Indonesian J our nal of Electrical Engineering and Computer Science V ol. 25, No. 2, February 2022, pp. 900 ∼ 909 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v25.i2.pp900-909 ❒ 900 Day-ahead solar irradiance f or ecast using sequence-to-sequence model with attention mechanism So wkarthika Subramanian 1 , Y asoda Kailasa Gounder 1 , Sumathi Linganathan 2 1 Department of Electrical and Electronics Engineering, Go v ernment Colle ge of T echnology , Coimbatore, India 2 Department of Computer Science and Engineering, Go v ernment Colle ge of T echnology , Coimbatore, India Article Inf o Article history: Recei v ed Jul 18, 2021 Re vised No v 22, 2021 Accepted Dec 9, 2021 K eyw ords: Attention Long short-term memory Sequence-to-sequence LSTM Solar irradiance forecast ABSTRA CT The increasing inte gration of distrib uted ener gy resources (D ERs) into po wer grid mak es it signicant to forecast solar irradiance for po wer sys tem planning. W ith the adv ent of deep learning techniques, it is possible to forecast solar irradiance accu- rately for a longer time. In this paper , day-ahead solar irradiance is forecasted using encoder -decoder sequence-to-sequence models with attention mechanism. This study formulates the problem as structured multi v ariate forecasting and comprehensi v e e x- periments are made with the data collected from National Solar Radiation Database (NSRDB). T w o error metrics are adopted to measure the errors of encoder -decoder sequence-to-sequence model and compared with smart persistence (SP), back prop- ag ation neural netw ork (BPNN), recurrent neural netw ork (RNN), long short term memory (LSTM) and encoder -de coder sequence-to-sequence LSTM with attention mechanism (Enc-Dec-LSTM). Compared with SP , BPNN and RNN, Enc-Dec-LSTM is more accurate and has reduced forecast error of 31.1%, 19.3% and 8.5% respecti v ely for day-ahead solar irradiance forecast with 31.07% as forecast skill. This is an open access article under the CC BY -SA license . Corresponding A uthor: So wkarthika Subramanian Department of Electrical and Electronics Engineering, Go v ernment Colle ge of T echnology Coimbatore, India Email: so wkarthika@gct.ac.in 1. INTR ODUCTION Inte gration of solar electricity kno wn as distrib uted ener gy resources (DERs) into po wer grid has g ained a rapid de v elopment in recent years due to reduction in manuf acturing cost and increased ef cienc y of photo v oltaic (PV) panels. The amount of electricity that can be generated from DERs is al w ays a stochastic in nature because of its dependenc y on weather parameters. This further leads to a challenge for grid operators in estimating generati o n, distrib ution and scheduling of po wer generation. Therefore, an accurate day-ahead forecast of solar irradiance with big data and deep learning model solv es this problem. F orecast models i n literature for solar i rradiance are persistence model, ph ysical model and statist ical model. V ery short-t erm forecast (seconds to less than 30 minutes) is popularly predicted with persistence model [1], [2]. As accurac y of persistence model decreases with increase in forecast horizon, it is not preferred for 24 hours day-ahead forecast. In ph ysical model or numerica l weather prediction models [1], [3], the state of the atmosphere is described by mathematical equations which require numerical methods to solv e. F orecast emplo yed with ph ysical model leads to erroneous result for sudden change in v alues of meterological v ariables such as relati v e humidity , wind speed and wind direction. Articial neural netw ork (ANN) based multilayer perceptron model [4], [5] with Le v enber g-Marquardt algorithm w as proposed to forecast 24 ho ur s ahead solar J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 901 irradiance and found that the usage of meterological parameters as input v ariables gi v es more accurac y in forecast. Input v ari ables with higher dimension [6]-[9] (up to 900 inputs) are used with ANN models of dif ferent architecture to predict short ter m global solar irradiance of 20% reduction in errors. Deep learning models are the subset of machine learning and these models on s olar irradiance forecast results with higher accurac y comapared to machine learning models. A method of day-ahead solar irradiance forecast using long short-term memory (LSTM) netw