Rapid Determination of the Gross Calorific Value of Coal Using Laser

Publication Date (Web): March 3, 2017. Copyright © 2017 American Chemical Society. *Telephone: +86-13925150807. Fax: +86-20-87110613. E-mail: ...
0 downloads 0 Views 1MB Size
Subscriber access provided by University of Newcastle, Australia

Article

Rapid Determination of Gross Calorific Value of Coal using LIBS Coupled with Artificial Neural Networks (ANN) and Genetic Algorithm (GA) Zhimin Lu, Juehui Mo, Shunchun Yao, Jingbo Zhao, and Jidong Lu Energy Fuels, Just Accepted Manuscript • DOI: 10.1021/acs.energyfuels.7b00025 • Publication Date (Web): 03 Mar 2017 Downloaded from http://pubs.acs.org on March 5, 2017

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Energy & Fuels is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Rapid Determination of Gross Calorific Value of Coal using LIBS Coupled with Artificial Neural Networks (ANN) and Genetic Algorithm (GA) Zhimin Lu, †‡§ Juehui Mo, †‡§ Shunchun Yao, *,†‡§ Jingbo Zhao†‡§, Jidong Lu†‡§

†School of Electric Power, South China University of Technology, Guangzhou, Guangdong, 510640, China ‡Guangdong Province Key Laboratory of Efficient and Clean Energy Utilization, Guangzhou, Guangdong, 510640, China §Guangdong Province Engineering Research Center of High Efficient and Low Pollution Energy Conversion, Guangzhou, Guangdong, 510640, China

ABSTRACT: On-line measurement for the gross calorific value (GCV) of coal is important in coal utilization industry. This paper proposed a rapid GCV determination method that combined laser-induced breakdown spectroscopy (LIBS) technique with artificial neural networks (ANN) and Genetic Algorithm (GA). Input variables were selected according to the physical mechanism and mathematical significance to improve the prediction of the ANN. GA was applied to

ACS Paragon Plus Environment

1

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 30

determine an optimal architecture for the network instead of trial and error method. As a result, the mean standard deviation (MSD) of the GCV for 4 prediction set samples is 0.38 MJ/kg in fifty trials (repetitions of training the ANN with the same input data but different random initial weights and biases), proving that the ANN model is able to provide a high modeling repeatability in the GCV analysis. The mean absolute error (MAE) of the GCV for the prediction set is 0.39 MJ/kg. The result meets the requirements (0.8 MJ/kg) for coal on-line analyses using neutron activation method in the Chinese national standard (GB/T 29161-2012).

KEYWORDS: gross calorific value of coal, laser-induced breakdown spectroscopy, rapid determination, artificial neural network, genetic algorithms

1. Introduction Gross calorific value (GCV) is an important indicator for the quality of coal. Accurate and timely GCV analysis of coal is an important step in mine processing and power plant operations for the purpose of operation optimization and emission control. Traditionally, a standard coal sample is obtained by sampling system and then GCV is determined using bomb calorimeter in laboratory.1 The classical laboratory method is laborious and time-consuming and the results fluctuate greatly with the sampling procedure. It prompts research efforts on the development of advanced techniques for online or rapid determination of the coal GCV. The commercially available online coal analyzer is based on the gamma-ray technique.2,3 The contents of organic elements (e.g C, H and O) in coal are determined and then used to calculate the GCV by an empirical formula, which is often valid within a specific range of coal. Moreover, the gamma-ray based analyzer is very bulky and has strict safety regulatory requirements for the neutron source

