Integrated solutions based on neural networks for optimizing energy management in a microgrid

Dragomir Otilia Automation, Information and Electrical Engineering Dept.

Valahia University of Targoviste Targoviste, Romania

drg_otilia@yahoo.com

Dragomir Florin Automation, Information and Electrical Engineering Dept.

Valahia University of Targoviste Targoviste, Romania

drg_florin@yahoo.com

Abstract—The article proposes the integration of an intelligent energy management system, in a microgrid with distributed production of energy from renewable sources (PD-RES), equipped with artificial intelligence elements (neural networks), capable of load forecasting on short tim horizon. In relation with identified or predicted situations, the decision support system that integrates these solutions, proposes, to consumer-producer (prosumer) of “green” energy, to act in a proactive manner for: reconfiguration of the microgrid’s architecture, improving profits, and reducing the microgrid’s vulnerability. From a practical point of view, the article results are based on data monitored form a three phase microgrid with 25kW installed power, equipped with energy storage elements, produced from renewable energy sources (wind and sun).

Keywords—renewable energy systems, energy system management, neural network, distributed power

I. INTRODUCTION Unlike conventional sources of energy, the non-

conventional energies are inexhaustible, but they generate high costs of integrating them into the existing power grid. This important economic aspect enhances the practitioner’s decision to orient towards RES. Two new important trends on the national energy market are to be noticed: first it’s the consumers involvement in the complex process of efficient management of the energy and second it can be noticed the increased attention, paid to both the technical plan and to the organizational and economical plan, for the energy production from RES. Also, the success of the national energy market entry, recently liberalized, will be conditioned to a large extent on the capacity of each producer to predict, through different methods, the future evolution of the amount of energy produced from renewable sources.

In this context, this paper: describes the European and national context for distributed power generation from renewable energy sources, identifies the opportunities and existing barriers in this field, provides a state of the art of microgrids with PD- RES problems identified in the scientific literature and, in the end, proposes two solutions based on: multilayer perceptron network (MLP) and a radial basis function (RBF) neural networks for optimizing energy management in the microgrid.

II. DISTRIBUTED POWER GENERATION FROM RENEWABLE ENERGY SOURCES CONTEXT

A. European context The current energy policy of the European Union (EU)

considers the security of supply, competitiveness and sustainability as central goals. In order to achieve these targets, through European strategies [1] are imposed a series of constraints ("objective 20-20-20"): 20% reduction in emissions of greenhouse gases compared to 1990, providing 20% of entire EU energy consumption by renewable energy sources (RES) and a reduction in energy use by 20% compared to a similar scenario in which no action regarding sustainability has been taken. To achieve these objectives and generate a "sustainable growth" a policy of encouraging distributed generation from RES, such as wind, biofuels and solar power must be followed.

Intense concerns at European level regarding the distributed generation (PD) from RES were materialized by setting up a giant group (cluster) of projects called Integration of Renewable Energy Sources and Distributed Generation into the European Electricity Grid (IRED cluster), cluster that integrates projects DISPOWER , MICROGRIDS, CRISP, DGFACTS, SUSTELNET, DGNET , INVESTIRE from FP6 [2].The main objectives of this consortium are to develop scientific and technical knowledge, development of indispensable standards under increased number of small producers and developing partnerships and projects under the 7th Framework Programme according to different technological platforms.

Within the European project DISPOWER, the German institute Fraunhofer ISE together with the project partners has developed a management system for power flows and its quality (PMS) that works on a pilot plant since 2005. The PMS is a partially decentralized control structure that represents a compromise between centralized and fully decentralized control and brings major advantages to the energy system.

The studies that were conducted following these projects highlighted the need for an energy management system from micro to macro level, the existing control strategies not being always successfully applied.

B. National context The share of RES in the electricity production in Romania

is currently 17.8%. EU has set for Romania a 24% target for energy generation from RES by 2020, but there have also been identified needs for investments and large operating costs as main barriers for the successful implementation of an increased generating capacity. Compatibility with EU objectives in the field of clean energy and national levels is achieved through regional policy [3].

European policies had a national resonance since 2003, when the draft for the project Strategy for the use of renewable energy, approved by the GD 1535 in December 2003, published in Official Gazette of Romania, Part I, Year XV - No. 8, January 7, 2004 [4]. In the capitalization strategy for SER, the declared wind potential is 14,000 MW (installed capacity), which can provide an amount of energy of about 23,000 GWh / year. These values are an estimate of the theoretical potential and should be nuanced according to the technical and economical possibilities for exploiting.

