Artigo SEPOPE - Redes Neurais - Ingles

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

Wind power forecasting with Neural Networks for Power


Smoothing using Energy Storage Systems

Pedro Rosas1, Gustavo Luna1, Luiz Henrique A. de Medeiros1, Pablo Tabosa1, Paola
Furlanetto2, Mariana Moraes2, Felipe Ramos2, Wendell Teixeira3, Renato Povia Silva3
1
Universidade Federal de Pernambuco – UFPE
2
Instituto Tecnológico Edson Mororó Moura - ITEMM
3
CPFL Energia
Brasil

SUMMARY

This paper presents a technique for prediction of wind generation for complex, circuits, parks and farms
that together the energy storage system can reduce fluctuations of wind power produced in the
interconnection with the power grid of transmission or distribution of energy. The methodology
employed is based on artificial intelligence with the use of neural network architecture NARX-type to
perform a prediction a step ahead of the instantaneous power generated in a wind farm. To calibrate this
neural network and perform the desired forecast time series is used the instant wind power generated.

KEYWORDS

Neural Network, NARX, Wind Generation, Energy Storage System

1. Introduction
The use of renewable sources for power generation has been growing in recent years, with an
emphasis on wind and solar energy, which are clean and inexhaustible sources [1]. Wind power has
been the one that’s been growing faster and it is estimated that by 2030 the total energy from wind
farms in the world to be 2182 TWh [2].
Despite the use of renewable sources bring benefits to the environment, the increase of its
participation in the electrical system can bring difficulties to the operation of the grid. The intermittent
nature of wind brings in power fluctuations that can, in specific situations, result in variations of
frequency and voltage in electrical grid.
In addition, the reduction in wind speed, as happens in moments of stillness, can result in a
reduction of the total power of the windfarm in a matter of minutes. The reduction of the power
generation can cause in a demand request from the grid that cannot be attained by the current power
production of the wind farms leading to long forays into low frequencies.
Due to such variations ensuring a good stability and reliability to the electrical system with high
concentration of wind power becomes a great challenge [3].

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

One of the ways to mitigate the effects caused by intermittence of renewable wind and solar is
the use of energy storage systems (ESS). Among all the ESS technologies available on the market,
battery-based system (BESS, Battery Energy Storage Systems) is the one that features higher growth,
especially with the opportunities that arise with the cost reduction of investments in Lithium batteries
with applications in electric cars [4].
For renewable energy, especially wind power, the BESS can act in various applications, the
noblest one being the power smoothing, wich basically smoothes the power variation on the wind farm
output caused by the wind intermittency [5].
Some control systems for power smoothing require the prediction of the power for the next
instant, which can be obtained throught several ways, one of them being through artificial neural
network (ANN). Thus, the main focus of this paper will be matters relating to neural network for
predicting wind power in a very short period.
It is already proven in the literature that the prediction made by neural networks presents better
results than other forecasting methods [6] [7]. Besides, neural networks are simpler, compared to other
forecasting methods, since it acts like a black box in which only receives data and presents results in its
output [8].
The case analysis presented here is based on a project of strategic R&D of ANEEL for storage
application.

2. The Problem

The control system of the BESS is divided into several components: dispatch system of BDM
(Battery Dispatch Module), BMS (Battery Monitoring System) and monitoring system PCS (Power
Conversion System).
The problematic involves in predicting the next power generated by the wind farm in order to
determine the power to be dispatched or absorbed by the storage system, wich in this case is the BESS.
In this article it is introduced the component that provide the predicted generation for the dispatch
system.
Thus, the problem addressed in this article is to develop a tool capable of predicting a step ahead
for the power produced in a wind farm.

