Professional Documents
Culture Documents
Abstract—This paper proposes a short-term forecast of the management. The works developed in the energy management
solar irradiation on the Gandon site in Senegal, based on field were focused on control command [4], optimization [5],
the multi-layer artificial neural networks, with the Levenberg- load forecast [6] and available forecast resources [7], [8]. The
Marquardt algorithm. The meteorological data are used as
input variables. Due to the random nature of the meteorological present work belongs to the last category.
variables, this forecast is necessary in order to plan the energy
resources on the short-term. The forecast horizon is fixed at ten The forecast resources can be developed on the short-term
minutes. A data analysis with the WEKA (Waikato Environment [7] and on the long-term [9]. Works [8] classified the eleven
for Knowledge Analysis) software permitted to select three input most used forecast methods in decreasing order: artificial
parameters (maximum temperature, mean radiation and time)
among the nine measured and to determine four hidden layers. neural network (ANN), support vector machine (SVM),
A comparative study is realized between the measured solar autoregressive integrated moving average (ARIMA), fuzzy
radiation on the site and the forecast results. The forecasted and logic (FL), linear regression (LR), genetic algorithm (GA),
measured data are correlated at 99% with a root mean square particle swarm optimisation (PSO), grey prediction (GM)
error (RMSE) of 0.0362, a time-statistical error of 9.25 and a and autoregressive moving average (ARMA). Moreover, the
normalized root mean square error (nRMSE) of 19.37%. These
results show the efficiency of this model and the relevance of the Kalman filter [10], interpolation methods [11], Thevenin
chosen approach. Equivalent Circuit Model [12] , Mann-Kendall Test [13] and
Index Terms—Forecast, solar irradiation, artificial neural net- Amstrong’s model [14] are used. In our work, the artificial
work, Levenberg-Marquardt algorithm, WEKA . neural network is used for the mean solar radiation forecast
on a horizontal surface.
I. I NTRODUCTION
Renewable energies has presented a growing interest for The artificial neural network is based on artificial
the past decade [1]. Several works examined the following intelligence. The learning process depends on the data quality
aspects: potential evaluation [2], system sizing [3] and energy and quantity that is used [15]. Some authors [7], [16], [17]
,&5(5$
,(((
WK,QWHUQDWLRQDO&RQIHUHQFHRQ5HQHZDEOH(QHUJ\5HVHDUFKDQG$SSOLFDWLRQV 3DULV)5$1&(2FW
,&5(5$
WK,QWHUQDWLRQDO&RQIHUHQFHRQ5HQHZDEOH(QHUJ\5HVHDUFKDQG$SSOLFDWLRQV 3DULV)5$1&(2FW
TABLE I
DATA MEASURED OVER 06 HOURS
h T Tm H Hm E Em P
(minute) (o C) (o C) (%) (%) (W/m2 ) (W/m2 ) (Pa )
460 18.49 18.65 72.6 73.4 4.327 7.02 101.2
470 18.44 18.52 72.2 73 14.27 21.05 101.2
480 18.31 18.52 73.3 73.6 22.92 28.07 101.2
490 18.13 18.25 74.1 74.5 42.8 49.12 101.2
500 18.17 18.31 74.1 74.3 62.34 84.2 101.2
510 18.33 18.58 74.6 75.3 90.6 105.3 101.2
520 18.79 19.05 73.2 74.5 102.1 133.3 101.2
530 19.17 19.39 71.7 72.4 125.5 140.3 101.2
540 19.02 19.32 72.6 73.8 166.1 189.5 101.2
550 19.23 19.58 73.2 74.2 195.5 210.5 101.2
560 19.92 20.19 70.1 71.7 241.8 280.7 101.3
570 20.21 20.39 68.81 69.69 302.1 322.8 101.3
580 20 20.12 69.55 70.2 347.7 371.9 101.3
590 20.48 21.06 67.56 69.36 394.2 414 101.3
600 21.41 21.73 63.57 65.14 416.3 435 101.3
610 21.49 21.73 62.86 63.74 430.5 463.1 101.3
620 21.32 21.52 63.64 64.07 476.6 512.2 101.3
630 21.86 22.46 62.1 63.8 525.3 547.3 242.9
640 22.86 23.13 58.27 59.51 548.2 561.3 1059 Fig. 2. methodology
650 23.35 23.53 56.66 57.7 546 582.3 1059
660 23.31 23.53 57.06 57.97 595.5 624.5 1059
670 23.31 23.66 57.25 57.91 634.5 645.5 1059
680 24.04 24.53 55.05 56.97 638.4 652.5 1059 B. Normalization and denormalization
690 25.11 25.6 51.28 53.35 628.5 694.6 1059
700 25.67 25.8 49.05 49.74 693.3 716 1059
710 25.43 25.6 49.56 50.01 716 730 1059 The data normalization consists of bringing the data to the
720 25.35 25.54 49.79 51.14 732 744 1059 same order of magnitude. It avoids undermining the ANN
730 25.19 25.4 54.08 55.36 744 751 1059 training process [17]. To provide a good data convergence
740 24.85 25 55.84 56.5 760 772 1059
750 25.16 25.33 54.52 56.03 752 772 1059 with the ANN, several works propose a data normalization
760 25.05 25.26 54.85 55.49 760 800 1059 [17], [32]. The normalization equation is given by (1).
