- Email: [email protected]
Application of artificial neural networks to co-combustion of hazelnut husk–lignite coal blends
Bioresource Technology 200 (2016) 42–47
Contents lists available at ScienceDirect
Bioresource Technology journal homepage: www.elsevier.com/locate/biortech
Application of artificial neural networks to co-combustion of hazelnut husk–lignite coal blends Zeynep Yıldız, Harun Uzun, Selim Ceylan ⇑, Yıldıray Topcu Ondokuz Mayıs University Chemical Engineering Department, 55139 Samsun, Turkey
h i g h l i g h t s First study on application of ANN model to co-combustion process. The coefficient of determination for the developed model was 0.9995. ANN model predicted TG curves for co-combustion of blends effectively.
a r t i c l e
i n f o
Article history: Received 17 August 2015 Accepted 27 September 2015 Available online 8 October 2015 Keywords: Co-combustion Hazelnut husk Lignite coal TGA Artificial neural network
a b s t r a c t The artificial neural network (ANN) theory is applied to thermal data obtained by non-isothermal thermogravimetric analysis (TGA) from room temperature to 1000 °C at different heating rates in air to study co-combustion of hazelnut husk (HH)–lignite coal (LC) blends of various composition. The heating rate, blend ratio and temperature were used in the ANN analysis to predict the TG curves of the blends as parameters that affect the thermal behavior during combustion. The ANN model provides a good prediction of the TG curves for co-combustion with a coefficient of determination for the developed model of 0.9995. The agreement between the experimental data and the predicted values substantiated the accuracy of the ANN calculation. Ó 2015 Elsevier Ltd. All rights reserved.
1. Introduction Coal remains the most important primary energy source in power generation, amounting for 36% of the total globallygenerated power. Turkey has rich lignite reserves that are used in thermal power stations. However, coal combustion also is one of the main causes of global warming. Coal combustion is globally responsible for 43% of CO2 emissions from fossil-fuel combustion, with 28% being emitted from coal-fired power plants (Authier and Moullec, 2013; Buratti et al., 2015). Co-combustion of coal with biomass provides a cost-effective means to reduce CO2 and other greenhouse gas emissions, fossil fuel consumption, and the amount of landfilled organic matter. Biomass combustion is considered carbon-neutral because the CO2 released during combustion is in principle withdrawn from the atmosphere by photosynthesis during the growth of plants (Hu et al., 2015; Buratti et al., 2015; Peng et al., 2015; Gil et al., 2015). Furthermore, biomass is often used to achieve better control of the burning process. In co-combustion, biomass increases the volatile matter ⇑ Corresponding author. E-mail address: [email protected] (S. Ceylan). http://dx.doi.org/10.1016/j.biortech.2015.09.114 0960-8524/Ó 2015 Elsevier Ltd. All rights reserved.
content, providing a more stable flame; moreover, many sources have lower ash content than coal, thus reducing waste and increasing efficiency. Although the higher proportion of oxygen and hydrogen to carbon atoms in biomass does result in a lower heating value for biomass compared to coal, the higher oxygen content of biomass also tends to give it a higher reactivity than coal, and thus a lower activation energy barrier for devolatilization and oxidation (Celaya et al., 2015). Turkey holds a vast potential of biomass energy resources, and especially of agricultural wastes. Although this potential is available, most of the waste biomass species could not be evaluated properly for energetic purposes (Bilgen et al., 2015). Turkey is the world’s leading hazelnut producer with production of approximately 500,000 tons per annum (Çepeliogullar and Pütün, 2013); an exorbitant amount of hazelnut husk (HH) is generated during processing of the hazelnut fruit as agricultural waste biomass. These residues have no alternative uses, and because of their large storage volumes they also cause a disposal problem (Ceylan and Topçu, 2014). To explore the possibilities of alleviating some of these technological burdens, in this study the potentials of HH as blending biomass material are studied. A knowledge of the co-combustion
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
characteristics is crucial to achieve effective design and operation of the co-combustion process. Thermogravimetry has been widely used to characterize the thermal process, and to describe the profiles and the effective kinetic parameters for various decomposition reactions in order to predict their thermal behavior (Peng et al., 2015; Toptas et al., 2015). Numerous studies dealing with the application of TG analysis to evaluate combustion behavior of coal and biomass alone and their blends in co-firing application have been reported prior to this study (Sahu et al., 2014; Liu et al., 2015; Chen et al., 2015; Yuanyuan et al., 2015; Goldfarb and Ceylan, 2015; Gil et al., 2015). In these studies, mathematical models were used to explain or predict the process behavior, and to analyze the effects of different process variables on the process performance. Several commonly used methods for kinetic parameter calculation have been established, notably the Kissinger–Aka hira–Sunose (KAS) method, the Flynn–Wall–Ozawa (FWO) method, and the Friedman method, among others. However, an inherent deficiency of these methods is their inapplicability to confirm three kinetic parameters simultaneously. Also, the complexity of co-combustion phenomena has handicapped the development of a unifying model suitable for design purposes. These methods fail particularly in modeling secondary and some complex multistep reactions (Carsky and Kuwornoo, 2001). Artificial intelligence systems, such as neural networks, are widely accepted as a technology that can be applied to non-linear problems, and (once trained) can be used for prediction and generalization at high speed. Artificial neural networks (ANN) models use a non-physical modelling approach which correlates the input and output data to form a process prediction model. The ANN models have proven their potential in prediction of process parameters in energy-related processes (Mikulandric et al., 2014; Ata, 2015; Gajic et al., 2015; Sßahin, 2015) but their potential to predict parameters of a biomass co-combustion process is yet to be explored. In this work the combustion behavior of HH, LC and their blends was investigated through TGA at different heating rates and blend ratios. Optimizing the operating conditions for such fuel blending requires a greater understanding of the thermal characteristics coal–biomass blends; only accurate prediction of thermal behavior can provide a better design and operation of industrial systems (Celaya et al., 2015). Thus, the aim of this work is to use ANN to obtain accurate thermal behavior of multi-step co-combustion reactions of HH–LC blends. 2. Methods 2.1. Proximate and ultimate analysis Coal samples were supplied from Elbistan thermal power plant (Kahramanmarasß, Turkey). HH was provided by a local company. Table 1 Proximate and ultimate analysis of HH and LC.