ork with weather v ariables as feature v ectors w as de v eloped [10] and results pro v e that LSTM outperforms all the other con v entional forecast methods in terms of forecast accurac y . Jeon et al. [11] proposed an LSTM based deep learning model for solar irradiance forecast with weather v ariables and also solar irradiance of the pre vious day as feature v ectors. Simulation result s ho ws the impro v ement in forecast accurac y if solar irradiance of pre vious day is also used as input feature. Gao et al. [12] proposed g ated recurrent unit (GR U) based model for hourly day-ahead solar irradiance forecast using weather v ariables. In this paper , studies are made to forecast day-ahead solar irradiance using LSTM based encoder - de- coder models with attention mechanism. Intially , datas are cleaned and con v erted into structured multi v ariate problem to train with encoder -decoder sequence-to-sequence models. Based on pearson correlation coef - cient, input v ariables a re selected from the list of meteorological parameters. Comprehensi v e e xperiments are made to determine the forecast accurac y considering meterological parameters as input v ariable. Experiments ha v e sho wn that LSTM based encoder -decoder sequence-to-sequence models with attention mechanism ha v e reduced errors comparati v ely . F orecast horizon [14] from the perspecti v e of decision making acti vity in mi- crogrid or smartgrid are classi ed as v ery short-term forecast, short-term forecast, medium-term forecast and long-term forecast. V ery short-term forecast is used in real time monitoring of photo v oltaic po wer and the forecast ho r izon is from fe w seconds to minutes ahead. Short-term forecast is used in decsion making applica- tions in v olv ed in po wer system operation such as economic dispatch, unit commitment. F orecast horizon for short-term forecast is up to 48 to 72 hours ahead. Schedule and maintenance of po wer plant are planned with medium-term forecast and its horizon is upto one week ahead. Long term forecast helps in the assessment of solar ener gy and its horizon is from months to years. Unit commitment [15], [16] for po wer plants such as biomass, nuclear , and coal, are one day-ahead and for po wer plants such as g as and oil are hour ahead. This time horizon is formul ated depending on their st artup and s h ut do wn times. In such a case with rene w able inte gration into grid, unit commitment and economic dispatch decisions v ary depending on solar forecasts. In this paper , day-ahead solar irradiance is forecasted using dif ferent deep-learning techniques. In day- ahead forecast pre vious day’ s data is used as input to forecast irradiance of ne xt 11 hours with a resolution of one hour . In general, geographical locations [17] also deter mine the forecast error and hence the models described here are tested for dif ferent locations with dif ferent climatic conditions also. This paper is or g anised as follo ws: methodology is described in section 2, description of data and preprocessing in section 3, e xperiments and results in section 4 and conclusion and future w ork in section 5. 2. METHOD Long short term memory netw ork is base for all the other models. Hence, architecture of LSTM and LSTM based encoder -decoder sequence-to-sequence models with attention mechanism and a benchmark algo- rithm are described in detail. Under benchmark algorithm, smart persistence is used to compare the proposed method. 2.1. Smart persistence-benchmark algorithm F orecast error v aries with dataset, location and horizon. Hence for a good comparison, benchmarking algorithm such as smart persistence model (SP) [18] or scaled persistence model is suggested. Smart persistence model suggest that the predicted v alue at the ne xt moment ˆ G ( t + h ) is the product of clear -sk y inde x k cs ( t ) and clear -sk y irradiance at ne xt moment G cs ( t + h ) . k cs ( t ) = G ( t ) G cs ( t ) (1) ˆ G ( t + h ) = k cs ( t ) G cs ( t + h ) (2) where k cs ( t ) is the clear -sk y inde x , G cs ( t ) is the clear -sk y irradiance and h in (2) is the forecast horizon. Day-ahead solar irr adiance for ecast using sequence-to-sequence model with ... (Sowkarthika Subr amanian) Evaluation Warning : The document was created with Spire.PDF for Python.