ACS Paragon Plus Environment

2

Page 3 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

presents potential health hazards. Therefore, there has been continuous interests in developing new technologies for online coal analysis. Among these new technologies, laser-induced breakdown spectroscopy (LIBS) has raised the greatest interests for its unique advantages, e.g., little or no sample preparation requirement, onsite and multi-elemental analysis. Thus, LIBS has been regarded as a “future superstar” in the application of online elemental analysis4 and coal analysis.5,6 To date, many efforts have been made to improve the performance of coal analysis by LIBS, such as analysis of the inorganic elements,7-10 organic elements11-16 and proximate analysis.17-22 However, limited publications can be found in gross calorific value (GCV) determination. Wang et al.23 put forward a multivariable calibrated model of coal GCV, to correct the spectrum deviation caused by the self-absorption effect, elements interference and matrix effect. Yuan et al.24 and Hou et al.15 proposed a nonlinear multivariate-dominant-factor based partial least square (PLS) model, taking into account the physical mechanisms. The results of GCV prediction demonstrated an overall improvement over conventional PLS model. Zhang et al.22 utilized a support vector regression (SVR) method combined with principal component analysis (PCA), which enabled a significant improvement in cross-validation accuracy on GCV measurement. The composition of coal is extremely complex and the chemical and physical properties vary greatly among different types of coal. There are many empirical formulas that use the elemental analysis results to calculate the coal GCV.25 The linear method using LIBS to extract the elementals concentrations so as to calculate the GCV can be found in Ref. 6. But this method is subject to strict application conditions and the error may be magnified in the two estimation processes. It is expected that a non-linear relationship will present between the LIBS spectral intensity and the coal GCV analysis, leading to the increasing complexity of data processing. The

ACS Paragon Plus Environment

3

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 30

non-linear multivariate analysis methods based on PLS and SVR have been employed to improve the performance of coal analysis by LIBS.14,22 Indeed, when LIBS was used to analyze different types of coal, there would be no fixed correlation between spectral intensity and GCV. A non-linear multivariate analysis method, with strong fault tolerance and robustness, is expected to be a better solution in the quantitative analysis of GCV. Inspired by biological neural networks, artificial neural network (ANN) is used to estimate or approximate generally unknown functions that can depend on a large number of inputs. Ghosh et al.

26,27

have proposed the application of the ANN to predict coal proximate parameters and

useful heat values from well logs and got promising results. In fact, estimating the coal GCV by ANN is proved to be an effective method using the proximate analysis or ultimate analysis as input variables. 28-31 Therefore, the non-linear relationship between the LIBS spectra intensity and the GCV seems to be a perfect application for ANN model, which can be used as an arbitrary function approximation mechanism that ‘learns’ from observed data. Several research groups have proposed the use of ANN for LIBS data processing, showing promising results that ANN can weaken the matrix effect

32

and self-absorption effect

33

to some extent. The

combination of ANN and LIBS has been successfully applied to the elemental analysis of oil,34 soil,35,36 steel

37,38

and bronze alloy.39,40 However, the network architecture has a huge influence

on the trade-off between predictive accuracy on the training dataset and generalization capability of the model on untrained data. The network architectures are usually determined by the empirical formula or trial and error method, which are time-consuming when dealing with the complex non-linear relationship. The aim of this work is to provide a rapid coal GCV determination method coupled with LIBS and ANN. Especially, the coal GCV was directly extract from LIBS spectral data rather than

ACS Paragon Plus Environment

4

Page 5 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

from the concentration of combustible elements. Genetic algorithm (GA) is used to generate an optimal architecture for the ANN model. To reduce the randomness of the ANN model, we establish the GCV analysis model with a certain number of networks and average the outputs of all networks as the final predicted value for an unknown sample. 2. Methodology The proposed gross calorific value (GCV) determination method combines the LIBS technique with the GA-ANN hybrid algorithm (Figure.1), including three steps: i) Data preprocessing, determine the input and output variables for ANN, ii) GA optimization process, determine an optimal architecture for ANN, and iii) Prediction process, determine an appropriate number of networks used to average the outputs.

2. 1.Data preprocessing The most widely used neural network, multi-layer perceptron artificial neural network (MLPANN), is chosen as the basis analysis model. Back propagation (BP) algorithm is used to search the optimal weights and biases for the ANN model. As a black-box model, it’s hard to study the contribution of different variables to the outputs in the ANN model. The selection of the input variables has a significant impact on the reliability and robustness of the ANN model.41 The strategy of using all the LIBS spectral lines as the inputs isn’t adopted, because the introduction of irrelative variables will result in increasing complexity and function overfitting of the ANN model.42 Input variables are selected according to their physical mechanism and mathematical significances. Firstly, in order to reduce the measurement uncertainties caused by laser pulse energy fluctuation, the intensities of analyte lines are processed by internal standard

