the = on for 0 and on the e with d used everal el for ferent e PV ferent shows days t step using ersion ns, X-initial


centr the num cons and outp T from insta deve appl to pr next T iden new pred C. A I simu day, A node mak used mean data The mak roids to be su data. The ap mber of cluste sistent with the

rainy) that h put.

F The second ste m the training,

ances that be eloping a pred ies the same m redict simultan t day. To predict a n tifies the NNs instance, an diction. Approach 3 (A In contrast ultaneously th A3 does this i A prediction m e) is built for kes a predictio d to make the ns that the pre and is appen last twenty s ke prediction f

upplied by the pplication of ers for each c

e expected 3 t have the mos

Fig. 4. Main step ep is assigning validation and elong to eac diction model method as A1 neously all ha new instance s ensemble tra nd then uses A3) to approache he half-hourly iteratively. model (an ens

a single step on for time h e prediction f edicted data fo nded to the en samples from for time h+1,

e user, but est X-Means ge case study – types of days st influence o ps of approach A g a cluster lab d testing sets, ch cluster. T for each clust – it creates an alf-hourly pow

from the tes ained for the c

this ensembl es A1 and PV power ou semble of NN ahead predic +1, and this for time h+2 or time h is co nd of the ava the appended and this con

timates them enerated the s

3. This resu (clear sky, clo on the PV p


bel to the insta and separatin The third ste ter. To do thi n ensemble of wer outputs fo ting data, A2 cluster label o le to produce A2 that pr

utputs for the Ns with one ou ction. At time prediction is 2 and so on. onsidered as a ailable actual d data are use

tinues for all from same ult is oudy, power ances ng the ep is is, A2 f NNs or the first of the e the redict next utput h, it then This actual data. ed to time poin the p the p fore appr elec T A2. mem T stud A. E T perf Rela w for d (day outp B. P T 20 h appr T mos term with com that A1 f the r and A size to ea nts from the f prediction for previously rec ecasted data fo roach has bee

tricity load in The NNs ense

The only diff mber – it has o V This section p dies. All report

Evaluation M To assess the formance mea ative Error (M where: and day d at time ys) in the test d puts for a day (

Performance E Table I presen hours from th roaches. The results sh st accurate, fo ms of average M h 1.05%, and mparison acros A3 outperform

for case study range of 16.86 16.92% - 17.5 TABLE I Case Study MAE kW 1 72.55 2 16.19 3 58.34 4 59.66 Avg. 51.69 A possible rea of the cluster ach cluster wa forecasting ho the time h+1 corded actual d or times 1 to h en shown to [23]. emble is built ference is in t one output nod . RESULTS AN presents and d

ted results are

Metrics e predictive ac sures, Mean A MRE), as define 1 1 1 1 d are the ac h, respectivel data; H is the (H=20 for our Evaluation nts the forecas e forecasting how that overa

ollowed by th MRE, A3 outp the relative r ss A1, A2, and med both A1 a y 1. The MRE 6% - 17.92% f 58% for A3. . ACCURACY OF T A1 E MRE % M k 5 16.86 78 9 17.92 17 4 17.24 63 6 17.65 61 9 17.42 55

ason for the lo rs. The numbe as: 48 for clus

orizon. This m of day d+1, w data till day d h for day d+1. be effective t in the same the architectur de, not 20 as in ND DISCUSSIO discusses the results on the accuracy, we Absolute Error ed below:

ctual and pred ly; D is the nu

total number r task) and R i

sting accuracy horizon, for t all the iterativ he non-iterati performed A2 w

results for M d A3 for each

and A2 for all

E for the four for A1, 18.26% THE PROPOSED A A2 MAE kW MRE % 8.72 18.30 7.66 19.54 3.12 18.65 1.72 18.26 5.31 18.69 ower accuracy er of training ster 1, 51 for c means that to m where h>1, w

and the previ A similar iter for forecastin way as for A re of the ense n A1 and A2. ON

results for all e testing data.