The strategy also proposes the installation of 120 MW by 2010 and a further 280MW by 2015. According to this evolution, the electricity produced from wind sources would provide about 1.6% of gross electricity consumption in 2010. Compared to the amount of energy provided from renewable sources without large hydro power, wind power could provide 12.3% of this amount. Re-analyzing the data in the strategy, we believe that there are sufficient reserves for a much more important development of the wind applications than the one stipulated. Compared to a technical harnessed potential of 3,600 MW (8,000 GWh / year), the target quotas for the wind applications can be 200 MW in 2010 and 600 MW in 2015.

The integration of RES in the national energy system structure has created a favorable context for the proposal of this article.

III. TOWARDS INTEGRATION OF INTERACTIVE TECHNOLOGIES IN MICROGRIDS WITH PD- RES ENERGY

MANAGEMENT SYSTEM

A. State-of-the art of microgrids with PD- RES problems The problematic in microgrids with PD- RES area is vast.

In summary, the identified problems, by reviewing the literature are: to increase conversion efficiency [5], [6]; to identify new opportunities complementary to the limitations of existing solutions [7]; to limit the variability of the power production through integration of RES [5]; to balance the insufficient number of providers reported to market needs [8]; to increase the number of consumers [9]; to provide and to promote sustainable buildings and housing projects [10]; to identify the quantities of renewable energy in order to highlight the risks and social and economic costs [11]; to understand the effects associated with the use of RES and the increasing complexity of systems [12]; to identify measures and solutions to ensure: the sustainability of renewable resources [13], the impact of renewable energy for the final beneficiaries in terms of cost / benefit ratio [14] and [15], the importance of being efficiently informed in order to make better decisions about energy options. [16].

B. Opportunities and existing barriers Unfortunately Romania follows a centralized approach of

the regional policy and although the country is covered adequately with electricity networks and the potential development for RES is high, mainly from wind energy, its aging infrastructure (30% of it was built in the 1960s) causes significant losses along the energy supply chain.

In addition, for a large number of energy resources, the current energy systems are hardly scalable.

The European Commission believes that the current energy infrastructure is inadequate to connect and serve all of Europe and recognizes the challenges in improving both the private sector as well as the national governments, therefore it proposes the introduction of EU-wide downward directives in order to modernize and sufficiently interconnect national energy networks, with the ultimate objective of a single European market [1].

Another key aspect of the current energy markets is the deficiency (and security) to supply energy from fossil fuels. Since most sources and global reserves are outside the EU, transport, financial and political problems could be barriers to the availability of fossil fuels.

A final negative aspect identified is the energy demands (and safety) during peak periods. These are (is) a problem during periods with more and more extreme phenomena, which require more power. A properly connected infrastructure could prevent future crises by supporting energy supply that can be easily transported through the European energy network. This however does not replace the need for the decentralization of the power system and the need for orientation towards microgrids with distributed renewable production.

Romania has little experience with EU funds and currently invests very little in energy, but the following actions are encouraged: extension of the existing electricity networks, creating a new infrastructure to maximize the energy potential from RES, energy quality improvement and connection of the Romanian research with the evolution and requirements of the international socio-economic environment.

IV. SOLUTIONS BASED ON NEURAL NETWORKS FOR OPTIMIZING ENERGY MANAGEMENT IN A MICROGRID

To overcome the mentioned barriers and the concerns identified, this article proposes to redefine the role of consumer of energy in "prosumer" (consumer and producer of energy from renewable sources) in the context of a reorganized decentralized energy market, now reported to intelligent microgrids (smart grids).

In this paper we will focus on short term load forecasting using ANNs, precisely on MLP and RBF, which are the most popular and widely-used paradigms in many applications, including energy forecasting. Integration of interactive technologies based on neural networks in the microgrids energy management system with PD- RES optimizes: functioning from an economical point of view, active control of

distributed generation, controlled consumption, loading the storage equipment.

A. Simulation conditions Define abbreviations and acronyms the first time they are

used in the text, even after they have been defined in the abstract. Abbreviations such as IEEE, SI, MKS, CGS, sc, dc, and rms do not have to be defined. Do not use abbreviations in the title or heads unless they are unavoidable.

The forecasting performances of MLP and RBF are illustrated using a dataset with 120 data points {y (t), u (t)}, from t=1 to t=120, representing the data from the monitored plant. All the experiments are carried out in MATLAB 7.1.The DPcg (difference between the electricity produced from renewable energy sources and consumed) (kW) is considered as output y(i) of the MLP/ RBF model and the radiation (W/m2) is input x(i) (see Table I).