902 ❒ ISSN: 2502-4752 2.2. Encoder -decoder sequence-to-sequence ar chitectur e T raditional neural netw ork lik e back propag ation neural netw ork (BPNN) , do not ha v e memory to understand and process sequential data. This w as o v ercome by recurrent neural netw ork (RNN) algorithm [19]. RNNs ha v e loops within them and mak es the informations to pe rsist. Ho we v er RNN suf fers from v anishing and e xploding gradient problems that pre v ents it from learning lar ge sequences. Hochreiter et al. proposed LSTM [20] netw ork that can process sequential data ef fecti v ely with recurrent neural netw ork as sho wn in Figure 1. Figure 1. LSTM cell structure Input v ariables of a single LSTM units are current time step input v ector X t , output of the pre vi ous LSTM unit h t − 1 and memory of the pre vious LSTM unit also called cell state c t − 1 . The outputs of a single LSTM unit are output of the hidden layer h t and memory at the current time step c t . Each LSTM unit proces ses the information through for get g ate ( f t ), input g ate ( i t ) and output g ate ( o t ) according to (3), (4) and (5). f t = σ ( W xf or x t + W hf or h t − 1 + b f or ) (3) i t = σ ( W xinp x t + W hinp h t − 1 + b inp ) (4) o t = σ ( W xout x t + W hout h t − 1 + b out ) (5) Where W xf or , W hf or are for get g ate’ s weight matrix, W xinp , W hinp are input g ate’ s weight matrix and W xout , W hout are output g ate’ s weight matrix; b f or , b inp and b out are bias v alues of for get g ate, input g ate and output g ate respecti v ely . σ represents sigmoid acti v ation function. F or get g ate ( f t ) decides, which part of the informations are to be erased and which part of the informations are to be retained and outputs a number between 0 and 1 through sigmoid function. Input g ate ( i t ) and for get g a te ( f t ) species the part of the informations to be added with the cell state. Finally , output g ate ( o t ) decides the information output from cell state. Cell state c t and current output of hidden layer are calculated by (6) and (7), c t = f t ⊙ c t − 1 + i t ⊙ tanh ( W xcel l x t + W hcel l h t − 1 + b cel l ) (6) h t = o t ⊙ tanh ( c t ) (7) where ⊙ represents the hadamard product that performs element-wise matrix multiplication. Encoder -decoder sequence-to-sequence architecture uses LSTM (Enc-Dec-LSTM) as encoder com- ponent, Luong’ s att ention layer , another LSTM netw ork as decoder component and a dense layer as sho wn in Figure 2. Encoder -decoder sequence-to-sequence archi tecture although de v eloped for nat ural language trans- lation, it had been succesfully applied for time-series forecasting [21] such as ai r -quality , and traf c prediction. Encoder [22] encodes the information from i n put into a x ed length v ector . The nal outputs of the encoder are Indonesian J Elec Eng & Comp Sci, V ol. 25, No. 2, February 2022: 900–909 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 903 discarded and the internal state and hidden state combinedly called x ed length v ector is fed into the decoder . Decoder is also gi v en pre vious hour of the tar get and trained to predict ne xt hour . This process of training is called teacher -forcing. Figure 2. Encoder -decoder sequence-to-sequence architecture Figure 3. Attention layer 2.3. Attention mechanism According to Luong et al. [23] the potential issue of the encoder is, by compressing all necessary information of input into a x ed-length v ector may f ail to generate long sequence from the decoder . Attention layer as sho wn in Figure 3 allo ws the model to access all the past hidden states of encoder instead of the last hidden layer alone. The alignment score e t,i for Luong’ s attention is calculated as in (8), e t,i = h T dt · h i (8) a t,i = exp ( e t,i ) P N j =1 exp ( e t,j ) (9) C t = N X i =1 a t,i h i (10) where h dt is current tar get state or t th hidden sta te of decoder and h i is i th hidden state of encoder . Attention weight a t,i as in (9) is calculated by softmaxing the alignment score to sum up to 1. Conte xt v ector as in (10) is computed by element-wise multiplication of ith hidden state of encoder and attention weight. The conte xt v ector is then concatenated with current tar get state h dt and is fed into a fully connected feed-forw ard netw ork Day-ahead solar irr adiance for ecast using sequence-to-sequence model with ... (Sowkarthika Subr amanian) Evaluation Warning : The document was created with Spire.PDF for Python.