770 25.25 25.6 54.51 55.23 796 800 1059 x represents the measured variable
780 25.66 26.07 53.38 53.82 813 828 1059 y is the normalized value of the measured variable
790 26.43 26.74 50.99 52.75 673.6 807 1059
800 26.5 26.67 50.46 51.07 739 800 1059
xmax is the maximum value of measured variable
810 26.27 26.4 50.77 51.14 723 765 1059 xmin is the minimum value of measured variable
820 26.09 26.21 51.03 51.54 771 786 1059 ymax is the maximum value of the normalized value
ymin is the minimum value of the normalized value
the ANN training process, and 30% for the test. One example y = (ymax − ymin )(x − xmin )/(xmax − xmin ) + ymin (1)
over six hours of these parameters on an unfavorable day
(January 15th , 2015) is illustrated in Tab.I. The day number
After the test process of the ANN, the data is denormalized
(Dn) is counted from the first day of the year in which data
using (2). Denormalization consists of bringing back the
measurements started.
reduced data to the real values .
III. M ETHODOLOGY
The methodology used (Fig.2)is divided into several steps: x = (xmax − xmin )(y − ymin )/(ymax − ymin ) + xmin (2)
variable of interest and number of layer selection for the ANN,
variable normalization and denormalization, performance eval-
IV. A RTIFICIAL NEURAL NETWORK
uation and mean radiation forecast results.
A. Variable of interest and number of layers selection for the A. Artificial neural network model
artificial neural network
On the nine collected parameters of the Gandon site, a The ANN type used is the feed-forward multilayer
preliminary study permitted the seclection of decision-making perception neural network, based on the Levenberg-Marquardt
variables. For this purpose, the correlation between each algorithm. The multilayer neural network architecture, with
variable measured at t time and the radiation value measured n input values, a hidden layer with Ni artificial neuron and
at t + ΔT time is realized using the WEKA software. The single neuron in the output layer is presented in Fig.3.
variables having a strong correlation on the forecast and easy The mathematical expression of the neural
to measure are taken as input variables for the ANN. network output o is given by (3) and (4).
,&5(5$
WK,QWHUQDWLRQDO&RQIHUHQFHRQ5HQHZDEOH(QHUJ\5HVHDUFKDQG$SSOLFDWLRQV 3DULV)5$1&(2FW
Ni represents the number of artificial neurons Fig. 4. Correlation obtained from WEKA between forecasted and measured
Wj,N i is the weight associated with each connection data on Gandon site.