ACS Paragon Plus Environment

5

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 30

method.10,44 Secondly, the analyte lines of carbon and hydrogen are selected as the input variables because gross calorific value (GCV) is strongly dependent on the combustion of organic compositions. Thirdly, the correlation coefficients (R) between the GCV and the intensities of different analyte lines are calculated respectively. The analyte line with a higher R value than a default value is selected as the input variable. 2.2. GA optimization process The network architecture also has a significant impact on the convergence speed and generalization capability of ANN model. Although a typical three-layer ANN model can approximate any continuous function with arbitrary precision, it should be emphasized that how to design the network architecture efficiently remains a debatable problem when it comes to a practical application. Too few neurons in the hidden layer can lead to lower predictive accuracy which means that the network cannot capture the non-linear relationships in the training dataset. On the other hand, too many neurons can also result in over-fitting of the training dataset and increasing computational time. Therefore, the GA optimization algorithm is applied to determine an optimal network architecture for GCV analysis. The network architecture includes the number of hidden layers, the activation function and number of neurons in each hidden layer. The activation function defines the output of a neuron by giving it a set of inputs, including two nonlinear forms: tansig and logsig. The information of network architecture is encoded in a 12-byte string by binary-code43:i) the first five and the sixth to tenth characters, respectively, decoded into integer between 0 and 31, represent the neurons number in the first and second hidden layer, ii) the eleventh and twelfth characters, respectively, represent the activation function in the first and second hidden layer, 0 stands for tansig while 1 stands for logsig.

ACS Paragon Plus Environment

6

Page 7 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

The second part of Figure.1 shows the combination of GA and ANN. The evolution algorithm starts from a population of randomly generated individuals. Then, the individuals are decoded into the architectures of different networks. Each network is trained by the training set. After training, the network is used to predict the calorific values in the validation set. The root mean square error (RMSE) between the measured values and predicted values of the validation set is used as objective function value to evaluate the generalization ability of each network. A series of optimization operation of selection, crossover, mutation and recombination is adopted to process the current population to form a new generation. As a result, the individuals having the smaller RMSE are preserved so that each generation is gradually optimized. The iterative process will continue until the maximum number of generations has been produced. 2.3. Prediction process. The initial weights and biases of ANN are randomly assigned, so the results of random trials (same input data but different initial random weights and biases) may trap at different local optima, which means that we will get different outputs for the same training set. To reduce the randomness, the calorific value analysis model is established with a certain number of networks. For predicting an unknown sample, the outputs of all networks are averaged as the final predicted value. 3. Experimental 28 bituminous coal samples were used as experimental samples. All samples were pulverized to less than 200 µm in diameter and air dried. Their GCV were measured by bomb calorimeter 1 and listed in Table 1. These 28 samples were divided into three subsets: i) the training set has 20 samples including the ones with the maximum and minimum GCV, to properly train the ANN model, ii) the validation set with 4 samples, to determine an optimal ANN architecture in the

ACS Paragon Plus Environment

7

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 8 of 30

optimization process, and iii) the prediction set with 4 samples is considered as unknown samples which were not presented in the training and optimization process, to evaluate the predictive accuracy of the model. The measurement was done in a LIBS experimental system with a particle flow system to simulate the on-site measurements. The details of the experimental system were described in our previous work.19 A Q-switch neodymium-doped yttrium aluminum garnet (Nd : YAG) laser (Elite-200; Beamtech Optronics, Beijing) operating at 1064 nm with a 4 ns pulse duration was adopted as the ablation source, and a dual-channel spectrometer (AvaSpec-2048; Avantes, Holland) was used to detect the spectrum with a resolution of 0.2–0.3 nm in the spectral coverage of 240–400 nm (channel 1) and 580–790 nm (channel 2). Each sample was excited for 1000 times. Since all of the samples get through the measuring point in the form of particle flow, so each shot was on a fresh spot. The average spectral data of 1000 shots was obtained to reduce unexpected measurement fluctuations.