used two stan r (MAE) and M


dicted power o umber of insta

of predicted p s the data rang

y, averaged ov the three prop ve approach, A ive A1 and A with 7.77% an AE are simil h case study s case studies ex case studies a % - 19.54% fo PPROACHES A3 MAE kW MRE % 72.77 16.92 15.61 17.27 58.10 17.17 59.43 17.58 51.48 17.24 y of A2 is the examples assi cluster 2 and 8 make we use ously rative ng of A1 and emble l case ndard Mean output ances power ge. ver all posed A3, is A2. In nd A1 ar. A shows xcept are in or A2, small igned 81 for


cluster 3, for the first case study. Using a larger training data is likely to improve the clustering results, and in turn, the predictive accuracy results

In terms of time required to build the prediction models, the more accurate approaches A1 and A3 needed less time than the

less accurate A2. This is because A1 and A3 create one

prediction model, while A2 first clusters the data and then

creates a separate prediction model for each cluster (3 clusters and 3 prediction models for our data). Thus, as expected, the time required to build a prediction model for A2 was about 3

times higher than the time for A1 and A3 (10 minutes compared

to 3 minutes). This time includes the training time for all n

ensemble members and the aggregation of their outputs.

C. Ensemble of NNs vs single NN

To assess the benefit of using an ensemble of NNs, we compared its performance with the performance of a single NN, using the same experimental setting. The accuracy results for A1, A2,and A3 with a single NN are shown in Table II. A

graphical comparison of the MRE results for the ensemble (Table I) and single NN (Table II) is presented in Fig. 5.

The results show that using an ensemble of NNs resulted in higher accuracy than using a single NN, for all three approaches. The improvements in terms of MRE averaged over all case studies are 9.43%, 36.48% and 7.18% for A1, A2,and

A3, respectively. The computational cost of training an

ensemble is higher than for a single NN, but suitable for both offline and online practical applications: 3-10 minutes for training an ensemble, compared to 10-30 sec for training a single NN.


A1 A2 A3 Case Study MAE kW MRE % MAE kW MRE % MAE kW MRE % 1 77.76 18.08 112.19 26.08 77.64 18.05 2 16.62 18.38 26.38 29.19 16.94 18.74 3 71.37 21.09 76.51 22.61 64.42 19.04 4 63.17 18.69 81.59 24.14 61.04 18.06 Avg. 57.23 19.06 74.17 25.51 55.01 18.47

Fig. 5. Ensemble of NNs vs single NN in the proposed approaches Hence, we can conclude that the use of ensemble of NNs instead of a single NN is beneficial for our proposed

approaches. A1, A2,and A3 achieved better prediction accuracy

using an ensemble at acceptable computational cost.

D. Comparative Study

We compared the performance of our approach with two persistence models (baselines) and a SVR based iterative approach.

The first persistence model, Bpday, considers the half-hourly

PV power outputs from the previous day d, as the predictions for the next day d+1, i.e. the predictions for , … , are given by , … , .

The second persistence model, Bweek, considers the

half-hourly PV power outputs from the same day, one week before as the predictions for the next day, i.e. the predictions for

, … , are given by , … , .

SVR, on the other hand, is one of the state-of-the art machine learning algorithms for solving regression problems. SVR based prediction models have been shown to achieve promising accuracy for solar power prediction in [9, 10]. In order to develop the SVR prediction model, we used the Weka’s implementation of the SMOreg algorithm with RBF kernel as described by Shevade et al. [24]. We generated the predictions iteratively, as in approach A3 (see Section IV-C).

Table III shows the performance of the three methods used for comparison. Fig. 6 visually compares the MRE results from Table I and Table III, for all methods.


SVR Bpday Bweek Case Study MAE kW MRE % MAE kW MRE % MAE kW MRE % 1 77.63 18.04 79.39 18.45 88.12 20.48 2 16.25 17.98 17.37 19.21 19.01 21.03 3 59.17 17.49 61.44 18.16 68.63 20.28 4 58.22 17.22 60.07 17.77 68.36 20.22 Avg. 52.82 17.68 54.57 18.40 61.03 20.50

Fig. 6. Comparison of MRE for different methods

We can see that A3 compares favorably with SVR in all

case studies except case study 4, where the performance of the two methods is similar. A3 also outperformed the two baselines

0.00 5.00 10.00 15.00 20.00 25.00 30.00 A1 A2 A3 MRE   (%) ensemble single NN 0 5 10 15 20 25

Case study 1 Case study 2 Case study 3 Case study 4





Bpday and Bpweek, achieving an improvement in MRE of6.32%

and 15.94%, respectively. The second best approach overall is A1, closely followed by the iterative SVR, and then Bpday, A2

and Bpweek. A2 is outperformed by the iterative SVR and also

slightly outperformed by the Bpday baseline. This highlights that

A2 did not performed well enough.