L1, L2 indicate which layer the weight belongs to
f, g indicate activation functions
o is the multilayer perceptron output
x1 , xj , .., xn are input data
y1 is the neuron 1 output
yj is the neuron Ni output
⎛ ⎞ Fig. 5. Layer number arhitecture
n
L1
yi = g⎝ (xj ) ∗ Wj,N i ⎠ (3)
j=1
⎛ ⎞ RM SE = (xi − x̂i )2 /N (9)
n
L2 i
o= f⎝ (yi ) ∗ Wj,N i ⎠ (4)
j=1
nRM SE = RM SE/E(x) (10)
B. Levenberg-Marquardt algorithm M BE = (1/N ) ∗ (x̂i − xi ) (11)
i
The Levenberg-Marquardt algorithm allows the system to
perform a learning process by minimizing the error. It is t − statistic = (N − 1)M BE 2 / (RM SE 2 − M BE 2 )
described by (7) and (5) [33]. The algorithm parameters b (12)
and Wj,N i are chosen using (6).
V. R ESULTS AND DISCUSSION
Snew = S − J T J + μI J T e (5) ymax and ymin values are chosen respectively as one and
ST = [Wj,N i ...b] (6) zero.
J= [∂e/∂Wj,N i ...∂e/∂b] (7) A. Data and number of layer selection
I is the identity matrix The correlation results between the meteorological variables
e is the error vector and the mean solar radiation are presented in Fig.4. The
J is the Jacobian matrix variables of maximum temperature, time and mean radiation,
Wj,N i indicates the weight being of easy access contribute to more than 99% (correlation
b is the bias of 0.99) of the forecast on the Gandon site (refer to Fig.4).
ST is the transpose value of vector S Therefore, they are chosen as the forecast input for the rest
μ is a scalar with initial value 0.001 of our work. The number of hidden layers retained is 4. The
the expressions of J, S and e are detailed in [33] obtained model is shown on Fig.5.
C. statistical criteria B. Artificial neural network
The statistical criteria are evaluated using (8) [19]; (9),(10) 1) Training process: Fig.6 presents the training results
[17]; (11),(12) [19]. obtained thanks to the 70% of the used data. It presents
xi is the i − th measured value the evolution of mean solar radiation as a function of the
x̂i is the i − th forecasted value input data. We can observe the correlation between training,
E(x) is the mean of test values validation and test phases. The validation indicates the end
N is the total number of test values of the training process. The training process performance
analysis is presented in Fig.7. It shows that the gradient
R2 = 1 − (xi − x̂i )2 / (xi )2 (8)
decreases according to the evolution of the epoch number
i i
,&5(5$
WK,QWHUQDWLRQDO&RQIHUHQFHRQ5HQHZDEOH(QHUJ\5HVHDUFKDQG$SSOLFDWLRQV 3DULV)5$1&(2FW
C. Performance indicators
The performance indicators are presented in Tab.II. It shows
the low values of RMSE, nRMSE, t-statistic and MBE.
600
Fig.8. It is obtained at the epoch 178.
2) Test: The test phase is the one following the training 400
,&5(5$
WK,QWHUQDWLRQDO&RQIHUHQFHRQ5HQHZDEOH(QHUJ\5HVHDUFKDQG$SSOLFDWLRQV 3DULV)5$1&(2FW
three have been used. As principal contribution, this paper [15] W. Cao, X. Wang, Z. Ming, and J. Gao, “A review on neural networks
presented: with random weights,” Neurocomputing, vol. 275, pp. 278–287, Jan.
2018.
• the different necessary meteorological variables, as well [16] M. Q. Raza, N. Mithulananthan, and A. Summerfield, “Solar output
as their correlation for a better forecast of mean solar power forecast using an ensemble framework with neural predictors and
Bayesian adaptive combination,” Solar Energy, vol. 166, pp. 226–241,
radiation on the Gandon site; May 2018.
• the future mean solar radiation forecast over a time period [17] E. F. Alsina, M. Bortolini, M. Gamberi, and A. Regattieri, “Artificial
of 10 minutes using maximum temperature, time and neural network optimisation for monthly average daily global solar
radiation prediction,” Energy Conversion and Management, vol. 120,
mean solar radiation recorded on the past time periods. pp. 320–329, July 2016.