MLP neural network[17], used for performing STLF, is a feedforward artificial neural network model that maps sets of input data onto a set of appropriate output. It consists of multiple layers of nodes in a directed graph, which is fully connected from one layer to the next. RBF neural network, used for performing STLF, has an input layer, one hidden layer and an output layer. The neurons in the hidden layer contain Gaussian transfer functions, whose outputs are inversely proportional to the distance from the center of the neuron. The characteristics of these neural networks are summarized in Table II. The numerical procedure used for training MLP/ RBF accepts only values lying in the (0, 1) range. Hence, the data used for training are normalized before starting the training session and de-normalized at the end of the training.

TABLE I. DATASET PARAMETERS

Measured data Units Mean value Standard deviation

Radiation W/m2 0.9255 97.6705 DPcg kW 0.8156 130.9313

TABLE II. MLP AND RBF PARAMETERS

Architecture MLP RBF

Number inputs 1 1 Number layers 1 hidden layer with 5

nodes 1 output layer with 1 node

1 hidden layer with 5 radbas neurons 1 output layer with with purelin neurons

Transfer functions

tansig- hidden layer purelin- output layer

gaussian - hidden layer purelin- output layer

Performance function

MSE MSE and MAE

Training mode Supervised Training method Levenberg-Marquardt

Train epochs 1000 Initial MSE goal 0.0098

Initial spread 0.02719 Learning rate (lr) Tolerance (goal)

0.05 10-3

B. Training the ANNs

The training method of MLP is Levenberg -Marquardt (which is a combination of gradient descent and Newton’s Method). This one is often the fastest backpropagation algorithm in the toolbox Matlab and it is used to minimize the error between its values and the target values. Two types of training have been implemented: In the first case (case 1) all data in a specified time period are used for training. In the second case (case 2), the data are split into two subsets, forming the training vector and the validation vector. In particular, the validation vector includes 20% of the data, while the other 80% forms the training vector.

The training results and the real outputs (targets), for supervised and controlled learning, are illustrated in Fig. 1.

The stop criterion for the training phase is typically implemented as maximum number of iterations (epochs) or training error lower than a specified threshold value. When using the controlled learning, the stop criterion changes, and the training stops when the error computed by using the validation vector starts to increase.

The performance indicators (regression analysis) concerning the training phase are summarized in Table III, comparing the two learning cases.

0 10 20 30 40 50 60-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

3MLP prediction- case 1

input

p

real outputMLP output

0 10 20 30 40 50 60-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

3MLP prediction- case 2

input

outp

ut

real outputMLP output

Fig. 1. MLP output in relation with real output in case of a) supervised learning and b) controlled learning.

For the RBF neural network the dataset is divided into train and test subsets. 60% of the data set is selected as the training data and remained data set is selected as the testing data. For each run, the number of neurons, deviations of the radial units, MAE (Mean Absolute Error) and MSE (Mean Square Error) are computed in order to reach the MSE goal 0.0115.

TABLE III. PERFORMANCE INDICATORS CONCERNING THE TRAINING PHASE OF MLP

Case 1 Case 2 Indicatorsa value value Gradient 0.035321 0.003255

Mu 0.001 0.001 M 0.9965 0.9961 B -0.1347 -0.1389 R 00.9972 0.9965

aGradient (the minimal gradient which is allowed is 1e-10), Mu (the current factor 0.001<mu< 1e10), R ( regression value R=1 means perfect correlation), M ( slope of the linear regression), B ( output

intercept of the linear regression).

a b

In training phase of RBF the following steps are repeated until the network's mean squared error falls below goal or the maximum number of neurons are reached: 1) the network is simulated, 2) the input vector with the greatest error is found 3) a radbas neuron is added with weights equal to that vector and 4) the purelin layer weights are redesigned to minimize error.

Firstly, it was investigated how the spread of the hidden layer base function affects the network’s performance (see Fig 2).

The initial downward trend of MSE due to spread growth isn’t the same all over training set. This indicates the need for consideration of a second parameter in the evaluation of RBF training performance. This is the number of neurons in the hidden layer.

The number of neurons in the hidden layer is very important in design issue of an RBF network. Therefore, the experiments have been conducted on different RBF networks which has 6 neurons to 25 neurons located in the hidden layer. Using more neurons than that is needed causes an over learned network and moreover, increases the complexity of the RBF network (see Table IV).