904 ❒ ISSN: 2502-4752 (FFN). Computation time [10], [12] for these netw orks are not critical as training is of ine b ut the forecast using trained netw ork is f ast. 3. DESCRIPTION OF D A T A AND PREPR OCESSING 3.1. Dataset Solar irradiance data can be obtained either from a measuring instrument installed at site or through satellite deri v ed irradiance dataset. Though satellite deri v ed dataset is less accurate compared to a dataset col- lected from a measuring instrument, satellite deri v ed dataset is often used by researchers [24], [25] because of its open access, ease of use, wide temporal and spatial co v erage and almost no data is missed. The data set containing real-w orld meterological v alues are collected from the National Rene w able Ener gy Laboratory’ s (NREL), National Solar Radiation Database (NSRDB) [26] for Ne w Delhi, India. Hourly data of global hori- zontal irradiance (GHI), temperature, pressure, relati v e humidity , wind direction and wind speed are obtained from the year 2009 to 2015. Solar i rradiance e xists only during daytime and hence the hours between 7:00 AM and 5:00 PM are considered. After analysing the dataset, solar irradiance peaks in the month of April and May comparati v ely for selected location and this sho ws its seasonal beha viour . Figure 4. Heat map with correlation coef cient between input v ariables 3.2. Data normalization The datas loaded into neural netw ork are normalized in the range of [0, 1]. According to (11) d i is the data before normalization, d ∗ i is the data after normalization, d min and d max are the minimum and maximum v alue of the v ariable. The aim of data normalization is to con v ert the numeric v alues in dataset to a common v alue. d ∗ i = d i − d min d max − d min (11) 3.3. Corr elation Linear relationship between tw o v ariables are measured commonly with the Pearson’ s correlation coef cient. Correlation between solar irradiance and other weather v ariables are measured using Pearson’ s Indonesian J Elec Eng & Comp Sci, V ol. 25, No. 2, February 2022: 900–909 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 905 correlation coef cient as sho wn in Figure 4. From the analyses of Pearson’ s correlation coef cient, temperature is found to be positi v ely correlated with GHI and relati v e humidity is found to be ne g ati v ely correlated with GHI. In literature, Ev ans [27] and Denes et al. [28] classied the absolute v alue for correlation f actor as v ery weak if v alue is between 0 and 0.19, weak if v alue is between 0.20 and 0.39, moderate if v al ue is between 0.40 and 0.59, strong if v alue is between 0.6 and 0.79 and v ery strong if v alue is between 0.8 and 0.99. As per the abo v e classication wind direction and wind speed can be ne glected as their correlation is v ery weak with GHI. Sliding windo w technique is used in preprocessing of data. 4. EXPERIMENTS AND RESUL TS 4.1. T raining and testing data The data from January 2009 to December 2013 are tak en as training set and the data from 2014 are tak en as test set. T raining and v alidation data are split using test train splitter which de v otes 80% of data to train and remaining 20% of data for v alidation. Input data with weather v ariables are in dif ferent range of v alues. Hence datasets are rescaled to lie in the range of [0, 1] and it is called normalization of datasets. Datasets are normalized using MinMaxScaler in scikit-learn according to (11). 4.2. Metrics Standard statistical measures such as root mean square error (RMSE) , and mean absolute error (MAE) are commonly used to measure the accurac y of forecast model [29], R M S E = v u u t 1 n n X i =1 ( Y pr ed,i − Y actual ,i ) 2 (12) M AE = 1 n i = n X i =1 | Y pr ed,i − Y actual ,i | (13) where Y pr ed is the predicted irradiance v alue and Y actual is the actual irradiance v alue. T o ha v e a good com- parision, forecast skill (FS) [18] is one of the most recommended metric in the w orld of forecast, where SP in 14 is smart persistence. F or ecastS k il l = 1 − R M S E pr oposed R M S E S P (14) 4.3. Experiments Experiments described here uses K eras v ersion 2.3.1 to implement BPNN, RNN, LSTM and LSTM based encoder -decoder sequence-to-sequence model with attention. Hyper -parameter for the abo v e models are tuned based on grid-search method. BPNN has 55 units and 95 units in hidden layer1 and hidden layer2 respecti v ely whereas RNN has 95 units and 105 units, LSTM has 85 units and 125 units in their repecti v e hidden layer1 and hidden layer2. Encoder and decoder layer has each 95 units in Enc-Dec-LSTM netw ork. Dropout of 0.2 is used in each of input layers as a re gularisation technique. Adam optimiser is used for optimization as it combines the best features of RMSprop and AdaGrad and batch size is set as 100 from grid-search method. Smart persistence model is free of training and tuning of parameters. 