3. Methodology and Development

An Artificial Neural network is a mathematical representation of human neurons and their


synapses. Just like human neurons, an RNA is able to learn based on the stimulus that receives. The
learning capacity of a RNA is one its most attractive point, because with that comes the possibility of
training it to solve complex problems or act in some applications, such as pattern recognition, signal
processing, control, complex functions approximation, time series analysis [9].
The activity of prediction and time series analysis is best performed by architectures that are
known as recurring, as these have better ability to interpret not statistically stationary signals, thereby,

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

being able to represent as result the dynamic behavior the signal [6]. Among all recurrent architectures,
NARX (Nonlinear Autoregressive Exogenous Model) is one of them.

3.1 NARX Architecture

NARX is a type of neural network architecture classified as recurrent network. In recurrent


networks (or feedback), the output of some neurons in the hidden layer is feedback to other neurons
that are present in the same layer, or in previous layers or even itself.
This way the network develops the ability of dynamic memory, which allows it to be able to
interpret and represent states for dynamic systems/signals. This characteristic is one of the most
interesting point of recurrent networks.
However, for NARX architecture, the feedback is very limited, because it has a single feedback
that comes from output layer neuron in which it is connected to the input layer neuron. However,
despite this limitation, it has already been proven that NARX network is just as computationally
powerful as any other fully feed back network [10].
Thus, NARX architecture can be considered simpler than others fully feedback recurrent
architectures but as computationally strong as them. Such qualities were very attractive in the decision
of what type of architecture should choose to perform the predictions proposed in this work.
Mathematically a NARX network is represented, as shown in Equation 1 [10].
(1)

Where u (n) ∈ ℜ and y (n) ∈ ℜ are functions that represent, respectively, the input and the output
of the network at a given discreet time step n, while the terms du and dy represent, repectially, the input
delay and the feedback delay. It’s important to mention that du>1 and dy>1. The function
f[•]corresponds to the mapping function of the RNA, wich it is the same used MLP (Mulitlaye
perceptron) network.
Considering Equation (1) and the connection of neurons assumed from the mapping function,
NARX architecture can be finally assembled and observed in Figure 1.

... ...

... ...

... ...

Figure 1 – NARX Network

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

The NARX Network presented can be configured in two ways: parallel or in series-parallel and
both settings can be assigned to Figure 2.

Input delay

Input delay
NARX NARX
NETWORK NETWORK
Feedback delay

Feedback delay
(a) (b)

Figure 2 - (a) NARX Network in parallel configuration, (b) NARX Network in series-parallel configuration
For the parallel configuration, the equation (1) can be represented as in equation (2) [6]:

(2)

Where the notation (^) means estimated/predicted value. In parallel configuration, the prediction
is obtained through the feedback from the output layer to the exegenous input (y(n)) present at the input
layer, as it can be seen in figure 2 (a). The feedback process continue until the network is able to reach
the desired output
Mathematically the series-parallel configuration is represented by the equation (3) [6].

(3)

The series-parallel configuration, does not use the estimated values obtained from output to make
the prediction in fact, the prediction is done by feeding the exogenous input with actual predicted
values.
In both configurations the nerwork is able to forecast, but the series-parallel is limited to forecast
one-step ahead only, while the the parallel configuration can predict n step ahead. However, it is
indicated that during the network training process to use the series-parallel configuration since the
NARX is provided with actual predicted values, this way achieving better predictions. After the
training step, the NARX’s loop can be closed in order to predict n step ahead. For this paper it was used
the series parallel configuration, since it was only necessary to know the next power generated.

3.2 Development

With the decision to use NARX with parallel-series configuration to perform the prediction, the
next step was to organize the data that would be used for its training. The data is the measured power of
a wind complex located at Rio Grande do Norte, Brazil. Such park is divided into wind turbines, each
with a nominal power of 2,1MW; circuits, that is the association of five or six wind turbines; the wind
farm, which is existing points within the wind farm complex where two or more circuits are linked; and
the wind complex corresponding to a union of all wind farms. Thanks to the project of strategic R&D,
we had access to the measured power of 6 months from the the circuits, wind farm and wind complex.