a db
Proximate analysisdb, %
Lignite coal (LC)
Hazelnut husk (HH)
Volatile matters Ash Fixed carbona
52.10 31.50 16.40
74.24 6.01 19.75
Ultimate analysisdb, % C H N S Oa H/C O/C Calorific value/MJ kg1
61.79 5.94 1.82 4.31 26.14 0.10 1.15 20.20
43.09 5.81 1.30 0.142 49.66 0.42 0.13 17.14
Calculated by difference. Dry basis.
43
Both HH and coal samples were dried in an oven at 105 °C for 8 h to remove water. The samples were ground and sieved to particle size of 63–125 lm to avoid the heat- and mass-transfer limitations during experiments. The materials were characterized by using conventional proximate and t ultimate analyses, and the results are reported in Table 1. Proximate analysis is performed by a Simultaneous DTA-TG Analyzer (Shimadzu, Japan) based on the ASTM D 5142-04 ‘‘Moisture Volatile Ash” standard; ultimate analysis was obtained by Leco CHN/S analyzer. Higher heating values (HHV) are measured using an IKA C2000 oxygen bomb calorimeter.
2.2. Thermal analysis The co-combustion was performed on the HH and LC blends prepared at weight ratio of 100:0, 90:10, 80:20, 70:30, 60:40, 50:50 and 0:100 and named as LC, 90LC10HH, 80LC20HH, 70LC30HH, 60LC40HH, 50LC50HH and HH. The combustion behavior of individual and blended fuel samples was analyzed by Simultaneous DTA-TG Analyzer (Shimadzu, Japan). The temperature was initially raised to 110 °C at a heating rate of 5 °C min1. Subsequently, the temperature was raised to 900 °C at different heating rates (5, 10, 20 and 50 °C min1). The air flow was kept constant at 20 mL min1 in all experiments. The mass loss percentage (TG) and mass loss rate (DTG) of the samples were recorded continuously under non-isothermal conditions. The analysis of each sample was repeated to confirm reproducibility.
2.3. ANN model development In this study a multi-layer perception (MLP)-based feed forward neural network model Levenberg–Marquardt (LM) back propagation algorithm was used to predict TGA curves obtained from combustion of individual or blended samples. This model is widely used due to its efficiency and simplicity (Vani et al., 2015). The MLP network consists of input, hidden and output layers (Fig. 1). All layers are connected by weights and biases. By adjusting these, non-linear functions can be modeled. The input layer included three neurons: heating rate, blending ratio and temperature dependent weight loss ratio. Hidden layers were used to carry out complex and non-linear functions on the network. The model the number of hidden layers was adjusted to achieve the maximum R2 value. The number of hidden layers, number of neurons in the hidden layers, training epochs, and activation functions were selected by trial and error. Training and testing performances of the network were determined with the root mean square error (RMSE; Eq. (1)), mean absolute error (MAE; Eq. (2)), and mean bias error (MBE; Eq. (3)) analysis methods. The higher values of R2 and lower values of both RMSE and MAE mean a better performance of the developed ANN (Sßahin, 2015).
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N u1 X RMSE ¼ t ðHi Hi;model Þ2 N i¼1
ð1Þ
MAE ¼
N 1X jHi Hi;model j N i¼1
ð2Þ
MBE ¼
N 1X ðHi Hi;model Þ N i¼1
ð3Þ
The ANN modeling was performed using the ANN toolbox in the Matlab mathematical software.
44
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
Fig. 1. Neural network diagram used for predicting co-combustion behavior of different HH–LC blends.
3. Results and discussion 3.1. Proximate and ultimate analysis The proximate analysis, calorific values and ultimate analysis of the samples are shown in Table 1. The coal and biomass samples have different properties due to differences in their composition. Being a biomass, HH have high volatile content, which is known to be removed at a lower temperature than that of coal (Kirtania and Bhattacharya, 2013). As seen from Table 1, coal has about 52% volatile content, while that of HH is 74%. LC has higher carbon content compared to HH and therefore lower H/C and O/C ratios due to the low hydrogen and oxygen content, respectively. The HHV of HH was lower than that of LC; this result is in line with the C contents of the samples. The higher H and O levels in HH lowers the HHV, however it also increases the thermal efficiency during combustion. Furthermore, if used in a co-combustion process, lower S and N content of HH can reduce the SOx and NOx emissions caused by pure coal combustion. The ash contents of LC and HH were respectively 31.50% and 6.01%. Therefore, blends of highly reactive fuels, such as HH, with coal could be beneficial for reduction of the wastes of combustion process. Higher ash
Fig. 3. TG–DTG profiles for the blending ratio of 50% at different heating rates.
Table 2 The comparisons performance of different ANN structure.
Fig. 2. Comparison of the TG–DTG curves of HH, LC and their blends at the heating rate of 20 °C min1; (a) TG curve (b) DTG curve.