4.3.1. F or ecast r esults F orecast is performed with temperature, relati v e humidity and pressure as input v ariables and Enc- Dec-LSTM model is compared with LSTM, RNN, BPNN and SP . Clearsk y GHI is used in smart persistence to forecast day-ahead irradiance. The hourly input v ariables from 8:00 am to 5:00 pm are considered and therefore for a day , 10 timesteps are accounted. Dif ferent lagging time from 10 hour to 22 hour are tested and found the model results with least error for a 10 hour lagging time. Thus for day-ahead forecast, pre vious day’ s 10 hours of data is gi v en as input to predict ne xt day’ s 10 hours of solar irradiance with a resolution of one hour . Day-ahead solar irr adiance for ecast using sequence-to-sequence model with ... (Sowkarthika Subr amanian) Evaluation Warning : The document was created with Spire.PDF for Python.
906 ❒ ISSN: 2502-4752 -1 0 0 0 100 200 300 400 500 600 700 8 10 12 14 16 8 10 12 14 16 8 10 12 14 16 8 10 12 14 16 8 10 12 14 16 GH I   ( W / m ^ 2 ) Ti m e   i n   h o u r s Ac t u a l   D a t a RN N LS T M En c - D e c Su n n y   D a y   - 1  Ja n   2014 Su n n y   D a y   - 2  Ja n   2 0 1 4 Cl o u d y   D a y   - 3  Ja n   2 0 1 4 Su n n y   D a y   - 4  Ja n   2 0 1 4 Su n n y   D a y   - 5  Ja n   2014 Figure 5. Comparision of hourly forecasted irradiance v alues T able 1. Performance comparision of dif ferent algorithms Algorithm RMSE ( W /m 2 ) MAE ( W /m 2 ) FS(%) Enc-Dec-LSTM 100.57 60.27 31.07 LSTM 104.52 61.88 28.37 RNN 109.95 64.37 24.64 BPNN 124.67 104.13 14.56 SP 145.91 79.77 0 F orecast results in terms of error metric are sho wn in T able 1. Enc-Dec-LSTM outperforms the other models and compared to SP , BPNN and RNN, RMSE is reduced by 31.1%, 19.3%, 8.5% respecti v ely and MAE is reduced by 24.4%, 42.1%, 6.4% respecti v ely . Less forecast skill indicates that the models performance is almost same as that of smart persistence model. Enc-Dec- LSTM model has the highest forecast skill of 31.07% which indicates that the model performs better than an y other model compared here. As sho wn in Figure 5, Enc-Dec-LSTM model’ s forecast is nearer to actual data e v en on a cloudy day and hence its o v erall error is less compared to the other models. A v erage monthly RMSE of the test dataset is sho wn in T able 2 and its seen that the error peaks during Monsoon season. As sho wn in Figure 4, GHI is highly correlated with temperature v ariable and thus the monthly correlation of temperature with GHI is tested on the test dataset. The correlation of temperature with GHI during the months July , August and September are lo w which results wit h highest error during Monsoon season. T able 2. A v erage monthly RMSE ( W /m 2 ) and MAE ( W /m 2 ) of test dataset Error Algorithm Jan Feb Mar Apr May Jun Jul Aug Sept Oct No v Dec RMSE Enc-Dec 91.41 127.82 121.57 81.42 96.69 105.02 131.45 143.63 102.66 48.68 24.83 57.46 -LSTM LSTM 98.95 137.23 124.82 80.61 99.17 104.74 136.11 145.98 106.86 54.82 24.06 68.41 RNN 123.18 155.47 130.17 82.13 98.71 102.13 138.43 153.29 108.72 51.37 24.63 65.19 BPNN 98.76 133.35 136.35 114.75 121.88 114.53 137.24 153.98 123.97 73.09 49.03 64.87 MAE Enc-Dec 59.52 86.20 79.86 50.87 56.02 58.64 88.21 95.95 63.06 30.32 19.23 34.70 -LSTM LSTM 60.85 85.63 76.18 47.38 59.34 64.02 93.31 96.05 67.22 36.06 17.89 37.76 RNN 69.90 92.07 80.70 52.60 61.95 62.73 93.05 101.06 70.23 32.22 18.21 36.97 BPNN 71.29 103.27 106.79 99.72 105.91 93.45 108.85 122.92 97.72 59.49 39.92 47.15 Indonesian J Elec Eng & Comp Sci, V ol. 25, No. 2, February 2022: 900–909 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 907 4.3.2. F or ecast r esults at differ ent location In addition, the geographical location and climatic conditions also determine the forecast accurac y and hence a test is