4. Results and discussion 4.1 Data preprocessing results The coal LIBS spectra indicates its chemical compositions, which in turn determines its GCV. The GCV mainly results from the combustion of organic compounds like fixed carbon and volatile matter. C,H and O are the main constituent elements of combustibles. The ash content also has direct impacts on the GCV. The GCV generally decreases with the increasing of ash content. The compositions in coal ash mainly include the oxides of Si, Al, Fe, Ca, Mg, K, Na and Ti. Therefore, 11 commonly used analyte lines of the elements mentioned above were preliminary selected as input variables, including C 247.86 nm, H 657.25 nm, O 777.33 nm, Si

ACS Paragon Plus Environment

8

Page 9 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

288.15 nm, Al 308.19 nm, Fe 374.85 nm, Ca 393.36 nm, Mg 285.21 nm, K 766.49 nm, Na 616.26 nm and Ti 334.91 nm. Carbon is the main component element and homogeneously distributed in the coal sample. The feasibility of using the spectral line of carbon at 247.68 nm as a reference line for the determination of inorganic elements in coal has been verified.10 Silicon is the most abundant element in coal ash, whose spectral line at 288.15 nm is easy to identify in the spectra with little interference.44 Thus, C 247.86 nm and Si 288.15 nm were chosen as the reference lines. All of the 11 analyte lines were respectively normalized by these two reference lines. Then the correlation coefficient (R) between the GCV and the intensity of different analyte lines were calculated. As shown in Table 2, the R values for the same analyte line were different with different reference lines. The analyte line with an absolute R value higher than 0.7 was considered to be highly correlated with the GCV. The organic elements lines (C 247.86 nm, H 657.25 nm and O 777.33 nm) normalized by Si 288.15 nm showed high positive correlations with the GCV, which meant that the GCV would increase with the increasing of the intensity of these 3 normalized analyte lines. The inorganic elements lines (Al 308.19 nm and Ti 334.9 nm) normalized by C 247.86 nm show high negative correlations with the GCV, which meant that the GCV would decrease with the increasing of the intensity of these 2 normalized analyte lines. Therefore, these 5 normalized analyte lines were selected as the input variables, which made sense in both physical mechanism and mathematical significance. To compare the influences of different input variables on the model, three sets of input variables were used to train a typical three-layer ANN model for 10 times. The ANN model had one hidden layer with 11 neurons and the activation function was tansig (Table 3). These three

ACS Paragon Plus Environment

9

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 30

set input variables are, respectively, a) all of the 11 analyte lines normalized by Si 288.15 nm, b) all of the 11 analyte lines normalized by C 247.86 nm, c) the 5 normalized analyte lines selected by the method described above. Standard deviation (SD) of the network outputs for 10 trials was used to evaluate the reproducibility in modeling of the ANN model. As shown in Figure. 2, the mean SD (MSD) of the validation set predicted values for the three set of inputs were 2.26 MJ/kg, 0.61 MJ/kg and 0.23 MJ/kg, respectively. The results showed that the reproducibility in modeling of ANN was improved after using the data preprocessing and variables selection method we proposed.

4.2 Architecture optimized results Figure 3 recorded the average and minimum objective function value of each generation in the optimization process. The average and minimum RMSE of validation set gradually decreased along with the iteration. After 50 iterations, we selected the individual with the smallest RMSE to decode into the information of the optimized ANN architecture (Table 3). The optimized ANN model had one hidden layers with 23 neurons and the activation function is logsig. In order to evaluate whether the optimization is effective to the ANN model, the typical ANN model and optimized ANN model were trained for 10 times. As shown in Figure. 4, the MSD of the validation set for the typical ANN model is 0.22 MJ/kg, while that for the optimized ANN model is 0.07 MJ/kg. The RMSE of the validation set for the typical ANN model is 1.28 MJ/kg, while that for the optimized ANN model is 1.16 MJ/kg. The optimized ANN model provides a higher reproducibility and measurement accuracy. 4.3 Appropriate networks number

ACS Paragon Plus Environment

10

Page 11 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

To determine how many networks are needed to average the outputs, we carried out 10 groups of computations. Each group of computation repeatedly trained the optimized ANN model for 150 times. After each time of training, we got a new network (same architecture but different weights and biases). The outputs of this new network were averaged with the outputs of the previous networks. Finally, 150 sets average outputs calculated with different network numbers were recorded (Eq. (1)).  =

∑   

(1)