Model
Inputs
Network topology
RMSE
MBE
MAE
R2
ANN1 ANN2 ANN3 ANN4 ANN5 ANN6 ANN7 ANN8 ANN9 ANN10 ANN11 ANN12 ANN13 ANN14 ANN15 ANN16 ANN17 ANN18 ANN19 ANN20 ANN21
Biomass–coal ratio. Heating rate. Temperature
3*3*1 3*5*1 3*7*1 3*10*1 3*15*1 3*17*1 3*20*1 5*3*1 5*5*1 5*7*1 5*10*1 5*15*1 5*17*1 5*20*1 3*1 5*1 7*1 10*1 15*1 17*1 20*1
1.985 1.298 1.904 1.337 3.504 1.145 2.493 1.969 1.894 1.033 3.370 0.624 0.770 0.806 14.369 1.902 2.308 1.944 2.333 1.157 0.850
1.560 1.001 1.517 0.995 2.491 0.847 1.839 1.540 1.492 0.812 2.521 0.484 0.585 0.563 10.184 1.491 1.851 1.513 1.818 0.863 0.693
0.00048 0.00012 0.00031 0.00047 0.00114 0.00034 0.00241 0.00087 0.00099 0.00102 0.00212 0.00028 0.00038 0.01385 2.23986 0.00056 0.00068 0.00181 0.00023 0.00010 0.00006
0.9940 0.9975 0.9945 0.9973 0.9814 0.9980 0.9906 0.9941 0.9946 0.9984 0.9828 0.9994 0.9991 0.9990 0.6955 0.9945 0.9920 0.9943 0.9918 0.9980 0.9989
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
45
Fig. 4. Schematic diagram of the proposed ANN model.
Fig. 5. The comparisons performance of ANN12 based model (a) training (b) testing.
amounts can also hinder the efficiency of the combustion process by forming aggregates and preventing effective mass and heat transfer. Therefore, study on these effects is very important for co-firing applications. 3.2. Combustion process of the pure materials The TG and DTG profiles, obtained at a heating rate of 20 °C min1, of individual combustion of HH and LC are shown in Fig. 2. It can be noted that there are significant differences in the thermal behavior of the samples. In particular, as a lignocellulosic biomass, three main stages were observed during the thermal degradation of HH. The first stage that extends up to the temperature of about 150 °C is related with the removal of moisture and release of low molecular weight compounds. In the range 150–360 °C a significant peak, corresponding to the decomposition of cellulose and hemicellulose, was detected. This second stage
represents the release of volatiles and their ignition, leading to char formation. Lignin degradation occurred in the third stage, in a wider temperature interval (390–550 °C). Lignin is the most thermally stable component of the biomass and it is the main component responsible for formation of biomass char (Lopez-González et al., 2015; Gil et al., 2015). In particular, the small peak of the weight loss rate confirms the low reactivity of lignin, since the reactivity is inversely proportional to the temperature corresponding to the peak (DTG peak) and directly proportional to the peak height. In previous works it has been reported that the decomposition of the main components of biomass (hemicellulose, cellulose, and lignin) occurs in the range of 225–325 °C, 305–375 °C, and 250–500 °C, respectively (Buratti et al., 2015). Our findings are line with the reported results. Lignite coal (LC) shows an initial mass loss in the temperature range 25–205 °C due to moisture evaporation. Further, the thermal degradation of the coal shows only one major peak over a wide
46
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
Fig. 6. Comparisons of testing performance of ANN12 Model for 80% blend ratio.
temperature range (205–585 °C) that corresponds to the release of carbon-containing volatile matter from coal (Buratti et al., 2015). The temperature corresponding to the maximum in the DTG curve is approximately 408 °C. The maximum weight loss rate is much lower than the one of HH due to lower volatile matter content of LC.
due to the non-isothermal behavior of the samples. Therefore, the heating rate enhances the heat transfer from the surface to the core of the particle.
3.3. Co-combustion of LC–HH blends
The training performance results of ANNs with different number of hidden layers and different number of neurons trained with LM are shown in Table 2. According to the training performance assessment, the LM algorithm provided a better result for ANN12 model. The total iterations in this model were set to 2000 and the performance goal was 105. The input layer and hidden layer have nonlinear activation neurons (tansig) and output layer has linear neuron (purelin) in network topology. The structure of ANN12 model is shown in Fig. 4. The ANN12 structure included 3 input parameters, namely heating rate, blending ratio and temperature, 5 hidden neurons in 15 hidden layer and 1 output parameter-mass loss percent. The output was assessed to predict the TG curves for different blend ratios. From Table 2, it can be seen that with the network structures with 15 hidden layers, the ANN12 network provided a better performance than others with respect to the training and testing results. As seen from Fig. 5, the coefficients of determination of this network were 0.9995 and 0.9993 for training and testing, respectively. The RMSE value of ANN12 model was 0.624. As a results the ANN12 network is more suitable for estimating the mass loss with temperature for the remaining blend ratios. Predictions of the mass loss percent with the ANN12 network as a function of the measured values are shown in Fig. 6. The agreement between the experimental data and predicted values justified the accuracy of the ANN calculation.