This model is a tool for decision-making concerning the solar [18] A. K. Yadav, H. Malik, and S. S. Chandel, “Selection of most relevant
photovoltaic power stations production forecast. Indeed, the input parameters using WEKA for artificial neural network based
solar radiation prediction models,” Renewable and Sustainable Energy
power station operators, who actually represent 20.3% of the Reviews, vol. 31, pp. 509–519, Mar. 2014.
total production, need a reliable production estimation at any [19] E. Mboumboue, D. Njomo, M. L. Ndiaye, P. A. N’diaye, M. F.
time. This tool can provide a contribution to this sort of need Ndiaye, and A. K. Tossa, “On the applicability of several conventional
regression models for the estimation of solar global radiation component
. in Cameroon and Senegal sub-Saharan tropical regions,” Journal of
In perspectives, we intend to: Renewable and Sustainable Energy, vol. 8, p. 025906, Mar. 2016.
[20] B. Amrouche and X. Le Pivert, “Artificial neural network based daily
• proceed the forecast overall national territory starting
local forecasting for global solar radiation,” Applied Energy, vol. 130,
from the geographical coordinates; pp. 333–341, Oct. 2014.
• investigate the correlation between different interval mea- [21] L. Mazorra Aguiar, B. Pereira, M. David, F. Díaz, and P. Lauret, “Use
of satellite data to improve solar radiation forecasting with Bayesian
surement times Δt on the mean solar radiation forecast Artificial Neural Networks,” Solar Energy, vol. 122, pp. 1309–1324,
at the given t time in the future. Dec. 2015.
[22] M. Zaiani, D. Djafer, and F. Chouireb, “New Approach to Establish a
R EFERENCES Clear Sky Global Solar Irradiance Model,” p. 9, 2017.
[1] M. J. Burke and J. C. Stephens, “Political power and renewable energy [23] N. Kumar, S. P. Sharma, U. K. Sinha, and Y. Nayak, “Prediction of
futures: A critical review,” Energy Research & Social Science, vol. 35, Solar Energy Based on Intelligent ANN Modeling,” p. 6, 2016.
pp. 78–93, Jan. 2018. [24] V. A. Jimenez, A. Barrionuevo, A. Will, and S. Rodríguez, “Neural Net-
[2] S. Pishgar-Komleh and A. Akram, “Evaluation of wind energy potential work for Estimating Daily Global Solar Radiation Using Temperature,
for different turbine models based on the wind speed data of Zabol Humidity and Pressure as Unique Climatic Input Variables,” Smart Grid
region, Iran,” Sustainable Energy Technologies and Assessments, vol. 22, and Renewable Energy, vol. 07, no. 03, pp. 94–103, 2016.
pp. 34–40, Aug. 2017. [25] K. Moustris, K. Kavvadias, A. Kokkosis, and A. Paliatsos, “One day-
[3] A. Thornton, S.-J. Kim, and S. Kara, “Sizing a Hybrid Renewable ahead forecasting of mean hourly global solar irradiation for energy
Energy System to Reduce Energy Costs at Various Levels of Robustness management systems purposes using artificial neural network modeling,”
for an Industrial Site,” Procedia CIRP, vol. 69, pp. 371–376, 2018. in Mediterranean Conference on Power Generation, Transmission, Dis-
[4] M. F. Ndiaye, F. Guérin, D. Lefebvre, and P. A. Ndiaye, “Model tribution and Energy Conversion (MedPower 2016), (Belgrade, Serbia),
Predictive Control and Generalized Adaptative PID for Load Sharing pp. 104 (6 .)–104 (6 .), Institution of Engineering and Technology, 2016.
In Systems of Multiple Sources of Energy,” IJARCCE, vol. 4, pp. 307– [26] M. Brenna, F. Foiadelli, M. Longo, and D. Zaninelli, “Solar radiation and
312, Oct. 2015. load power consumption forecasting using neural network,” in 2017 6th
[5] A. B. Forough and R. Roshandel, “Lifetime optimization framework International Conference on Clean Electrical Power (ICCEP), (Santa
for a hybrid renewable energy system based on receding horizon Margherita Ligure, Italy), pp. 726–731, IEEE, June 2017.
optimization,” Energy, vol. 150, pp. 617–630, May 2018. [27] L. Ciabattoni, G. Ippoliti, S. Longhi, M. Cavalletti, and M. Rocchetti,
[6] A. Kaur, L. Nonnenmacher, and C. F. Coimbra, “Net load forecasting for “Solar irradiation forecasting using RBF networks for PV systems
high renewable energy penetration grids,” Energy, vol. 114, pp. 1073– with storage,” in 2012 IEEE International Conference on Industrial
1084, Nov. 2016. Technology, (Athens), pp. 699–704, IEEE, Mar. 2012.