Here, where the  are the outputs of the i-th network, the  are the average outputs of the first n networks. The range of n is from 1 to 150. The validation set RMSEs of the 150 sets average outputs was calculated for the 10 groups of computations. As shown in Figure.5, when the networks number was less than 25, the RMSEs varied widely from different group of computation and the RMSEs for each group of computation fluctuated within a relatively large range. When the networks number approached 50, RMSEs fluctuated within a small range and become stable. Therefore, the average outputs of 50 networks can reduce the impact of randomly assigning initial weights and biases of the ANN model to some extent. Specifically, to construct a GCV analysis model, we would train an optimized ANN for 50 times and saved the weights and biases after each training so as to get 50 trained ANN models. Therefore, we established the GCV analysis model with 50 trained ANN models. The regression graphs (Figure. 6) illustrated that the R2 for the training set is 0.964 and the RMSE is 0.14 MJ/kg, showing a high goodness of fit between the measured and predicted GCV of the model. We used the proposed model to predict the GCV of 4 prediction set samples. The prediction MSD is 0.38 MJ/kg for the outputs of the 50 networks. The average prediction outputs

ACS Paragon Plus Environment

11

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 12 of 30

were shown in Table 4. The prediction RMSE is 0.27 MJ/kg. The maximum and minimum absolute error is 0.97 MJ/kg and -0.11 MJ/kg, respectively, with the mean absolute error (MAE) of these 4 samples equaling to 0.39 MJ/kg. Compared to the dynamic precision requirement (0.8 MJ/kg) in a Chinese national standard 45 for neutron activation coal on-line analyzer, most of the predictions can meet the demand. The method we proposed, combining 50 GA-optimized ANN models to establish the GCV analysis model, can achieve acceptable reproducibility and measurement accuracy.

5. Conclusion In this paper, we have proposed a rapid gross calorific value determination method based on the LIBS spectra data of coal. ANN is used to establish the quantitative analysis model to weaken the matrix effects and the non-linear behavior in the LIBS. By the procedure of data preprocessing and GA optimization, we overcame the random effect of ANN to some extent. Three sets of coal samples are respectively used to train, validate and test the generalization ability of the model. The results show that the measurement accuracy is acceptable. With more and more samples adding to the training set, the reliability and robustness of the gross calorific value analysis model would be improved.

AUTHOR INFORMATION Corresponding Author *Email: [email protected]. Phone: +86-13925150807. Fax: +86-20-87110613

ACS Paragon Plus Environment

12

Page 13 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

ACKNOWLEDGEMENTS The work was supported by National Natural Science Funds of China (51206055, 51676073, 51476061),

Pearl

River

S&T

Nova

Program

of

Guangzhou

(2014J2200054),

Guangdong Province train high-level personnel special support program (2014TQ01N334) and Science and Technology Project of Guangdong Province (2015A020215005). REFERENCE 1.

DIN 5900-2. Determining the gross calorific value of solid and liquid fuels using the isoperibol or static jacket calorimeter and calculation of net calorific value; 2003.

2.

Lim C S, Abernethy D A. On-line coal analysis using fast neutron-induced gamma-rays[J]. Applied radiation and isotopes, 2005, 63(5): 697-704.

3.

Sowerby B D. Nuclear techniques for the on-line bulk analysis of carbon in coal-fired power stations[J]. Applied Radiation and Isotopes, 2009, 67(9): 1638-1643.

4.

Winefordner J D, Gornushkin I B, Correll T, et al. Comparing several atomic spectrometric methods to the super stars: special emphasis on laser induced breakdown spectrometry, LIBS, a future super star[J]. Journal of Analytical Atomic Spectrometry, 2004, 19(9): 10611083.

5.

Gaft M, Sapir-Sofer I, Modiano H, et al. Laser induced breakdown spectroscopy for bulk minerals online analyses[J]. Spectrochimica Acta Part B: Atomic Spectroscopy, 2007, 62(12): 1496-1503.

6.

Yin W, Zhang L, Dong L, et al. Design of a laser-induced breakdown spectroscopy system for on-line quality analysis of pulverized coal in power plants[J]. Applied spectroscopy, 2009, 63(8): 865-872.

7.

Wallis F J, Chadwick B L, Morrison R J S. Analysis of lignite using laser-induced

ACS Paragon Plus Environment

13

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 14 of 30

breakdown spectroscopy[J]. Applied Spectroscopy, 2000, 54(8): 1231-1235. 8.