The TG and DTG profiles of the blends at the heating rate of 20 °C min1 are also shown in Fig. 2. In the DTG curves of the blends two oxidation stages can be seen—the first one between 220 and 310 °C is due to degradation of hemicellulose and cellulose, and a broader peak in the second stage, up to about 600 °C, which is mainly due to the release of volatile matter in the coal that burns slowly over the whole temperature range, together with the char. All the curves of the blends lie between the ones of the pure samples. 3.4. Influence of heating rate Typical DTG curves at different heating rates (5, 10, 15, and 50 °C min1) for the blending ratio of 50% are shown in Fig. 3. As the heating rate increased, the DTG curves were shifted to higher temperatures, in accord with previously reported results (Buratti et al., 2015; Goldfarb and Ceylan, 2015; Jiang et al., 2015). Both stages of the combustion process are developed in a wider temperature range, when increasing the heating rate. Therefore, heating rate has an important role in devolatilization and combustion processes of the blends, because it can accelerate devolatilization rate, however, the complete oxidation of char requires more time. The fact that the maximum mass loss rate shifts toward higher temperatures proves that the combustion intensity can be improved by increasing heating rate. Higher heating rate implies shorter time to achieve certain ambient temperature, which leads to relatively larger temperature difference between the ambient and the inner core of a particle,
3.5. Application of ANN to co-combustion of HH–LC
4. Conclusion The characteristics of the co-combustion process of HH and LC was investigated using TG analysis. Furthermore, ANN was applied to estimate TG curves of blends. The agreement between
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
experimental data and predicted values justified the accuracy of ANN calculation. The results obtained indicate that the dynamic ANN model developed in this study is capable of direct prediction of the co-combustion behavior, provided that the heating rate, blend ratio, and temperature and known. Thus, ANN can provide useful information for the design and optimization of co-combustion systems by predicting the thermal behavior of any blend ratio of interest. Acknowledgement Authors want to thank to Associate Professor. Pance Naumov (NYU Abu Dhabi) for his helps for reviewing the manuscript. References Ata, R., 2015. Artificial neural networks applications in wind energy systems: a review. Renewable Sustainable Energy Rev. 49, 534–562. Authier, O., Moullec, Y.L., 2013. Coal chemical-looping combustion for electricity generation: investigation for a 250 MW Power plant. Energy Proc. 37, 588–597. _ Kaygusuz, K., 2015. A perspective for potential and Bilgen, S., Kelesß, S., Sarıkaya, I., technology of bioenergy in Turkey: present case and future view. Renewable Sustainable Energy Rev. 48, 228–239. Buratti, C., Barbanera, M., Bartocci, P., Fantozzi, F., 2015. Thermogravimetric analysis of the behavior of sub-bituminous coal and cellulosic ethanol residue during co-combustion. Bioresour. Technol. 186, 154–162. Carsky, M., Kuwornoo, D.K., 2001. Neural network modelling of coal pyrolysis. Fuel 80, 1021–1027. Celaya, A.M., Lade, A.T., Goldfarb, J.L., 2015. Co-combustion of brewer’s spent grains and Illinois No. 6 coal: impact of blend ratio on pyrolysis and oxidation behavior. Fuel Process. Technol. 129, 39–51. Ceylan, S., Topçu, Y., 2014. Pyrolysis kinetics of hazelnut husk using thermogravimetric analysis. Bioresour. Technol. 156, 182–188. Chen, J., Mu, L., Cai, J., Yin, H., Song, X., Li, A., 2015. Thermal characteristics and kinetics of refining and chemicals wastewater, lignite and their blends during combustion. Energy Convers. Manage. 100, 201–211. Çepeliog˘ullar, Ö., Pütün, A.E., 2013. Thermal and kinetic behaviors of biomass and plastic wastes in co-pyrolysis. Energy Convers. Manage. 75, 263–270.
47
Gajic, D., Savic-Gajic, I., Savic, I., Georgieva, O., Gennaro, S., 2015. Modelling of electrical energy consumption in an electric arc furnace using artificial neural networks. Energy. http://dx.doi.org/10.1016/j.energy.2015.07.068. Gil, M.V., García, R., Pevida, C., Rubiera, F., 2015. Grindability and combustion behavior of coal and torrefied biomass blends. Bioresour. Technol. 191, 205–212. Goldfarb, J.L., Ceylan, S., 2015. Second-generation sustainability: application of the distributed activation energy model to the pyrolysis of locally sourced biomass– coal blends for use in co-firing scenarios. Fuel 160, 297–308. Hu, S., Ma, X., Lin, Y., Yu, Z., Fang, S., 2015. Thermogravimetric analysis of the cocombustion of paper mill sludge and municipal solid waste. Energy Convers. Manage. 99, 112–118. Jiang, L., Xiao, H.H., He, J.J., Suna, Q., Gong, L., Sun, J.H., 2015. Application of genetic algorithm to pyrolysis of typical polymers. Fuel Process. Technol. 138, 48–55. Kirtania, K., Bhattacharya, S., 2013. Pyrolysis kinetics and reactivity of algae–coal blends. Biomass Bioenergy 55, 291–298. Lopez-González, D., Avalos-Ramirez, A., Giroir-Fendler, A., Godbout, S., FernandezLopez, M., Sanchez-Silva, L., Valverde, J.L., 2015. Combustion kinetic study of woody and herbaceous crops by thermal analysis coupled to mass spectrometry. Energy. http://dx.doi.org/10.1016/j.energy.2015.06.134. Liu, X., Chen, M., Wei, Y., 2015. Kinetics based on two-stage scheme for cocombustion of herbaceous biomass and bituminous coal. Fuel 143, 577–585. Mikulandric, R., Loncar, Drazen., Böhning, D., Böhme, R., Beckmann, M., 2014. Artificial neural network modelling approach for a biomass gasification process in fixed bed gasifiers. Energy Convers. Manage. 87, 1210–1223. Peng, X., Ma, X., Xu, Z., 2015. Thermogravimetric analysis of co-combustion between microalgae and textile dyeing sludge. Bioresour. Technol. 180, 288–295. Sahu, S.G., Chakraborty, N., Sarkar, P., 2014. Coal–biomass co-combustion: an overview. Renewable Sustainable Energy Rev. 39, 575–586. Sßahin, F., 2015. Effects of engine parameters on ionization current and modeling of excess air coefficient by artificial neural network. Appl. Therm. Eng. 90, 94–101. Toptas, A., Yildirim, Y., Duman, G., Yanik, J., 2015. Combustion behavior of different kinds of torrefied biomass and their blends with lignite. Bioresour. Technol. 177, 328–336. Vani, S., Sukumaran, R.K., Savithri, S., 2015. Prediction of sugar yields during hydrolysis of lignocellulosic biomass using artificial neural network modeling. Bioresour. Technol. 188, 128–135. Yuanyuan, Z., Yanxia, G., Fangqin, C., Kezhou, Y., Yan, C., 2015. Investigation of combustion characteristics and kinetics of coal gangue with different feedstock properties by thermogravimetric analysis. Thermochim. Acta 614, 137–148.
