8

2012-MLP Neural Network based Prediction of Coal Quality Categories.pdf

Embed Size (px)

Citation preview

  • International Conference on Innovative Technologies

    IN-TECH 2012

    Rijeka

    Proceedings

  • IN-TECH 2012

    Proceedings of International Conference on Innovative Technologies

    Editors: Zlatan Car Croatia Czech Republic Toma Pepelnjak Slovenia

    IN-TECH 2012 Organization committee:

    Zlatan Car Croatia Czech Republic Toma Pepelnjak - Slovenia Leon ikulec Croatia Marko Krulja Croatia Hrvoje Radelja Croatia Petr Dranar Czech Republic Vladislava Ostr Czech Republic

    Publisher: Faculty of Engineering University of Rijeka Printed by: Grafika Helvetica d.o.o. Printed in 250 copies. IN-TECH 2012 International Conference on Innovative Technologies runs from 26.09.2012 to 29.09.2012 in Rijeka, Croatia. E-mail: [email protected] URL: http://www.in-tech.info The CIP record is available in the computer catalogue of the University Library Rijeka under No. 121219001.

    ISBN 978-953-6326-77-8

  • International Conference on Innovative Technologies, IN TECH 2012,Rijeka, 26. 28.09.2012

    MLP NEURAL NETWORK BASED PREDICTION OF COAL QUALITY CATEGORIES

    M. Suljic1, and L. Banjanovic-Mehmedovic2 1JP Elektroprivreda BiH d.d. Sarajevo, ZD Rudnici ''Kreka'' d.o.o.-Tuzla,

    Bosnia and Herzegovina 2Faculty of Electrical Engineering, University of Tuzla

    Bosnia and Herzegovina

    Keywords: coal quality, coal calorific value, neural network prediction. Abstract. The significant coal deposits which represent the primary source of energy could be found on the territory of Bosnia and Herzegovina, Tuzla region especially. Therefore, the prediction of energy values is the most important task aiming to secure the optimal usage of coal energy value. The coal quality is an important property that depends on its chemical composition and points at useful energy value in coal. It is described with gross caloric value (GCV) and net caloric value (NCV). There are different methods for prediction the gross calorific value (GCV) of a coal sample based upon its proximate and/or ultimate analyses. One of intelligent method is neural network (ANN)-based method for prediction of the coal quality class. It is shown in this paper, that multi-layer perceptron (MLP)-based model based on coal and other data from technical control department of Coal mines ''Kreka'', achieved good reasonable prediction accuracy in prediction categories of coal (net calorific value, NCV).

    Introduction

    The coal has been major factor in intensive economic development of Europe as well as Bosnia and Herzegovina (BiH) during the last 40 years. Many regions have achieved their economic and industrial development thank to coal as a source of energy. According the statistics data, coal as a source of fossil energy secure about 30% of total energy production in the EU and around 59% in Bosnia and Herzegovina respectively [1,2]. It is well-proved fact that despite introduction of new renewable sources of energy,coal stays irreplaceable in energy supply of many countries as well as in Bosnia and Herzegovina. According to the structure and natural characteristics of Bosnia and Herzegovina (BIH), it is a complicated natural complex with significant wealth in coal. Total geological coal reserves are estimated at 5,5x109tons of which 3,2x109have proven reserves based on Law on geological reserves [3]. These reserves represent significant basis for electricity production by using either conventional(traditional) technologies or new combustion ones. The largest coal deposits are located in Tuzla region and they are escavated by a group of coal mines called ''Kreka'' Coal Mine Company with their headquarters located in city Tuzla. Lignite deposits are of pliocene age and they are devided into two synclines (Northern and Southern) with coal layers oriented in the direction northwest-soutwest shown in Fig. 1.

    Fig. 1 Kreka lignite basin

    Four coal layers seem to be present in the major part of the basin and they are called: floor,main,I intermediate and II roof layer.Thickness of coal layers varies from 8 to 25 m with dip angle ranging from small up to steep one.The depth of the layers is about 1250m below the surface,but only the layers which lay at the shallow part have been escavated by using open-pit and underground-pit methods.There are four mines operating within the Kreka Mines Company,as follows: ikulje (open pit mine), Dubrave (open pit mine), Mramor (underground pit mine) and Bukinje (underground pit mine). The measure of the amount of energy that a given quantity of coal will produce when burned is known as calorific value or heating value. Heating value is a rank parameter and a complex function of the elemental composition of the coal, but it is also dependent on the maceral and mineral composition.