Chadwick B L, Body D. Development and commercial evaluation of laser-induced breakdown spectroscopy chemical analysis technology in the coal power generation industry[J]. Applied spectroscopy, 2002, 56(1): 70-74.

9.

Mateo M P, Nicolas G, Yanez A. Characterization of inorganic species in coal by laserinduced breakdown spectroscopy using UV and IR radiations[J]. Applied Surface Science, 2007, 254(4): 868-872.

10. Yao S, Xu J, Bai K, et al. Improved Measurement Performance of Inorganic Elements in Coal by Laser-Induced Breakdown Spectroscopy Coupled with Internal Standardization[J]. Plasma Science and Technology, 2015 (11): 938-943. 11. Li X, Mao X, Wang Z. Quantitative Analysis of Carbon Content in Bituminous Coal by Laser-Induced Breakdown Spectroscopy Using UV Laser Radiation[J]. Plasma Science and Technology, 2015, (11): 928-932. 12. Feng J, Wang Z, West L, et al. A PLS model based on dominant factor for coal analysis using laser-induced breakdown spectroscopy[J]. Analytical and bioanalytical chemistry, 2011, 400(10): 3261-3271. 13. Zhang L, Dong L, Dou H, et al. Laser-induced breakdown spectroscopy for determination of the organic oxygen content in anthracite coal under atmospheric conditions[J]. Applied spectroscopy, 2008, 62(4): 458-463. 14. Feng J, Wang Z, Li L, et al. A Nonlinearized Multivariate Dominant Factor–Based Partial Least Squares (PLS) Model for Coal Analysis by Using Laser-Induced Breakdown Spectroscopy[J]. Applied spectroscopy, 2013, 67(3): 291-310. 15. Hou Z, Wang Z, Yuan T, et al. A hybrid quantification model and its application for coal

ACS Paragon Plus Environment

14

Page 15 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

analysis using laser induced breakdown spectroscopy[J]. Journal of Analytical Atomic Spectrometry, 2016, 31(3): 722-736. 16. M. Gaft, L. Nagli, I. Fasaki, M. Kompitsas, G. Wilsch. Laser-induced breakdown spectroscopy

for

on-line

sulfur

analyses

of

minerals

in

ambient

conditions.

SpectrochimicaActa Part B, 64(10),2009:1098-1104 17. Shunchun Yao, Jidong Lu, Meirong Dong, Kai Chen, Junyan Li, Jun Li. Extracting Coal Ash Content from Laser-Induced Breakdown Spectroscopy Spectra by Multivariate Analysis. Applied Spectroscopy, 2011,65, 1197-1201. 18. Meirong Dong, Jidong Lu, Shunchun Yao, Jun Li, Junyan Li, ZimingZhong, Weiye. Application of LIBS for Direct Determination of Volatile Matter Content in Coal. Journal of Analytical Atomic Spectrometry, 2011, 26(11), 2183-2188. 19. Zheng Jianping, Lu Jidong, Zhang Bo, Dong Meirong, Yao Shunchun, Lu Weiye, Dong Xuan. Experimental Study of Laser-Induced Breakdown Spectroscopy (LIBS) for Direct Analysis of Coal Particle Flow. Applied Spectroscopy, 2014(68):672-679. 20. Haider A, Rony M A, Abedin K M. Determination of the Ash Content of Coal without Ashing: A Simple Technique Using Laser-Induced Breakdown Spectroscopy[J]. Energy & Fuels, 2013, 27(7): 3725-3729. 21. X. Wang, L. Zhang, J.J. Fan, Y.F. Li, Y. Gong, L. Dong, W.G. Ma, W.B. Ying, S.T. Jia. Parameters Optimization of Laser-Induced Breakdown Spectroscopy Experimental Setup for the Case with Beam Expander[J]. Plasma Science and Technology, 914, 2015 (11): 914-918. 22. Zhang L, Gong Y, Li Y, et al. Development of a coal quality analyzer for application to power plants based on laser-induced breakdown spectroscopy[J]. SpectrochimicaActa Part B: Atomic Spectroscopy, 2015, 113: 167-173.