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.002

0.004

0.006

0.008

0.01

spread

MS

E

RBF training

Fig. 2. MSE in relation with spread for training phase

TABLE IV. TRAIN RESULTS OF RBF SIMULATIONS.

Train

Spread H MSE (*10-3) 0.027 25 7.4966 0.127 15 7.8996 0.227 11 9.6653 0.327 8 6.4121 0.427 7 9.3638 0.527 7 8.6819 0.627 6 8.8989 0.727 7 6.1955

C. Test and validation The vector used for MLP to test the performance of the

network before its effective use. Validation vectors are used to stop training early if the network performance on the validation vectors fails to improve. In first case, the best validation performance is 0.051975 reached at epoch 5 and in the second case, is 0.0070157 reached at epoch 4 (see Fig.3).

Fig. 3. The best validation performance a) supervised learning and b) controlled learning.

The goal of the tests for RBF is, given training and test data, to choose the input parameters MSE goal, spread and Hmax (hidden layer neurons number) to minimize MSE value. Thus, the input parameters have been initialized with: MSE goal= 0.0098, the minimum distance between clusters of different classes MNDST =0.8156, spread0 = 0.2719 and Hmax0 = 60.

D. Results analysis The choice of error measures to help comparing forecasting

methods has been much discussed. A major part of them have been summarized by Dragomir [3]. Because of manner of the penalization of large errors, measures based on mean squared error (MSE) and mean absolute error (MAE) are often employed in performance evaluation Also, it is generally recognized that error measures should be easy to understand and closely related to the needs of the decision-makers. In consequence, we have used MAE a,d MSE to evaluate MLP and RBF performances in short term load forecasting.

Fig. 4 indicates the training errors in both cases. Precisely, it shows the evolution of MAE during the training and testing processes.

The MSE is used to determine how well the predicted output fits the desired output. More epochs generally provide higher correlation coefficient and smaller MSE values. Fig. 5 shows the evolution of MSE during the training and testing processes, which stops when the validation error starts to increase. As expected, the error values (MAE and MSE) in controlled learning (case 2) of MLP are higher. They put in relation target and network’s output on training and testing data sets and give an overview of generalization and memorization abilities of MPL.

0 10 20 30 40 50 600

0.1

0.2

0.3

0.4

0.5

0.6

0.7MAE- case 1 and case 2

input

erro

r

MAE case 1MAE case 2

Fig. 4. The evolution of MAE during the training and testing processes

a b

0 10 20 30 40 50 600

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5MSE- case 1 and case 2

input

erro

r

MSE case 1MSE case 2

Fig. 5. The evolution of MSE during the training and testing of MLP

The error measures are only intended as summaries for error distribution. This distribution is usually expected to be normal white noise in a forecasting problem, but it will probably not be so in a complex problem like load forecasting. No single error measure could possibly be enough to summarize it. The shape of the distribution should be suggested. A total error low means keeping the model simple.

In both learning cases, the computed mean and standard deviation (Mean_case1= -0.0199, Mean_case2=-0.0086 and std_case1=0.1235, std_case1=0.0927) are very close one with each other. Keeping a constant standard deviation of the prediction error gives confidence to this MLP architecture and recommends it as well fitted for short-term forecasting.

The predictions made by RBF neural network over the test dataset in relation with the measured outputs (targets) are illustrated in Fig. 6. As we can see, the output of the RBF network is a measure of distance from a decision hyper plane, rather than a probabilistic confidence level. As we can observe, the small number of test data has a bad influence over the forecasting. accuracy.

0 2 4 6 8 10 12-1.2

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4RBF prediction (testing phase)

input

outp

ut

real outputRBF output

Fig. 6. RBF outputs vs. targets in testing phase.

The quality of the possible solutions are calculated using MSE and MAE between the actual output of the RBF network and the desired output (see Table V).

TABLE V. RESULTS OF RBF SIMULATIONS IN TEST PHASE

Test

Spread MAE (*10-1) MSE (*10-1) 0.027 4.9542 6.4668 0.127 1.9653 1.4197 0.227 0.6805 0.1354 0.327 0.5751 0.0766 0.427 0.5456 0.0786 0.527 0.8884 0.1448 0.627 0.5706 0.0730 0.727 0.9046 0.1332

Fig. 7 indicates the RBF testing errors with the help of MAE and MSE.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

spread

erro

r

MAE and MSE for RBF testing phase

MSEMAE

Fig. 7. MAE and MSE values of RBF in testing phase.