In this paper, we presented three approaches (A1, A2, and

A3) for forecasting the half-hourly PV power output for the

next day. A1 uses an ensemble of NNs to predict all 20 outputs

for the next day at the same time. A2 also predicts all outputs

for the next day simultaneously, but firstly partitions the data into a set of clusters using the X-Means algorithm and then builds an ensemble of NNs for each cluster. A3 uses an iterative

methodology, where an ensemble of NNs forecasts one output at a time, which is then used for the predictions of the next points from the forecasting horizon.

We conducted a comprehensive evaluation of the proposed approaches A1, A2 and A3 using four Australian solar datasets

for one year, and compared their performance with an iterative SVR approach and two baselines.

We found that A3 was the most accurate approach, and that

A1 also showed promising results. The MRE of A1 and A3

averaged for 1-20 steps ahead prediction was 16.86% - 17.92% and 16.92% - 17.58%, respectively. A2, on the other hand,

showed relatively poor performance and was outperformed by the baseline Bpday for 3 out of 4 case studies. However, the

accuracy of A2 is likely to be improved by increasing the

number of training instances for each cluster. We also investigated the effect of using an ensemble of NNs instead of a single NN, and showed that the use of ensemble improved accuracy.

The iterative SVR also achieved good accuracy; its MRE was between 17.22% and 18.04%. The good performance of the two iterative approaches, A3 and SVR, indicates that

iterative approaches are better option for multiple step ahead prediction than non-iterative approaches.

In conclusion, considering both accuracy and time to train the prediction model, we found that A3 was the most promising

method for practical applications. Our future studies will investigate the application of wavelet based approaches for solar power forecasting.


This research was partially supported by a research award from the Clean Energy and Intelligent Networks Research Cluster at the University of Sydney.


[1] M. H. Nehrir, C. Wang, K. Strunz, A. Aki, R. Ramakumar, J. Bing, et al., "A review of hybrid rebewable/alternative energy systems for electric power generation: configurations, control and applications,"

IEEE Transactions on Sustainable Energy, vol. 2, pp. 392-403, 2011. [2] C. Chen, S. Duan, T. Cai, and B. Liu, "Online 24-h solar power

forecasting based on weather type classification using artificial neural networks," Solar Energy, vol. 85, pp. 2856-2870, 2011.

[3] A. Mellit and A. M. Pavan, "A 24-h forecast of solar irradiance using artificial neural network: application for performance prediction of a

grid-connected PV plant at Trieste, Italy," Solar Energy, vol. 84, pp. 807-821, 2010.

[4] (2013). Australian Renewable Energy Agency (ARENA). Available: http://arena.gov.au/about-renewable-energy/solar-energy/

[5] H. T. C. Pedro and C. F. M. Coimbra, "Assessment of forecasting techniques for solar power production with no exogeous inputs," Solar Energy, vol. 86, pp. 2017-2028, 2012.

[6] Y. Dazhi, P. Jirutitijaroen, and W. M. Walsh, "Hourly solar irradiance time series forecasting using clound cover index," Solar Energy, vol. 86, pp. 3531-3543, 2012.

[7] P. Mandal, S. T. S. Madhira, A. U. Hague, J. Meng, and R. L. Pineda, "Forecasting power output of solar photovooltaic system using wavelet transform and artificial intelligence techniques," Procedia Computer Science, vol. 12, pp. 332-337, 2012.

[8] S. K. Chow, E. W. Lee, and D. H. Li, "Short-term prediction of photovoltaic energy generation by intelligent approach," Energy and Buildings, vol. 55, pp. 660-667, 2012.

[9] J. Shi, W.-J. Lee, Y. Lin, Y. Yang, and P. Wang, "Forecasting power output of photovoltaic systems based on weather classification and support vector machines," IEEE Transactions on Industry Applications,

vol. 48, pp. 1064-1069, 2012.