ACS Paragon Plus Environment

15

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 16 of 30

23. Di W, Ji-dong L, Mei-rong D, et al. A New Calibrated Model of Coal Calorific Value Detection with LIBS[J]. Spectroscopy and Spectral Analysis, 2016, 36(8): 2607-2612. 24. Yuan T, Wang Z, Lui S L, et al. Coal property analysis using laser-induced breakdown spectroscopy[J]. Journal of Analytical Atomic Spectrometry, 2013, 28(7): 1045-1053. 25. Wenmin Chen. Calorific Value Calculation of Various Kinds of Coal in China by Using Industrial Analysis and Elemental Analysis [J]. Coal Conversion, 1981, 1. 26. Ghosh S, Chatterjee R, Shanker P. Estimation of ash, moisture content and detection of coal lithofacies from well logs using regression and artificial neural network modelling[J]. Fuel, 2016, 177: 279-287. 27. Ghosh S, Chatterjee R, Shanker P. Prediction of Coal Proximate Parameters and Useful Heat Value of Coal from Well Logs of the Bishrampur Coalfield, India, Using Regression and Artificial Neural Network Modeling[J]. Energy & Fuels, 2016, 30(9): 7055-7064. 28. Kavšek D, Bednárová A, Biro M, et al. Characterization of Slovenian coal and estimation of coal heating value based on proximate analysis using regression and artificial neural networks[J]. Open Chemistry, 2013, 11(9): 1481-1491. 29. Patel S U, Kumar B J, Badhe Y P, et al. Estimation of gross calorific value of coals using artificial neural networks[J]. Fuel, 2007, 86(3): 334-344. 30. Erik N Y, Yilmaz I. On the use of conventional and Soft Computing Models for prediction of gross calorific value (GCV) of coal[J]. International Journal of Coal Preparation and Utilization, 2011, 31(1): 32-59. 31. Yilmaz I, Erik N Y, Kaynar O. Different types of learning algorithms of artificial neural network (ANN) models for prediction of gross calorific value (GCV) of coals[J]. Scientific Research and Essays, 2010, 5(16): 2242-2249.

ACS Paragon Plus Environment

16

Page 17 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

32. PM. Mukhono, KH. Angeyo, et al. Laser induced breakdown spectroscopy and characterization of environmental matrices utilizing multivariate chemometrics [J]. SpectrochimicaActa Part B: Atomic Spectroscopy, 2013, 87:81-85 33. F.

Rezaei,

P.

karimi,

SH.

Tavassoli.

Effect

of

self-absorption

correction

on LIBS measurements by training curve and artificial neural network [J]. Applied Physics B -Lasers andOptics, 2014, 114 (4):591-600 34. Tarazona J L, Guerrero J, Cabanzo R, et al. Construction of a predictive model for concentration of nickel and vanadium in vacuum residues of crude oils using artificial neural networks and LIBS[J]. Applied optics, 2012, 51(7); B108-B114. 35. J. El Haddad, D. Bruyère, A. Ismaël, et al. Application of a series of artificial neural networks to on-site quantitative analysis of lead into real soil samples by laser induced breakdown spectroscopy [J]. SpectrochimicaActa Part B: Atomic Spectroscopy, 2014, 97:57-64 36. J. El Haddad, M. Villot-Kadri, A. Ismaël, et al. Artificial neural network for on-site quantitative analysis of soils using laser induced breakdown spectroscopy [J]. SpectrochimicaActa Part B: Atomic Spectroscopy, 2013, 79-80:51-57 37. Li K, Guo L, Li C. Analytical-performance improvement of laser-induced breakdown spectroscopy for steel using multi-spectral-line training with an artificial neural network [J]. Journal of Analytical Atomic Spectrometry, 2015, 30 (7):1623-1628 38. Lorenzetti G, Legnaioli S, Grifoni E, et al. Laser-based continuous monitoring and resolution of steel grades in sequence casting machines[J]. SpectrochimicaActa Part B: Atomic Spectroscopy, 2015, 112: 1-5. 39. D’Andrea E, Pagnotta S, Grifoni E, et al. An artificial neural network approach to laser-