Fig. 7 shows that, the growth of spread values until 0.3 has a big influence over MAE and MSE values. These ones decreas a lot, from 4.9542 to 0.5751 MAE and from 6.4668 to 0.0766 MSE. The trend change when the spread reach 0.327 value. The error values increase and indicate that the optimal values for spread and number of neurons in hidden layer has to be locate in this area.

At the beginning, Hmax was equal with the number of training points. The training tests with variable number of neurons in hidden layer have showed that 8 is the optimum number for the neurons in hidden layer, much less than the number of training points. At the end of RBF training, the optimum spread value found is 0.327.

V. CONCLUSION Power system operators are now forced to deal with an

increasing number of problems, largely related to increased number of loads, new environmental policies and economic pressures of the market. Large scale adoption of RES, able to contribute to the world’s energy needs and partly to resolve these problems, is slow. This article may be viewed as another form to promote the proliferation of alternative energy sources, developing reliable forecasting models dedicated to macro energy harvesting (RES harvesting).

The paper dealt with MLP and RBF neural networks, which are the most popular and widely-used paradigms in many applications, including energy forecasting, in order to obtain accurate load predictions on short term horizon. The proposed networks architectures illustrated and discussed using a data based obtain from an experimental plan, show that RBF neural networks are able to surpass the multilayer feedforward neural networks as they are simple structure and have the ability to model any nonlinear function in a straight forward way. RBF networks are local approximators with nonlinear input-output mapping. Theirs knowledge representation is localized. Thus, RBF network are able to learn faster and suffer less from interference, as compared to multilayer feedforward neural networks for example.

ACKNOWLEDGMENT This work was supported by a grant of the Romanian

National Authority for Scientific Research, CNDI– UEFISCDI, project code PN-II-PT-PCCA-2011-3.2-1616.

REFERENCES [1] http://ec.europa.eu/europe2020/europe-2020-in-a-

nutshell/priorities/index_ro.htm [2] IRED cluster http://www.ired-cluster.org/, DISPOWER project

www.dispower.org şi MICROGRIDS project http://microgrids.power.ece.ntua.gr/micro/default.php, CRISP project http://www.ecn.nl/crisp, DGFACTS project http://dgfacts.labein.es/dgfacts/index.jsp, SUSTELNET project http://www.sustelnet.net/, DGNET project http://www.dgnet.org/, INVESTIRE project http://investire-network.com/

[3] http://ec.europa.eu/regional_policy/activity/energy/index_ro.cfm [4] http://www.anre.ro/documente.php?id=393 [5] G. Boyle, Renewable Energy: Power for a Sustainable Future,

Oxford University Press, 2012 [6] B. Sorensen, Renewable Energy, Fourth Edition: Physics,

Engineering, Environmental Impacts, Economics & Planning, Academic Press, 2010

[7] G.Masters, Renewable and Efficient Electric Power Systems, Wiley-IEEE Press, New Jersey, 2004

[8] H. Kohl, The Development, R. Wengenmayr, & T. Buhrke, Renewable Energy, Weinheim: Wiley-VCH, 2008, pp. 4-14

[9] A.V. Da Rosa, Fundamentals of Renewable Energy Processes, Oxford: Academic Press, 2012

[10] W. Kemp, The Renewable Energy Handbook, Revised Edition: The Updated Comprehensive Guide to Renewable Energy and Independent Living. Tamworth: Aztext Press, 2009.

[11] D. MacKay, Sustainable Energy - Without the Hot Air, Cambridge: UIT Cambridge Ltd., 2009

[12] M. S. Kaltschmitt., Renewable Energy: Technology, Economics and Environment. Berlin: Springer, 2010

[13] R. Bryce, Power Hungry: The Myths of "Green" Energy and the Real Fuels of the Future, New York: PublicAffairs, 2011

[14] D. Chiras, The Homeowner's Guide to Renewable Energy: Achieving Energy Independence Through Solar, Wind, Biomass, and Hydropower, Gabriola Island: New Society Publishers, 2011

[15] G. Boyle, Renewable Electricity and the Grid, London, 2007. [16] R. Rapier, Power Plays: Energy Options in the Age of Peak Oil,

Apress, 2012 [17] Program CE – Altener – “Photovoltaic Training Courses for

Candidate Countries”, proiect SolTrain, 2002- 2005 [18] O. Dragomir, F. Dragomir, E. Minca and I. Brezeanu, “MLP

neural network as load forecasting tool on short- term horizon”, in

Proc. The 19th Mediterranean Conference on Control and Automation, (MED'11), 2011, Corfu, Grecia