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

Important to note that during the six months of data, the wind complex went through a period of
maintenance, thereby affecting the overall quality of the time series. However, there were periods in
which were not affected, such as, for example, one that corresponds to a full month of data, and the
other resulting from 45 days in a row.
Such datas were separated into two sets: the training set, wich is the set that contains the time
series of the interval of 45 days; and the simulation set, which has the data pertinent of an entire month.
The data belonging to the training set were partitioned into two parts: one containing 80% of the total
of the data (this one, also separated into subgroups called training, validation and test, with a proportion
of 70%, 15%, 15% for each the from the total data partitioned), which is responsible for training the
network itself; and the other one containing the remaining 20%, which is responsible for verification
process of the training. All this processed happened for circuits, wind farm and wind complex.
After the separation, the NARX was built with only only 5 neurons on the occult layer, then
trained using Levenberg Marquad learning algorithm After the traning process, it was used the
verification dataset to verify the training. The verification dataset is responsiable for checking if the
NARX is able to predicted with data that has similar characteristics of the dataset used during the
training process.
To ensure good results, ten cycles of training and verification were performed. For each
verification process it was observed the mean squared error defined by the equation 4 bellow.

The mean squared error calculates the difference between the estimator (yi) and the estimated
(y^i) squared on the total size of a time series (n). The mean squared error will always be a positive
value, and the closer to zero, the better is the quality of prediction. Thanks to the mean squared error
the number of neurons present on the ocult layer can be determined by observing wich verification
result in the lowest error. In this work good results were achieved with only 10 neurons, the error rate
can be seem on the analysis of case section.
After choosing the NARX network that resulted on lowest error rate from verification process,
the dataset corresponded for the simulation process was used. The result from simulation can be seen
on the analysis of case section.

4. Analysis of case
The results of one step ahead prediction for the circuits, wind farm and wind complex obtained
through the previously presented neural networks, can be seen in this section.
Figure 3 present the one steap ahead prediction from the simulation process for the data
corresponding to the circuit.
The forecast can be observed better on the zoomed image. The actual power (estimator)of the
circuit is represented by the solid yellow line, while the predicted power (estimated) represented by the
dashed line.

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

Figure 3 - One step ahead power prediction for a circuit

Apparently, both lines overlap, but this is not the case, as can be seen in Figure 4, the estimated
values and estimator are very close.

Figure 4 - Difference between actual power and predicted power

This is a good sign, because it means that the network is being able to predict with low error
rate, this represents that the networks isn’t overfitting. The error rate can be seen in Table 1.

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL


Table 1 - Mean squared error for a circuit.

Step MSE

Verification (45 days) 2,1156E-11

Simulation (30 days) 2,4490E-11

Table 1 shows a comparison between the mean squared error obtained through verification and
simulation step.
The error rate during the simulation step is larger than the verification step, but that's to be
expected, since the data used in the simulation are different from the data used during training,
however, the difference between the errors of each step is minimal.
Confirmed that the neural network was working correctly, it was decided to simulate it with
data obtained from another circuit of wind turbines. In Figure 5 are the result of one step ahead
prediction for the period of one month for another circuit.
Observing the enlarged image in Figure 5, it is possible to note that the network is still capable
of carrying out the forecast one step ahead with data from a completely different circuit.

Figure 5 - One step ahead power prediction for another circuit for a month of data

The Table 2 presents the mean squared error.


Table 2. Root mean square error in another circuit.

Step MSE

Simulation (30 days) 1,7952E-09

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

The error value is slightly larger, because the network is predicting the power produced a
completely different circuit than the one that was used for traning, verification and simulation.
Despite the increase in error, it is still small, meaning that the neural network is still performing a a
good quality forecast.
It was intentioned to use that same network to predict the power of the wind farm, as well as for
the complex. However, the network would not be able to perform the forecast since the data from the
wind farm and the wind complex are very different from circuit wich was used in training. So it was
decided to create two new NARX networks, and train one with the time series of the wind farm and
train the another one with the time series of the wind farm complex.
In Figure 6 are the result of trained neural network trained with the data of the wind power farm.