made on three dif ferent location with dif ferent climatic conditions to study and compare the feasibility of Enc-Dec-LSTM model. LSTM and Enc-Dec-LSTM models are compared for the datasets collected from NSRDB at dif ferent locations for dif ferent climatic conditions according to K oppen-Geiger climate classication. The data from year 2009 to 2013 are set as training dataset and 2014 as testing dataset. T able 3 lists the day-ahead RMSE of LSTM and Enc-Dec-LSTM in which Enc-Dec-LSTM has least error in all dif ferent locations with dif ferent climatic conditions. According to K oppen-Geiger climate classication Csa, Bsh and A w as listed in T able 3 denotes hot-summer mediterranean climate, hot semi-arid (steppe) climate, tropical sa v anna wet climate respecti v ely . T able 3. Day-ahead RMSE of forecast model at dif ferent locations Latitude Longitude Climate LSTM RMSE ( W /m 2 ) Enc-Dec-LSTM RMSE ( W /m 2 ) 23.25 77.35 Csa 101.42 98.47 26.25 73.05 Bsh 88.44 85.5 22.65 88.45 A w 121.46 117.55 4.3.3. Comparision with r ecently published papers A comparision of recently published w orks in one day-ahead solar irradiance forecast is made in T able 4. Emer ging deep learning techniques sho ws grea t impro v ement in accurac y for day-ahead solar irradiance forecast. F orecast er ror can also depend on geographical location and climatic condition and therefore forecast skill as de v eloped by Y ang [18] can be the best reference to compare with other models. As per forecast skill comparision in T able 4, LSTM based encoder -decoder sequence-to-sequence with att ention mechanism has highest skill of 31.07% and thus it outperforms the other models. T able 4. Comparision of day-ahead solar irradiance with recently published w orks Author Algorithm Location RMSE FS(%) Larson et al. [30] LSO a and NWP b San Die go, USA 27.5 % 24 Aryaputera et al. [31] WRF c and ETS d Station 500, Sing apore 188 ( W /m 2 ) 12.9 as per 14 Hai et al. [32] DFT e Qingdao, China 127.3 ( W /m 2 ) 6.3 Qing and Niu et al. [10] LSTM Cape V erde, Santiago 122.72 ( W /m 2 ) – Gao et al. [12] GR U Den v er , USA 122.45 ( W /m 2 ) 28.4 Present w ork Enc-Dec-LSTM Ne w Delhi, India 100.57 ( W /m 2 ) 31.07 Abbre viations: a least-squares optimization, b numerical weather prediction, c weather research forecasting, d e xponential smoothing, e discrete fourier transform 5. CONCLUSION AND FUTURE W ORK This paper attempts to study the encoder -decoder sequence-to-sequence models with attention for solar irradiance forecast which w as originally de v eloped for natural language processing. Initially datas are collected from NSRDB site and processed with sliding windo w technique and then normalised before applying to deep-learning model s to impro v e accurac y . Unw anted features of data are remo v ed usi n g pearson’ s corre- lation method. Fi v e years of data are supplied for training and one year of data is supplied for testing. Based on the e xperimental results, LSTM based encoder -decoder sequence-to-sequence models with attention mech- anism outperforms the other techniques as it combines both encoder -decoder f acility and attention mechanism which reduces error and impro v es accurac y , though the computation time of Enc-Dec-LSTM model is higher than LSTM. Further the rec ently de v eloped CNN based h ybrid models and transformer models could also be studied for solar irradiance forecast. REFERENCES [1] P . Mathiesen and J. Kleissl, “Ev aluation of numerical weather prediction for intra-day solar forecasting in the continental United States, ” Solar Ener gy , v ol. 85, no. 5, pp. 967–977, May 2011, doi: 10.1016/j.solener .2011.02.013. [2] S. Dutta et al. , “Load and rene w able ener gy forecasting for a microgrid using persistence technique, ” Ener gy Procedia , v ol. 143, pp. 617–622, Dec. 2017, doi: 10.1016/j.e gypro.2017.12.736. [3] S. Pelland, G. Galanis, and G. Kallos, “Solar and photo v oltaic forecasting through post-processing of the gl obal en vironmental multiscale numerical weather prediction model, ” Progress in Photo v oltaics: Research and Applications , v ol. 21, no. 3, pp. 284–296, No v . 2011, doi: 10.1002/pip.1180. Day-ahead solar irr adiance for ecast using sequence-to-sequence model with ... (Sowkarthika Subr amanian) Evaluation Warning : The document was created with Spire.PDF for Python.