[7] F.-V. Gutierrez-Corea, M.-A. Manso-Callejo, M.-P. Moreno-Regidor, and [28] M. Cococcioni, E. D’Andrea, and B. Lazzerini, “24-hour-ahead fore-
M.-T. Manrique-Sancho, “Forecasting short-term solar irradiance based casting of energy production in solar PV systems,” in 2011 11th In-
on artificial neural networks and data from neighboring meteorological ternational Conference on Intelligent Systems Design and Applications,
stations,” Solar Energy, vol. 134, pp. 119–131, Sept. 2016. (Cordoba, Spain), pp. 1276–1281, IEEE, Nov. 2011.
[8] K. B. Debnath and M. Mourshed, “Forecasting methods in energy [29] Shubham and K. Padmanabh, “Solar power prediction using deduced
planning models,” Renewable and Sustainable Energy Reviews, vol. 88, feature of visibility index and artificial neural network,” in 2017 Inter-
pp. 297–325, May 2018. national Conference on Advances in Computing, Communications and
[9] P.-C. Hsu, B.-J. Huang, P.-H. Wu, W.-H. Wu, M.-J. Lee, J.-F. Yeh, Y.-H. Informatics (ICACCI), (Udupi), pp. 97–102, IEEE, Sept. 2017.
Wang, J.-H. Tsai, K. Li, and K.-Y. Lee, “Long-term Energy Generation [30] M. F. NDIAYE, A. K. MBODJI, O. CISSE, E. MBOUMBOUE, M. L.
Efficiency of Solar PV System for Self-consumption,” Energy Procedia, NDIAYE, and P. A. NDIAYE, “Prévision à court terme du potentiel
vol. 141, pp. 91–95, Dec. 2017. éolien par la méthode de prédiction linéaire : application au site de
[10] T. Soubdhan, J. Ndong, H. Ould-Baba, and M.-T. Do, “A robust Gandon au Sénégal,” p. 8, 2016.
forecasting framework based on the Kalman filtering approach with a [31] M. F. Ndiaye, F. Guerin, D. Lefebvre, and P. A. Ndiaye, “Optimal load
twofold parameter tuning procedure: Application to solar and photo- sharing strategy in system of multiple source based on model control
voltaic prediction,” Solar Energy, vol. 131, pp. 246–259, June 2016. predictive,” in 29th European Simulation and Modelling Conference -
[11] C. Xiu, T. Wang, M. Tian, Y. Li, and Y. Cheng, “Short-term prediction ESM’2015, (Leicester, United Kingdom), pp. 237–244, ESM’2015, Oct.
method of wind speed series based on fractal interpolation,” Chaos, 2015.
Solitons & Fractals, vol. 68, pp. 89–97, Nov. 2014. [32] S. M. Mousavi, E. S. Mostafavi, and P. Jiao, “Next generation prediction
[12] M. A.-E.-H. Mohamed, “Solar Irradiance Estimation of Photovoltaic model for daily solar radiation on horizontal surface using a hybrid
Module based on Thevenin Equivalent Circuit Model,” p. 7, 2015. neural network and simulated annealing method,” Energy Conversion
[13] J. Singh and M. Kumar, “Solar Radiation over Four Cities of India: and Management, vol. 153, pp. 671–682, Dec. 2017.
Trend Analysis using Mann-Kendall Test,” p. 11, 2016. [33] H. T. Nguyen-Truong and H. M. Le, “An implementation of the Leven-
[14] A. David and N. R. Ngwa, “Global Solar Radiation of Some Regions berg–Marquardt algorithm for simultaneous-energy-gradient fitting using
of Cameroon using the linear Angstrom and non-linear Polynomial two-layer feed-forward neural networks,” Chemical Physics Letters,
Relations (Part I) Model Development,” p. 9, 2013. vol. 629, pp. 40–45, June 2015.
,&5(5$