Figure 6. One step ahead power prediction for the wind farm.

The new neural network is capable of predicting a step ahead perfectly, similar to that achieved in
the neural network presented earlier. The error rate is given in Table 3 below.

Table 3 - Mean squared error for the wind farm.

Step MSE

Verification (45 days) 9,3347E-11

Simulation (30 days) 1,3657E-10

As expected, the error rate is higher during the simulation step, because the network is using
different data to that that were used in training. With low error rate as presented in table 3, it can be
said that the neural network practically predicts a step ahead most of the time.
Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br
XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

Finally, in Figure 7 are the result of prediction of the power for the wind farm complex. The
prediction is performed by another neural network, that was trained with data regarding the power
produced by wind power complex.

Figure 7. One step ahead power prediction for the wind farm complex.

This new neural network is capable of predicting a step ahead, this time for the wind farm
complex. The quality of the forecast is be given in Table 4.
Table 4 – Mean squared error for wind power complex

Step MSE

Verification (45 days) 3,8613E-10

Simulation (30 days) 4,1911E-10

The minimum difference between the errors of the verification step and the step shows that this
simulation NARX network is capable of predicting a step ahead with quality.

5 Conclusions

With low values of error, NARX network proved to be able to predict the power one step ahead,
as was desired.
The values obtained through these predictions can help power smoothing control systems that
need to know the values of power beforehand to be able to act properly, as for example, in energy
storage systems batteries, wich nowadays is increasingly becoming common. If you wish, the network
still has the ability to be reconfigured in parallel, to achieve n steps ahead predictions.

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

BIBLIOGRAPHY

[1] Beudin, M., Zareipour, H., Schellenberglabe, A., & Rosehart, W. (2010). Energy storage for
mitigating the variability of renewable electricity sources: An updated review. Energy for
sustainable development(14), 302-314.
[2] Zhao, H., Wu, Q., Hu, S., Xu, H., & Nygaard Rasmussen, C. (2014). Review of energy storage
system for Wind power integration support. Applied Energy.
[3] Luo, X., Wang, J., Dooner, M., & Clarke, J. (2015). Overview of current development in electrical
energy storage technologies and application potencial in power system operation. Applied Energy,
511-536.
[4] IRENA. (2015). BATTERY STORAGE FOR RENEWABLES: MARKET STATUS AND
TECHNOLOGY OUTLOOK.
[5] Eyer, J., & Corey, G. (2010). Energy Storage for the Electricity Grid: Benefits and Market Potential
Assessment Guide . Sandia National Laboratories , Albuquerque.
[6] P. Mandic, D., & A. chambers, J. (2001). Recurrent naural networks for prediction. John Wiley &
Sons, LTD.
[7] Parai, A., Nassiri Moghaddam, H., Asadpour, S., Bahrampour, J., & Mansoori, G. (2010). A
comparison of artificial neural networks with other statistical approaches for the prediction of true
metabolizable of meat and bone meal. Poultry Science, 89(7), 1562-1568.
[8] Sfetsos, A. (2000). A comparison of various forecasting techniques applied to mean hourly wind
speed time series. Renewable Energy(23-25), 21.

[9] Haykin, S. (1999). Neural Networks - A comprehensive foundation. Prentice Hall International.
[10] Di Piazza, A., Di Piazza, M., & Vitale, G. (2016). Solar forecasting by NARX neural networks.
Renew. Energy Environ. Sustain (39).

SHORT BIOGRAPHY

Pedro Rosas, Ph.D.