    273

  • International Conference on Innovative Technologies, IN TECH 2012,Rijeka, 26. 28.09.2012

    The coal quality predicting is the application of science and technology to predict the energy value in coal. Therefore, the prediction of energy value is the task of enormous importance, which should ensure optimal utilization of coal with aim to increase energy efficiency. The coal quality is an important property that depends on its chemical composition and points at useful energy value in coal,so we use two measures for describing it: gross caloric value(GCV) and net caloric value(NCV).At present there are two methods for determination od these values[4]:proximate analysis and ultimate analysis. The objective of proximate analysis is to determine the relative amounts of moisture, ash, volatile matter and fixed carbon content of the coal. On the other hand, the objective of coal ultimate analysis is to determine the constituent of coal, in a form of its basic chemical elements: carbon(C), hydrogen (H), oxygen (O), sulfur(S) and other elements within the coal sample. The ultimate analysis is performed in a properly equipped laboratory by a skilled chemist, while proximate analysis can be determined with a simple apparatus [4]. Different researchers have proposed different equations for prediction the gross calorific value (GCV) of a coal sample based upon its proximate and/or ultimate analyses[5,6,7,8]. These correlations are mainly linear in character although there are indications that the relationship between the GCV and a few constituents of the proximate and ultimate analyses could be nonlinear. In recent years a researcher looking for simpler and cheaper methods for prediction the coal calorific value has begun to use nonlinear models such as artificial neural network[8,10,11]. Neural networks are powerful tools that have the abilities to identify nonlinear patterns between input and output values and can solve complex problems. Owing to their wide range of applicability and their ability to learn complex and non-linear relationships including noisy or less precise information ANNs are very well suited to solving problems in particularly towards the analysis of predicting of energy value in coal. Patel et al. (2007) developed seven nonlinear models for prediction of GCV with a special focus on Indian coals, using neural network analyses based on coal properties. It has been found that the performances of the ANN models are much better than those of their linear counterparts. Mesroghli et al. (2009) investigated the relationships of ultimate analysis and proximate analysis with GCV of U.S. coal samples by regression analysis and artificial neural network methods. Three set of inputs were used for the prediction of GCV values and the satisfactory prediction was obtained.Additionally, ANN concept is being successfully appliedin solvingwide variety of prediction [12,13], estimation and controlling issues that are encountered in power-plant. The objective of the present work is to develop ANN-based model by using coal and other data from technical control department of Coal mines ''Kreka'', for the predicting the categories of net calorific value (NCV).

    Data settings

    Data used in this paper to predict NCV have been taken from the technical control department of Coal mines ''Kreka'', which are achieved in laboratory conditions during the period 2005-2010.The samples with more than 50% ash, were excluded from the database. Analysis results for a total of 33256 coal samples were used. The number of samples and range of NCV for four operations within the Coal mines ''Kreka'' are shown in Table 1.

    Table 1.Range of NCV (as-received) for different Coal mines ''Kreka''

    Mine Number of samples Range of NCV (kJ/kg) Bukinje 949 5327 15019 Dubrave 14163 5657 13776 Mramor 5843 7056 15379 ikulje 12298 7112 12323

    The part of data which has been measured on sample of every 300 t was shown in Table 2.

    Table 2.Coal variables

    No. Coal variables in the every 300t frame Range of values 1 NRudnik (Mine) A, B, C,D 2 Eksplo (Type of excavation) 0, 1 3 Sinklin (Syncline) J, S 4 NAsortiman (Commercial grade of coal) M, O, S 5 Dan (Day) 1-7 6 BrVagona (Number of waggons) 1-15 7 GVlaga (Free moisture quantity) 13.4 52.4 8 HVlaga (Hydroscopic moisture quantity) 0-33.1 9 Pepeo (Ash) 1-50.2

    The selected subset of data contains ten attributes of which the nine are conditional and one attribute is predicting. Rows with Null values are ommited because it is extremly difficult to recover missing values. The attribute that predicts the Coal quality can belong to one of seven classes: A, B, C, D, E, F and G. Cross-validation methods are commonly used in examining the robustness of classifiers. In this study, a 10-fold cross validation was used: we split data set randomly into 10 subsets of equal size. Eight subsets were used for training, one subset for cross validating and one for measuring the predictive accuracy of the final constructed network. This procedure was performed 10 times so that each subset was tested once. Test results were averaged over 10 tenfold cross-validation runs. Data splitting was done without sampling stratification. One of the data set was used for testing MLP, while the remaining was used for training. The training and test sets consisted of 33256samples.