[10] J. Zeng and W. Qiao, "Short-term solar power prediction using a support vector machine," Renewable Energy, vol. 52, pp. 118-127, 2013. [11] H.-T. Yang, C.-M. Huang, Y.-C. Huang, and Y.-S. Huang, "A

weather-based hybrid method for 1-day ahead hourly forecasting of PV power output," IEEE Transactions on Sustainable Energy, vol. 5, pp. 917-926, 2014.

[12] I. Koprinska, M. Rana, A. Troncoso, and F. Martínez-Álvarez, "Combining pattern sequence similarity with neural networks for forecasting electricity demand time series," in International Joint Conference on Neural Networks (IJCNN 2013), Dallas, USA, 2013. [13] P. J. L. Adeodato, A. L. Arnaud, G. C. Vasconcelos, R. C. L. V. Cunha,

and D. S. M. P. Monteiro, "MLP ensembles improve long term prediction accuracy over single networks," International Journal of Forecasting, pp. 661-671, 2011.

[14] R. H. Inman, H. T. C. Pedro, and C. F. M. Coimbra, "Solar forecasting methods for renewable energy integration," Progress in Energy and Combustion Science, vol. 39, pp. 535-576, Jul 2013.

[15] A. Mellit, M. Benghanem, and S. Kalogirou, "An adaptive wavelet-network model for forecasting daily total solar-radiation," Applied Energy, vol. 83, pp. 705-722, 2006.

[16] S. Jafarzadeh, M. S. Fadali, and C. Y. Evrenosoglu, "Solar power prediction using interval type-2 TSK modeling," IEEE Transactions on Sustainable Energy, vol. 4, pp. 333-339, 2013.

[17] A. Yona, T. Senjyu, T. Funabashi, and C.-H. Kim, "Determination method of insolation prediction with fuzzy and applying neural network for long-term ahead PV power output correction," IEEE Transactions on Sustainable Energy, vol. 4, pp. 527-533, 2013.

[18] (2014). University of Queensland Solar Data. Available: http://www.uq.edu.au/solarenergy/

[19] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning vol. 2. New York: Springer, 2009.

[20] T. A. E. Ferreira, G. C. Vasconcelos, and P. J. L. Adeodato, "A new intelligent system methodology for time series forecasting with artificial neural networks," Neural Processing Letters, vol. 28, pp. 113-129, 2008. [21] M. T. Hagan and M. B. Menhaj, "Training feedforward networks with the marquardt algorithm," IEEE Transactions on Neural Networks, vol. 5, pp. 989-993, 1994.

[22] D. Pelleg and A. W. Moore, "X-means: Extending K-means with efficient estimation of the number of clusters," in 17th International Conference on Machine Learning (ICML), 2000, pp. 727-734.

[23] M. Rana, I. Koprinska, and A. Troncoso, "Forecasting hourly electricity load profile using neural networks," presented at the International Joint Conference on Neural Networks (IJCNN), Beijing, China, 2014. [24] S. K. Shevade, S. S. Keerthi, C. Bhattacharyya, and K. R. K. Murthy,

"Improvements to the SMO algorithm for SVM regression," IEEE Transactions on Neural Networks, vol. 11, 2000.


Related documents

But we are concerned that other aspects of these practices, such as casting tributes and other objects into the water at the same time as the ashes, could harm the environment or

A water collection basin or impoundment is a primary means of reducing potential water quality problems. It should be the goal of each container nursery operation that no

ed C entre (NSAC) Newham C ollege of F urther E duc ation North W est K ent C ollege of T echnology Oaklands C ollege Richmond Upon Thames C ollege South Thames C

It is clear that Lieutenant Kotler is getting close to Bruno’s mother because he is always in the living room making jokes with her; whenever the Commandant is away, he hangs around

Contract 55-15smc, Transportation Planning, Engineering & Consulting Services 3 Following successful negotiations, the Owner will prepare a purchase order for the agreed

The purpose of this study was to investigate the effect of perceived organizational support on affective commitment and job performance, with mediating role of

Error of Omission: One of the domain theory rules used to create a compiled rule needs to be.. generalized by dropping a condition that is not present in a performance example

While business associates were not referenced in HIPAA as first enacted, the Secretary proposed the regulation of "business partners"- parties that maintained