He received his Ph.D. from Technical University of Denmark in wind energy and power systems. Was
researcher at Risoe National Lab from 1999 to 2003. Actually he is full professor at electrical
engineering department UFPE and has supervised several Works on MSc and Ph.D thesis as well as
coordinated projects on R&D wind power related. Lately has been working with Moura and CPFL
developing applications for batteries and power systems.

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

Gustavo José Luna Filho, M.Sc.

Holds a Master’s degree in Energy Processing and PhD in progress in storage systems connected to the
electric grid at UFPE and MBA in project management at Estácio University Center of Recife.
Graduated in Electrical Engineering, specializing in Supply Engineering in oil and natural gas. He is
currently engaged in research and development projects in renewable energy systems for generation,
transmission and distribution companies. He is a professor and coordinator of the Bachelor's and Post-
graduate courses in Electrical Engineering at the Centro Universitário Estácio do Recife

Luiz Henrique A. de Medeiros, Ph.D.

He received the B.S. and M.Sc. degree in electrical engineering from the Universidade Federal de Santa
Catarina, Brazil, in 1992 and 1994, respectively, and the Ph.D. degree from Institut National
Polytechnique de Grenoble – INP Grenoble, France, in 1998. He has been with the Universidade
Federal de Pernambuco since 1998, where is a Full Professor of the Departamento de Engenharia
Elétrica and the head of the LCMag – Electromagnetic Compatibility and Applied Electromagnetic
Laboratory. His research interests concern developing numerical methods, techniques and
measurements to improve electromagnetic compatibility, electromagnetic pollution, biological effects
of electromagnetic fields, power quality, electromagnetic detection of corrosion in electric systems and
BESS systems.

Pablo Luiz Tabosa Silva

He received his BSc in electrical engineering from Instituto Federal da Paraíba (IFPB) 2016. He is
currently doing his MSc on this Project at Universidade Federal de Pernambuco (UFPE).

Paola Furlanetto

Electrical Engineer with emphasis on Power Systems by the Federal University of Campina Grande and
Master of Science in Energy Systems at Northeastern University. As a professional, focus on power
systems, specially renewable energy, being able to address the challenges related to grid integration and
energy storage. She is currently a Technology and engineering researcher at ITEMM, developing
solutions for the automotive and energy markets. The scope of work focus on supporting ITEMM's
partners and clients to address challenges related to vehicular electrification and modernization of
Brazilian electric system.

Mariana Moraes

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br


XIV SYMPOSIUM OF SPECIALISTS IN ELECTRIC OPERATIONAL AND EXPANSION PLANNING

SEPTEMBER 30TH THRU OCTOBER 3RD OF 2018 / RECIFE / PE /BRASIL

Holds BSc in electrical engineering from Universidade Federal de Pernambuco (UFPE), Brazil, in
2015. She is currently taking a MBA degree at UFPE. Moreover, she is a researcher in R&D projects of
energy storage systems at ITEMM.

Wendell William Teixeira

Holds a Master's degree in electrical engineering from UNICAMP, he holds a degree in Electrical
Engineering from the Paulista State University Júlio de Mesquita Filho (2014). He is currently an
innovation analyst for the CPFL Energia Group. He has professional experience as a Product
Development Engineer in France, where he studied Industrial Systems Engineering. In Brazil, it has
been working in the development and management of ANEEL PD & I projects since 2014, with
projects related to new technologies applied to the electric sector.

Renato Povia Silva

Graduated in Economics at Universidade Estadual de Campinas – UNICAMP. The professional


trajectory was marked by strategy consultant Roland Berger, where he had the opportunity to develop
projects for companies in the electric sector, such as CPFL, Cemig, Eletronorte, Ampla and State Grid
Brasil Holding . At CPFL since 2011, he has been a specialist in the content and positioning area of the
Corporate Communication and Institutional Relations Board. In 2014, he was promoted to Strategy
Manager and in 2017 he took over the Innovation Management of the CPFL Energia Group.

Pedro A. C. Rosas, UFPE - pedro.rosas@ufpe.br

You might also like