    Neural networks

    Neural Network (NN) represents a new generation of systems of information processing that shows characteristics of learning,memorizing and generalizations based on data trained.They are very efficient in tasks such as classification,function

    274

  • International Conference on Innovative Technologies, IN TECH 2012,Rijeka, 26. 28.09.2012

    approximation,optimisation and data clustering.Great populatity and success of neural network methods are resultat of their features that help us to solve complex tasks with high accuracy.The basis of neural networking is usage of the principles on which human brain functions and their application in resolving different tasks.Artificial neural networks(ANN) are connected with environment in two ways: first through inputs that have certain impact on the networks and secondly through output network which in turn affects the environment.It consists of potentially great number of processing elements,normally appearing in layers.Each element is connected with elements in the previous layers by means of adaptable strength or weight.The adaptation of these weights is performed by learning algoritham and this adaptation process improves the learning capability of the the system,thereby enabling it to generalize for new situation.The ANN model in our paper is based on one of the neural network architecture, named multi-layer perceptron[14].

    Multi-Layer Perceptron (MLP)

    This is maybe the most popular network architecture in use today. Each of its units performs biased weight sum of their inputs and passes this activation level through a transfer function to produce their input, and the units are arranged in a layered feed forward topology. The network thus have a simple interpretation as a form of input-output model, with the weights and thresholds (biases) the free parameters of the model. Such networks can model functions of almost arbitrary complexity with the number of layers and the number of units in each layer, determining the function complexity. Important issues of Multi-layer perceptron design take into account specification of hidden layers and the number of units in these layers. Once the number of hidden layers and number of units in each layers have been selected, the network's weight and the threshold must be set so as to minimalize the prediction error made by the network. This is the role of the training algorithms. The best known example of a neural network training algorithm is back propagation [15].

    Evolving MLP Structures With this background we designed and trained this network as below. The three-layer network with sigmoid transfer function for hidden layer and linear transfer function for output layer can represent any functional relationship between inputs and outputs, if the sigmoid layer has enough neurons [14], so we selected this three layer structure. MLP had 14 input and 7 output neurons. The gradient based back-propagation learning rule was used to determine the optimal MLP structure: learning rate ([0 ... 0,9] with step 0,1), learning momentum ([0 ... 0,9] with step 0,1) and hidden unit number ([1 ... 45] with step 1). Initial value of parameters MLP before optimization are: number of epochs is 300, the learning rate is 0.3, the learning momentum is 0.2 and hidden unit number is 9. The accuracy of training and test steps according to learning rate and learning momentum are shown in Fig. 2 and 3, respectively. The best result of training and test steps were obtained when the learning rate was 0.2 and learning momentum was chosen as 0.1.

    Fig. 2 The classification accuracy of training step according to Learning rate

    Fig. 3 The classification accuracy of training step according to Learning momentum

    Initial value of parameters MLP before optimalization has been as following: number of epoch was 300, the learning rate was 0.2 and the learning momentum was 0.1. The unit number in the hidden layer has been increased from 1 to 45 (incremented by 1). The performace of MLP depends heavily on initial conditions. Hence,the training and testing processes were repated 45 times. The results were obtined by averaging experimental results of 10-fold cross validation dataset. The accuracy of training and test steps according to hidden unit numbers (H) is shown in Fig. 4. The best result of training and test steps were obtained when the hidden unit number (H) was chosen as 24. Besides, the root means square error (RMSE) for training and test steps according to hidden unit numbers are given in Fig.7. As shown in Fig. 5, the minimum MSE errors for training and test processes were acquired when the hidden unit number (H) was 25 and 24, respectively. The best result of training and test processes are shown while the hidden units are 24. Then, the optimal MLP structure is14-24-7for the input, hidden and output layer respectively, the learning rate is 0.2and the momentum of learning0.1.

    Fig. 4 The classification accuracy according to hidden unit number

    Fig. 5 The RMSC according to hidden unit number

    85,085,285,485,685,886,086,286,486,686,887,0

    0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9

    Acc

    urac

    y L

    evel

    (%

    )

    Learning rate

    85,085,285,485,685,886,086,286,486,686,887,0

    0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9Acc

    urac

    y L

    evel

    (%

    )

    Learning momentum

    85,0

    85,586,0

    86,5

    87,0

    87,5

    88,0

    1 5 10152024252630354045

    Acc

    urac

    y L

    evel

    (%

    )

    Hiden unit number

    15,015,516,016,517,017,518,0

    1 5 10 15 20 24 25 26 30 35 40 45

    RM

    SE

    Val

    ue

    Hiden unit number

    275

  • International Conference on Innovative Technologies, IN TECH 2012,Rijeka, 26. 28.09.2012