Contents lists available at ScienceDirect
Bioresource Technology journal homepage: www.elsevier.com/locate/biortech
Application of artificial neural networks to co-combustion of hazelnut husk–lignite coal blends Zeynep Yıldız, Harun Uzun, Selim Ceylan ⇑, Yıldıray Topcu Ondokuz Mayıs University Chemical Engineering Department, 55139 Samsun, Turkey
h i g h l i g h t s First study on application of ANN model to co-combustion process. The coefficient of determination for the developed model was 0.9995. ANN model predicted TG curves for co-combustion of blends effectively.
a r t i c l e
i n f o
Article history: Received 17 August 2015 Accepted 27 September 2015 Available online 8 October 2015 Keywords: Co-combustion Hazelnut husk Lignite coal TGA Artificial neural network
a b s t r a c t The artificial neural network (ANN) theory is applied to thermal data obtained by non-isothermal thermogravimetric analysis (TGA) from room temperature to 1000 °C at different heating rates in air to study co-combustion of hazelnut husk (HH)–lignite coal (LC) blends of various composition. The heating rate, blend ratio and temperature were used in the ANN analysis to predict the TG curves of the blends as parameters that affect the thermal behavior during combustion. The ANN model provides a good prediction of the TG curves for co-combustion with a coefficient of determination for the developed model of 0.9995. The agreement between the experimental data and the predicted values substantiated the accuracy of the ANN calculation. Ó 2015 Elsevier Ltd. All rights reserved.
1. Introduction Coal remains the most important primary energy source in power generation, amounting for 36% of the total globallygenerated power. Turkey has rich lignite reserves that are used in thermal power stations. However, coal combustion also is one of the main causes of global warming. Coal combustion is globally responsible for 43% of CO2 emissions from fossil-fuel combustion, with 28% being emitted from coal-fired power plants (Authier and Moullec, 2013; Buratti et al., 2015). Co-combustion of coal with biomass provides a cost-effective means to reduce CO2 and other greenhouse gas emissions, fossil fuel consumption, and the amount of landfilled organic matter. Biomass combustion is considered carbon-neutral because the CO2 released during combustion is in principle withdrawn from the atmosphere by photosynthesis during the growth of plants (Hu et al., 2015; Buratti et al., 2015; Peng et al., 2015; Gil et al., 2015). Furthermore, biomass is often used to achieve better control of the burning process. In co-combustion, biomass increases the volatile matter ⇑ Corresponding author. E-mail address: [email protected] (S. Ceylan). http://dx.doi.org/10.1016/j.biortech.2015.09.114 0960-8524/Ó 2015 Elsevier Ltd. All rights reserved.
content, providing a more stable flame; moreover, many sources have lower ash content than coal, thus reducing waste and increasing efficiency. Although the higher proportion of oxygen and hydrogen to carbon atoms in biomass does result in a lower heating value for biomass compared to coal, the higher oxygen content of biomass also tends to give it a higher reactivity than coal, and thus a lower activation energy barrier for devolatilization and oxidation (Celaya et al., 2015). Turkey holds a vast potential of biomass energy resources, and especially of agricultural wastes. Although this potential is available, most of the waste biomass species could not be evaluated properly for energetic purposes (Bilgen et al., 2015). Turkey is the world’s leading hazelnut producer with production of approximately 500,000 tons per annum (Çepeliogullar and Pütün, 2013); an exorbitant amount of hazelnut husk (HH) is generated during processing of the hazelnut fruit as agricultural waste biomass. These residues have no alternative uses, and because of their large storage volumes they also cause a disposal problem (Ceylan and Topçu, 2014). To explore the possibilities of alleviating some of these technological burdens, in this study the potentials of HH as blending biomass material are studied. A knowledge of the co-combustion
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
characteristics is crucial to achieve effective design and operation of the co-combustion process. Thermogravimetry has been widely used to characterize the thermal process, and to describe the profiles and the effective kinetic parameters for various decomposition reactions in order to predict their thermal behavior (Peng et al., 2015; Toptas et al., 2015). Numerous studies dealing with the application of TG analysis to evaluate combustion behavior of coal and biomass alone and their blends in co-firing application have been reported prior to this study (Sahu et al., 2014; Liu et al., 2015; Chen et al., 2015; Yuanyuan et al., 2015; Goldfarb and Ceylan, 2015; Gil et al., 2015). In these studies, mathematical models were used to explain or predict the process behavior, and to analyze the effects of different process variables on the process performance. Several commonly used methods for kinetic parameter calculation have been established, notably the Kissinger–Aka hira–Sunose (KAS) method, the Flynn–Wall–Ozawa (FWO) method, and the Friedman method, among others. However, an inherent deficiency of these methods is their inapplicability to confirm three kinetic parameters simultaneously. Also, the complexity of co-combustion phenomena has handicapped the development of a unifying model suitable for design purposes. These methods fail particularly in modeling secondary and some complex multistep reactions (Carsky and Kuwornoo, 2001). Artificial intelligence systems, such as neural networks, are widely accepted as a technology that can be applied to non-linear problems, and (once trained) can be used for prediction and generalization at high speed. Artificial neural networks (ANN) models use a non-physical modelling approach which correlates the input and output data to form a process prediction model. The ANN models have proven their potential in prediction of process parameters in energy-related processes (Mikulandric et al., 2014; Ata, 2015; Gajic et al., 2015; Sßahin, 2015) but their potential to predict parameters of a biomass co-combustion process is yet to be explored. In this work the combustion behavior of HH, LC and their blends was investigated through TGA at different heating rates and blend ratios. Optimizing the operating conditions for such fuel blending requires a greater understanding of the thermal characteristics coal–biomass blends; only accurate prediction of thermal behavior can provide a better design and operation of industrial systems (Celaya et al., 2015). Thus, the aim of this work is to use ANN to obtain accurate thermal behavior of multi-step co-combustion reactions of HH–LC blends. 2. Methods 2.1. Proximate and ultimate analysis Coal samples were supplied from Elbistan thermal power plant (Kahramanmarasß, Turkey). HH was provided by a local company. Table 1 Proximate and ultimate analysis of HH and LC.
a db
Proximate analysisdb, %
Lignite coal (LC)
Hazelnut husk (HH)
Volatile matters Ash Fixed carbona
52.10 31.50 16.40
74.24 6.01 19.75
Ultimate analysisdb, % C H N S Oa H/C O/C Calorific value/MJ kg1
61.79 5.94 1.82 4.31 26.14 0.10 1.15 20.20
43.09 5.81 1.30 0.142 49.66 0.42 0.13 17.14
Calculated by difference. Dry basis.