908 ❒ ISSN: 2502-4752 [4] A. Mellit and AM. P a v an, ”A 24-h forecast of solar irradiance using articial neural netw or k: Application for performance prediction of a grid-connected PV plant at T rieste, Italy , ” Solar Ener gy , v ol. 84, no. 5, pp. 807–821, May . 2010, doi: 10.1016/j.solener .2010.02.006. [5] H. A. Mohammad, A. Jehad, and Y . Ola, “Solar photo v oltaic po wer forecasting in Jordan using articial neural net- w orks, ” International Journal of Electrical and Computer Engineering (IJECE) , v ol. 8, no. 1, pp. 497–504, Feb . 2018, doi: 10.11591/ijece.v8i1.pp497-504. [6] R . Meenal, P . A. Michael, D. P amela, and E. Rajasekaran, “W eather prediction using random forest machine learning model, ” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS) , v ol. 22, no. 2, pp. 1208–1215, May 2021, doi: 10.11591/ijeecs.v22.i2.pp1208-1215. [7] Y . Amellas, O. E. Bakkali, A. Djebil, and A. Echchelh, “Short-term wi nd speed predi ction based on MLP and N ARX netw ork models, ” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS) , v ol. 18, no. 1, pp. 150–157, April 2020, doi: 10.11591/ijeecs.v18.i1.pp150-157. [8] F . V . Gutierrez-Corea, M. A. Manso-Callejo, M. P . Moreno-Re gidor , and M. T . Manrique-Sanc ho, “F orecasting short-term solar irradiance based on articial neural netw orks and data from neighboring meteorological stations, ” Solar Ener gy , v ol. 134, pp. 119– 131, Sept. 2016, doi: 10.1016/j.solener .2016.04.020. [9] N. Susithra, G. Santhanamari, M. Deepa, P . Reba, K. C. Ramya, and L. Gar g, “Deep learning-based acti vity monitoring for smart en vironment using radar , ” In: Challenges and Solutions for Sustainable Smart City De v elopment. EAI Springer Inno v ations in Communication and Computing. Springer , Cham. , May 2021, doi: 10.1007/978-3-030-70183-3 5. [10] X. Qing and Y . Niu, “Hourly day-ahead solar irradiance prediction using weather forecasts by LSTM, ” Ener gy , v ol. 148, pp. 461– 468, April 2018, doi: 10.1016/j.ener gy .2018.01.177. [11] B. K. Jeon and E. J. Kim, “Ne xt-Day Prediction of Hourly Solar Irradiance Using Local W eather F orecasts and LSTM T rained with Non-Local Data, ” Ener gies , v ol. 13, no. 20, Oct. 2020, doi: 10.3390/en13205258. [12] B . Gao, X. Huang, J. Shi, Y . T ai, and X. Rui, “Predicting day-ahead solar irradiance through g ated recurrent unit using weather forecasting data, ” Journal of Rene w able and Sustainable Ener gy , v ol. 11, no. 4, Aug. 2019, doi: 10.1063/1.5110223. [13] M . Alhussein, S. I. Haider , and, K. Aurangzeb, “Microgrid-le v el ener gy management approach based on short-term forecasting of wind speed and solar irradiance, ” Ener gies , v ol. 12, no. 8, April 2019, doi: 10.3390/en12081487. [14] A . Sharma and A. Kakkar , “F orecasting daily global solar irradiance generation using machine learning, ” Rene w able and Sustainable Ener gy Re vie ws , v ol. 82, pp. 2254–2269, Feb . 2018, doi: 10.1016/j.rser .2017.08.066. [15] J . Zhang et al. , “Baseline and tar get v alues for re gional and point PV po wer forecasts: T o w ard impro v ed solar forecasting, ” Solar Ener gy , v ol. 122, pp. 804–819, 2015, doi: 10.1016/j.solener .2015.09.047. [16] G. Ganesh, G. V . K umar , A. R. V ijayBab u, G. S. Rao, and Y . R. T agore, “Performance Analysis and MPPT Control of a Standalone Hybrid Po wer Generation System”, Journal of Electrical Engineering , v ol. 15, no. 1, pp. 334–343, June 2015. [17] C. V o yant, T . Soubdhan, P . Lauret, M. Da vid, and M. Muselli, “Statistical parameters as a means to a priori assess the accurac y of solar forecasting models, ” Ener gy , v ol. 90, pp. 671–679, Aug. 2015, doi: 10.1016/j.ener gy .2015.07.089. [18] D. Y ang, “ A guideline to solar forecasting research practice: Reproducible, operational, probabilistic or ph ysically-based, ensemble, and skill (R OPES), ” Journal of Rene w able and Sustainable Ener gy , v ol. 11, no. 2, April 2019, doi: 10.1063/1.5087462. [19] A. T ealab, “T ime seri es forecasting using articial neural netw orks methodologies: A systematic re vie w , ” Future Computing and Informatics Journal , v ol. 3, no. 2, pp. 334–340, Dec. 2018, doi: 10.1016/j.fcij.2018.10.003. [20] S. Hochrei ter and J. Schmidhuber , “Long short-term memory , ” Neural Computation , v ol. 9, no. 8, pp. 1735–1780, No v . 