    Result and discussion

    With a preliminary analysis we have developed the optimal MLP neural network that give the best results. Testing optimal MLP neural network was carried out with the 1000 new samples. The exact and predicted values for each unseen sample by MLP network is shown in Table 3 and Fig. 6. Table 3.The results of prediction for each quality class

    MLP neural network TP Rate FP Rate Precision Recall Class

    training time: 100 learning rate:0.2 learningmomentum:0.1 hidden unit number: 24

    95.00 1.00 91.40 95.00 A 86.10 1.40 87.90 86.10 B 87.20 1.30 90.30 88.70 C 94.90 5.60 92.00 94.90 D 87.60 2.20 92.50 87.60 E 93.80 0.40 88.20 93.80 F

    0 0 0 0 G 91.30 3.20 91.30 91.30 AVG

    Fig. 6 Comparison between exact and predicted values

    Therefore the proposed ANN model with the developed structure shown in Table 3 can perform good predicting the categories of coal quality in Coal Company ''Kreka''.

    Conclusion

    The results of MLP neural network model used for predicting categories of coal in Coal Mine Company ''Kreka'' have shown that MLP neural network was achieved good performance and reasonable prediction accuracy for this model. The results suggest that this neural network could be an important tool in predicting coal quality. For future work, the experiment can be extended with more distinctive attributes from geological database to get more accurate results useful in improving the outcome. Also,experiments could be done by using other algorithams in order to get more accurate energy value of coal.

    References

    [1] European Association for Coal and Lignite. (2012, January) EURACOAL - European Association for Coal and Lignite. [Online]. http://www.euracoal.be/pages/home.php?idpage=1,

    [2] Plan Bleu. (2012, January) Plan Bleu - Regional Activity Center. [Online]. http://www.planbleu.org/publications/atelier_energie/BA_Summary.pdf

    [3] Vlada Federacije Bosne i Hercegovine. (2012, Januay) Strateski plan i program razvoja energetskog sektora FBiH. [Online]. http://www.fbihvlada.gov.ba/bosanski/izdvajamo/SPP-sept-08-PRIJEDLOG.pdf

    [4] B.G. Miller, Coal Energy Systems. London, United Kingdom: Elsevier Academic Press, 2005.

    [5] P.H. Given, D. Weldon, and J.H. Zoeller, "Calculation of calorific values of coals from ultimate analyses: theoretical basis and geochemical implications," Fuel, vol. 65, pp. 849854, 1986.

    [6] D.M. Mason and K.N. Gandhi, "Formulas for calculating the calorific value of coal and coal chars: Development, tests, and uses," Fuel Processing Technology, vol. 7, pp. 11-22, 1983.

    [7] T. Cordero, F. Marquez, J. Rodriquez-Mirasol, and J.J. Rodriguez, "Predicting heating values of lignocellulosic and carbonaceous materials from proximate analysis," Fuel, vol. 80, pp. 15671571, 2001.

    [8] Sh. Mesroghli, E. Jorjani, and S. Chehreh Chelgani, "Estimation of gross calorific value based on coal analysis using regression and artificial neural networks," International Journal of Coal Geology, vol. 79, pp. 4954, 2009.

    [9] S.U. Patel et al., "Estimation of gross calorific value of coals using artificialneural networks, Volume 86, Issue 3,.," Fuel, vol. 86, no. 3, pp. 334-344, 2007.

    [10] S. Chehreh Chelgani, J.C. Hower, E. Jorjani, Sh. Mesroghli, and A.H. Bagherieh, "Prediction of coal grindability based on petrography, proximate and ultimate analysis using multiple regression and artificial neural network models," FuelProcessing Technology, vol. 89, no. 1, pp. 13-20, 2008.

    [11] H. Salehfar and S. A. Benson, "Neural Network Based Power Plant Coal Quality Analysis," in IEEE, 29th Annual North American Power

    6000

    7000

    8000

    9000

    10000

    11000

    12000

    13000

    1 51 101 151

    Ene

    rgy

    valu

    e

    Sample

    Predicted K Exact K

    276

  • International Conference on Innovative Technologies, IN TECH 2012,Rijeka, 26. 28.09.2012

    Symposium, Laramie, 1997.

    [12] Y. Ozel, I. Guney, and E. Arca, "Application of Neural Network to the Cogeneration System by Using Coal," INTERNATIONAL JOURNAL OF ENERGY, vol. 4, no. 1, 2007.

    [13] M.T. Hagan, H.B. Demuth, and M.H. Beale, Neural Network Design. Boston, Massachusetts: PWS Publishing, 1996.

    [14] S. Haykin, Neural Networks: A Comprehensive Foundation. New York: Macmillan College Publishing, 1994.

    277