43
Both HH and coal samples were dried in an oven at 105 °C for 8 h to remove water. The samples were ground and sieved to particle size of 63–125 lm to avoid the heat- and mass-transfer limitations during experiments. The materials were characterized by using conventional proximate and t ultimate analyses, and the results are reported in Table 1. Proximate analysis is performed by a Simultaneous DTA-TG Analyzer (Shimadzu, Japan) based on the ASTM D 5142-04 ‘‘Moisture Volatile Ash” standard; ultimate analysis was obtained by Leco CHN/S analyzer. Higher heating values (HHV) are measured using an IKA C2000 oxygen bomb calorimeter.
2.2. Thermal analysis The co-combustion was performed on the HH and LC blends prepared at weight ratio of 100:0, 90:10, 80:20, 70:30, 60:40, 50:50 and 0:100 and named as LC, 90LC10HH, 80LC20HH, 70LC30HH, 60LC40HH, 50LC50HH and HH. The combustion behavior of individual and blended fuel samples was analyzed by Simultaneous DTA-TG Analyzer (Shimadzu, Japan). The temperature was initially raised to 110 °C at a heating rate of 5 °C min1. Subsequently, the temperature was raised to 900 °C at different heating rates (5, 10, 20 and 50 °C min1). The air flow was kept constant at 20 mL min1 in all experiments. The mass loss percentage (TG) and mass loss rate (DTG) of the samples were recorded continuously under non-isothermal conditions. The analysis of each sample was repeated to confirm reproducibility.
2.3. ANN model development In this study a multi-layer perception (MLP)-based feed forward neural network model Levenberg–Marquardt (LM) back propagation algorithm was used to predict TGA curves obtained from combustion of individual or blended samples. This model is widely used due to its efficiency and simplicity (Vani et al., 2015). The MLP network consists of input, hidden and output layers (Fig. 1). All layers are connected by weights and biases. By adjusting these, non-linear functions can be modeled. The input layer included three neurons: heating rate, blending ratio and temperature dependent weight loss ratio. Hidden layers were used to carry out complex and non-linear functions on the network. The model the number of hidden layers was adjusted to achieve the maximum R2 value. The number of hidden layers, number of neurons in the hidden layers, training epochs, and activation functions were selected by trial and error. Training and testing performances of the network were determined with the root mean square error (RMSE; Eq. (1)), mean absolute error (MAE; Eq. (2)), and mean bias error (MBE; Eq. (3)) analysis methods. The higher values of R2 and lower values of both RMSE and MAE mean a better performance of the developed ANN (Sßahin, 2015).
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N u1 X RMSE ¼ t ðHi Hi;model Þ2 N i¼1
ð1Þ
MAE ¼
N 1X jHi Hi;model j N i¼1
ð2Þ
MBE ¼
N 1X ðHi Hi;model Þ N i¼1
ð3Þ
The ANN modeling was performed using the ANN toolbox in the Matlab mathematical software.
44
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
Fig. 1. Neural network diagram used for predicting co-combustion behavior of different HH–LC blends.
3. Results and discussion 3.1. Proximate and ultimate analysis The proximate analysis, calorific values and ultimate analysis of the samples are shown in Table 1. The coal and biomass samples have different properties due to differences in their composition. Being a biomass, HH have high volatile content, which is known to be removed at a lower temperature than that of coal (Kirtania and Bhattacharya, 2013). As seen from Table 1, coal has about 52% volatile content, while that of HH is 74%. LC has higher carbon content compared to HH and therefore lower H/C and O/C ratios due to the low hydrogen and oxygen content, respectively. The HHV of HH was lower than that of LC; this result is in line with the C contents of the samples. The higher H and O levels in HH lowers the HHV, however it also increases the thermal efficiency during combustion. Furthermore, if used in a co-combustion process, lower S and N content of HH can reduce the SOx and NOx emissions caused by pure coal combustion. The ash contents of LC and HH were respectively 31.50% and 6.01%. Therefore, blends of highly reactive fuels, such as HH, with coal could be beneficial for reduction of the wastes of combustion process. Higher ash
Fig. 3. TG–DTG profiles for the blending ratio of 50% at different heating rates.
Table 2 The comparisons performance of different ANN structure.
Fig. 2. Comparison of the TG–DTG curves of HH, LC and their blends at the heating rate of 20 °C min1; (a) TG curve (b) DTG curve.