1997, doi: 10.1162/neco.1997.9.8.1735. [21] S. Du, T . Li, Y . Y ang, and S. J. Horng, “Multi v ariate time series forecasting via attention-based enc oder–decoder frame w ork, ” Neurocomputing , v ol. 388, pp. 269–279, May 2020, doi: 10.1016/j.neucom.2019.12.118. [22] I. Sutsk e v er , O. V in yals, and Q. V . Le, “Sequence to Sequence Learning with Neural Netw orks, ” arXi v preprint , Dec. 2014, arxi v:1409.3215v3. [23] M. Luong, H. Pham, and C. D. Manning, “Ef fecti v e Approaches to Attention-based Neural Machine T ranslation, ” arXi v preprint , Sept. 2015, arXi v:1508.04025v5. [24] D. Y ang, “ A correct v alidation of the National Solar Radiation Data Base (NSRDB), ” Rene w able and Sustainable Ener gy Re vie ws , v ol. 97, pp. 152–155, Dec. 2018, doi: 10.1016/j.rser .2018.08.023. [25] G. M. Y agli, D. Y ang, and D. Srini v asan, “ Automatic hourly solar forecasting using machine lea rning models, ” Rene w able and Sustainable Ener gy Re vie ws , v ol. 105, pp. 487–498, May 2019, doi: 10.1016/j.rser .2019.02.006. [26] NREL. [online]. A v ailable: https://maps.nrel.go v/nsrdb-vie wer [accessed December 2020] [27] J. D. Ev ans, “Straightforw ard statistics for the beha vioral sciences”, Thomson Brooks/Cole Publishing Compan y , 1996. [28] K. A. Denes, S. M. Yhan, P . C. Bernardes, and C. A. Conte-Junior , “Relationship between CO VID-19 and weather: Case study in a tropical country , ” International Journal of Hygiene and En vironmental Health , v ol. 229, Aug. 2020, doi: 10.1016/j.ijheh.2020.113587. [29] R. J. Hyndman and A. B. K oehler , “ Another look at m easures of forecast accurac y , ” International Journal of F orecasting , v ol. 22, no. 4, pp. 679–688, Dec. 2006, doi: 10.1016/j.ijforecast.2006.03.001. [30] D. P . Larson, L. Nonnenmacher , and C. F . M. Coimbra, “Day-ahead forecasting of solar po wer output from photo v oltaic plants in the American Southwest, ” Rene w able Ener gy , v ol. 91, pp. 11–20, June 2016, doi: 10.1016/j.renene.2016.01.039. [31] A. W . Aryaputera, D. Y ang, and W . M. W alsh, “Day-Ahead Solar Irradiance F orecasting in a T ropical En vironment, ” Journal of Solar Ener gy Engineering , v ol. 137, no. 5, July 2015, doi: 10.1115/1.4030231. [32] L. Hai, C. Zhang, Y . Y . Hong, Y . He, and S. W en, “Day-ahead spatiotemporal solar irradiation forecasting using frequenc y- based h ybrid principal component analysis and neural netw ork, ” Applied Ener gy , v ol. 247, pp. 389–402, Aug. 2019, doi: 10.1016/j.apener gy .2019.04.056. Indonesian J Elec Eng & Comp Sci, V ol. 25, No. 2, February 2022: 900–909 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 909 BIOGRAPHIES OF A UTHORS So wkarthika Subramanian recei v ed the B.E de gree in electrical engineering from K u- maraguru Colle ge of T echnology and M.E de gree in po wer electronics and dri v es in PSG Colle ge of T echnology . She is currently w orking a s assistant professor in Go v ernment Colle ge of T echnology . Her research interests include AI based po wer electronics applications to rene w able ener gy systems. She can be contacted at email: so wkarthika@gct.ac.in. Y asoda Kailasa Gounder recei v ed the B.E. in Electrical Engineering at Coimbatore Insti- tute of T ec hnology , Coimbatore, India, M.E. de gree in Po wer Electronics and Dri v es from Alag appa Chettiar Colle ge Engineering and T echnology , Kar aikudi, India and Ph.D. from Anna Uni v ersity , Chennai, India. Currently , she is w orking as an assistant professor (senior grade ) in Department of Electrical Engineering at Go v ernment Colle ge of T echnology , Coimbatore, India. Her research inter - ests are wind ener gy con v ersion systems, po wer electronics and micro grids. She can be contacted at email: yasoda@gct.ac.in. Sumathi Linganathan recei v ed the B.T ech de gree in Information T echnology from VLB Jannakiammal Colle ge of Engineering and T echnology and M.E. in Computer Science and Engineer - ing from K umaraguru Colle ge of T echnology . She w ork ed as softw are engineer in Infosys Pvt. Ltd. from 2007 to 2008. She is currently w orking as assistant professor in Go v ernment Colle ge of T ech- nology . Her research interests include Internet of Things, Machine Learning, and Cyber Security . She can be contacted at email: lsumathi@gct.ac.in. Day-ahead solar irr adiance for ecast using sequence-to-sequence model with ... (Sowkarthika Subr amanian) Evaluation Warning : The document was created with Spire.PDF for Python.