Model
Inputs
Network topology
RMSE
MBE
MAE
R2
ANN1 ANN2 ANN3 ANN4 ANN5 ANN6 ANN7 ANN8 ANN9 ANN10 ANN11 ANN12 ANN13 ANN14 ANN15 ANN16 ANN17 ANN18 ANN19 ANN20 ANN21
Biomass–coal ratio. Heating rate. Temperature
3*3*1 3*5*1 3*7*1 3*10*1 3*15*1 3*17*1 3*20*1 5*3*1 5*5*1 5*7*1 5*10*1 5*15*1 5*17*1 5*20*1 3*1 5*1 7*1 10*1 15*1 17*1 20*1
1.985 1.298 1.904 1.337 3.504 1.145 2.493 1.969 1.894 1.033 3.370 0.624 0.770 0.806 14.369 1.902 2.308 1.944 2.333 1.157 0.850
1.560 1.001 1.517 0.995 2.491 0.847 1.839 1.540 1.492 0.812 2.521 0.484 0.585 0.563 10.184 1.491 1.851 1.513 1.818 0.863 0.693
0.00048 0.00012 0.00031 0.00047 0.00114 0.00034 0.00241 0.00087 0.00099 0.00102 0.00212 0.00028 0.00038 0.01385 2.23986 0.00056 0.00068 0.00181 0.00023 0.00010 0.00006
0.9940 0.9975 0.9945 0.9973 0.9814 0.9980 0.9906 0.9941 0.9946 0.9984 0.9828 0.9994 0.9991 0.9990 0.6955 0.9945 0.9920 0.9943 0.9918 0.9980 0.9989
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
45
Fig. 4. Schematic diagram of the proposed ANN model.
Fig. 5. The comparisons performance of ANN12 based model (a) training (b) testing.
amounts can also hinder the efficiency of the combustion process by forming aggregates and preventing effective mass and heat transfer. Therefore, study on these effects is very important for co-firing applications. 3.2. Combustion process of the pure materials The TG and DTG profiles, obtained at a heating rate of 20 °C min1, of individual combustion of HH and LC are shown in Fig. 2. It can be noted that there are significant differences in the thermal behavior of the samples. In particular, as a lignocellulosic biomass, three main stages were observed during the thermal degradation of HH. The first stage that extends up to the temperature of about 150 °C is related with the removal of moisture and release of low molecular weight compounds. In the range 150–360 °C a significant peak, corresponding to the decomposition of cellulose and hemicellulose, was detected. This second stage
represents the release of volatiles and their ignition, leading to char formation. Lignin degradation occurred in the third stage, in a wider temperature interval (390–550 °C). Lignin is the most thermally stable component of the biomass and it is the main component responsible for formation of biomass char (Lopez-González et al., 2015; Gil et al., 2015). In particular, the small peak of the weight loss rate confirms the low reactivity of lignin, since the reactivity is inversely proportional to the temperature corresponding to the peak (DTG peak) and directly proportional to the peak height. In previous works it has been reported that the decomposition of the main components of biomass (hemicellulose, cellulose, and lignin) occurs in the range of 225–325 °C, 305–375 °C, and 250–500 °C, respectively (Buratti et al., 2015). Our findings are line with the reported results. Lignite coal (LC) shows an initial mass loss in the temperature range 25–205 °C due to moisture evaporation. Further, the thermal degradation of the coal shows only one major peak over a wide
46
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
Fig. 6. Comparisons of testing performance of ANN12 Model for 80% blend ratio.
temperature range (205–585 °C) that corresponds to the release of carbon-containing volatile matter from coal (Buratti et al., 2015). The temperature corresponding to the maximum in the DTG curve is approximately 408 °C. The maximum weight loss rate is much lower than the one of HH due to lower volatile matter content of LC.
due to the non-isothermal behavior of the samples. Therefore, the heating rate enhances the heat transfer from the surface to the core of the particle.
3.3. Co-combustion of LC–HH blends
The training performance results of ANNs with different number of hidden layers and different number of neurons trained with LM are shown in Table 2. According to the training performance assessment, the LM algorithm provided a better result for ANN12 model. The total iterations in this model were set to 2000 and the performance goal was 105. The input layer and hidden layer have nonlinear activation neurons (tansig) and output layer has linear neuron (purelin) in network topology. The structure of ANN12 model is shown in Fig. 4. The ANN12 structure included 3 input parameters, namely heating rate, blending ratio and temperature, 5 hidden neurons in 15 hidden layer and 1 output parameter-mass loss percent. The output was assessed to predict the TG curves for different blend ratios. From Table 2, it can be seen that with the network structures with 15 hidden layers, the ANN12 network provided a better performance than others with respect to the training and testing results. As seen from Fig. 5, the coefficients of determination of this network were 0.9995 and 0.9993 for training and testing, respectively. The RMSE value of ANN12 model was 0.624. As a results the ANN12 network is more suitable for estimating the mass loss with temperature for the remaining blend ratios. Predictions of the mass loss percent with the ANN12 network as a function of the measured values are shown in Fig. 6. The agreement between the experimental data and predicted values justified the accuracy of the ANN calculation.
The TG and DTG profiles of the blends at the heating rate of 20 °C min1 are also shown in Fig. 2. In the DTG curves of the blends two oxidation stages can be seen—the first one between 220 and 310 °C is due to degradation of hemicellulose and cellulose, and a broader peak in the second stage, up to about 600 °C, which is mainly due to the release of volatile matter in the coal that burns slowly over the whole temperature range, together with the char. All the curves of the blends lie between the ones of the pure samples. 3.4. Influence of heating rate Typical DTG curves at different heating rates (5, 10, 15, and 50 °C min1) for the blending ratio of 50% are shown in Fig. 3. As the heating rate increased, the DTG curves were shifted to higher temperatures, in accord with previously reported results (Buratti et al., 2015; Goldfarb and Ceylan, 2015; Jiang et al., 2015). Both stages of the combustion process are developed in a wider temperature range, when increasing the heating rate. Therefore, heating rate has an important role in devolatilization and combustion processes of the blends, because it can accelerate devolatilization rate, however, the complete oxidation of char requires more time. The fact that the maximum mass loss rate shifts toward higher temperatures proves that the combustion intensity can be improved by increasing heating rate. Higher heating rate implies shorter time to achieve certain ambient temperature, which leads to relatively larger temperature difference between the ambient and the inner core of a particle,
3.5. Application of ANN to co-combustion of HH–LC
4. Conclusion The characteristics of the co-combustion process of HH and LC was investigated using TG analysis. Furthermore, ANN was applied to estimate TG curves of blends. The agreement between
Z. Yıldız et al. / Bioresource Technology 200 (2016) 42–47
experimental data and predicted values justified the accuracy of ANN calculation. The results obtained indicate that the dynamic ANN model developed in this study is capable of direct prediction of the co-combustion behavior, provided that the heating rate, blend ratio, and temperature and known. Thus, ANN can provide useful information for the design and optimization of co-combustion systems by predicting the thermal behavior of any blend ratio of interest. Acknowledgement Authors want to thank to Associate Professor. Pance Naumov (NYU Abu Dhabi) for his helps for reviewing the manuscript. References Ata, R., 2015. Artificial neural networks applications in wind energy systems: a review. Renewable Sustainable Energy Rev. 49, 534–562. Authier, O., Moullec, Y.L., 2013. Coal chemical-looping combustion for electricity generation: investigation for a 250 MW Power plant. Energy Proc. 37, 588–597. _ Kaygusuz, K., 2015. A perspective for potential and Bilgen, S., Kelesß, S., Sarıkaya, I., technology of bioenergy in Turkey: present case and future view. Renewable Sustainable Energy Rev. 48, 228–239. Buratti, C., Barbanera, M., Bartocci, P., Fantozzi, F., 2015. Thermogravimetric analysis of the behavior of sub-bituminous coal and cellulosic ethanol residue during co-combustion. Bioresour. Technol. 186, 154–162. Carsky, M., Kuwornoo, D.K., 2001. Neural network modelling of coal pyrolysis. Fuel 80, 1021–1027. Celaya, A.M., Lade, A.T., Goldfarb, J.L., 2015. Co-combustion of brewer’s spent grains and Illinois No. 6 coal: impact of blend ratio on pyrolysis and oxidation behavior. Fuel Process. Technol. 129, 39–51. Ceylan, S., Topçu, Y., 2014. Pyrolysis kinetics of hazelnut husk using thermogravimetric analysis. Bioresour. Technol. 156, 182–188. Chen, J., Mu, L., Cai, J., Yin, H., Song, X., Li, A., 2015. Thermal characteristics and kinetics of refining and chemicals wastewater, lignite and their blends during combustion. Energy Convers. Manage. 100, 201–211. Çepeliog˘ullar, Ö., Pütün, A.E., 2013. Thermal and kinetic behaviors of biomass and plastic wastes in co-pyrolysis. Energy Convers. Manage. 75, 263–270.
47
Gajic, D., Savic-Gajic, I., Savic, I., Georgieva, O., Gennaro, S., 2015. Modelling of electrical energy consumption in an electric arc furnace using artificial neural networks. Energy. http://dx.doi.org/10.1016/j.energy.2015.07.068. Gil, M.V., García, R., Pevida, C., Rubiera, F., 2015. Grindability and combustion behavior of coal and torrefied biomass blends. Bioresour. Technol. 191, 205–212. Goldfarb, J.L., Ceylan, S., 2015. Second-generation sustainability: application of the distributed activation energy model to the pyrolysis of locally sourced biomass– coal blends for use in co-firing scenarios. Fuel 160, 297–308. Hu, S., Ma, X., Lin, Y., Yu, Z., Fang, S., 2015. Thermogravimetric analysis of the cocombustion of paper mill sludge and municipal solid waste. Energy Convers. Manage. 99, 112–118. Jiang, L., Xiao, H.H., He, J.J., Suna, Q., Gong, L., Sun, J.H., 2015. Application of genetic algorithm to pyrolysis of typical polymers. Fuel Process. Technol. 138, 48–55. Kirtania, K., Bhattacharya, S., 2013. Pyrolysis kinetics and reactivity of algae–coal blends. Biomass Bioenergy 55, 291–298. Lopez-González, D., Avalos-Ramirez, A., Giroir-Fendler, A., Godbout, S., FernandezLopez, M., Sanchez-Silva, L., Valverde, J.L., 2015. Combustion kinetic study of woody and herbaceous crops by thermal analysis coupled to mass spectrometry. Energy. http://dx.doi.org/10.1016/j.energy.2015.06.134. Liu, X., Chen, M., Wei, Y., 2015. Kinetics based on two-stage scheme for cocombustion of herbaceous biomass and bituminous coal. Fuel 143, 577–585. Mikulandric, R., Loncar, Drazen., Böhning, D., Böhme, R., Beckmann, M., 2014. Artificial neural network modelling approach for a biomass gasification process in fixed bed gasifiers. Energy Convers. Manage. 87, 1210–1223. Peng, X., Ma, X., Xu, Z., 2015. Thermogravimetric analysis of co-combustion between microalgae and textile dyeing sludge. Bioresour. Technol. 180, 288–295. Sahu, S.G., Chakraborty, N., Sarkar, P., 2014. Coal–biomass co-combustion: an overview. Renewable Sustainable Energy Rev. 39, 575–586. Sßahin, F., 2015. Effects of engine parameters on ionization current and modeling of excess air coefficient by artificial neural network. Appl. Therm. Eng. 90, 94–101. Toptas, A., Yildirim, Y., Duman, G., Yanik, J., 2015. Combustion behavior of different kinds of torrefied biomass and their blends with lignite. Bioresour. Technol. 177, 328–336. Vani, S., Sukumaran, R.K., Savithri, S., 2015. Prediction of sugar yields during hydrolysis of lignocellulosic biomass using artificial neural network modeling. Bioresour. Technol. 188, 128–135. Yuanyuan, Z., Yanxia, G., Fangqin, C., Kezhou, Y., Yan, C., 2015. Investigation of combustion characteristics and kinetics of coal gangue with different feedstock properties by thermogravimetric analysis. Thermochim. Acta 614, 137–148.