ACS Paragon Plus Environment

17

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 30

induced breakdown spectroscopy quantitative analysis [J]. SpectrochimicaActa Part B: Atomic Spectroscopy, 2014, 99: 52-58. 40. E. D’Andrea, S. Pagnotta, E. Grifoni, et al. A hybrid training-free/artificial neural networks approach to the quantitative analysis of LIBS spectra [J]. Applied Physics B -Lasers and Optics, 2015, 118 (3):353-600 41. He X, P. Niyogi. Loaclity preserving projections [C]. Proceedings of Neural Information Processing System.2003:153-160 42. Marcos-Martinez D, Ayala J A, Izquierdo-Hornillos R C, et al. Identification and discrimination of bacterial strains by laser induced breakdown spectroscopy and neural networks[J]. Talanta, 2011, 84(3):730-737. 43. YANG J, WENG S, ZHAO H, et al. An Optimized BP Network Model Using Genetic Algorithm for Predicting the Ignition-Stability Index of Pulverized Coal[J]. Journal of Power Engineering, 2006, 1: 017. 44. Liu Y, Lu J, Li P, et al. Determination of carbon content in pulverized coal with laserinduced breakdown spectroscopy by internal standard method[J]. Proc. CSEE, 2009, 29(5). 45. GB/T 29161-2012, Specifications of on-line analyzer for coal based on neutron activation analysis [S]

ACS Paragon Plus Environment

18

Page 19 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Captions Figure.1. Schematic diagram of rapid GCV determination method. Figure.2. The validation set SD (10 trials) of the ANN outputs using different input variables. Figure.3. Variation of the average and minimum objective value (RMSE) of the validation set for each generation Figure.4. The performance of the ANN models. a. Typical ANN; b. Optimized ANN Figure.5. The validation set RMSEs for 10 groups of computation. Figure.6. Regression graphs of the GCV analysis model that combining 50 trained optimized ANN models.

ACS Paragon Plus Environment

19

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 20 of 30

Table 1. The measured GCV of 28 coal samples (MJ/kg)

Grouping

Sample calorific Sample calorific Sample calorific Sample calorific value value value value ID ID ID ID T1

23.29

T2

22.54

T3

22.18

T4

29.15

T5

21.16

T6

27.24

T7

22.58

T8

23.92

T9

22.53

T10

22.08

T11

24.71

T12

23.66

T13

22.70

T14

23.10

T15

25.17

T16

21.58

T17

23.55

T18

23.42

T19

23.98

T20

20.86

validation set

V1

25.30

V2

24.72

V3

22.41

V4

23.94

prediction set

P1

22.73

P2

23.32

P3

26.30

P4

21.27

training set

* Here, , and are, respectively, the measured calorific values of the training set, the validation set and the prediction set associated with the sample i.

ACS Paragon Plus Environment

20

Page 21 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Table 2. The correlation coefficients (R) between the GCV and analyte lines normalized by C 247.86 nm or Si 288.15 nm R

R

(C247.869)

(Si 288.150)

C 247.869

——

0.76

H 657.251

-0.46

0.75

O 777.335

-0.34

0.75

Si288.150

-0.66

——

Al 308.197

-0.75

-0.17

Fe374.856

-0.12

0.64

Ca393.368

0.22

0.69

Mg285.215

-0.35

0.57

K766.459

-0.47

0.21

Na616.26

0.25

0.69

Ti 334.906

-0.73

-0.2

Reference line Analyte line

ACS Paragon Plus Environment

21

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 22 of 30

Table 3. The architecture of typical ANN and optimized ANN Typical ANN

Optimized ANN

Neurons number

activation function

1

11

2

——

No. of hidden layer

activation function

No. of

Neurons

hidden layer

number

tansig

1

23

logsig

——

2

——

——

ACS Paragon Plus Environment

22

Page 23 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Table 4. GCV (MJ/kg) predictive results for the 4 prediction set samples Sample Reference Predicted Absolute ID value value error P1

22.73

22.93

0.20

P2

23.32

24.29

0.97

P3

26.30

26.19

-0.11

P5

21.27

21.53

0.26

ACS Paragon Plus Environment

23

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure.1. Schematic diagram of rapid calorific value determination method. The proposed calorific value d 105x86mm (300 x 300 DPI)

ACS Paragon Plus Environment

Page 24 of 30

Page 25 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Figure.2. The validation set SD (10 trials) of the ANN outputs using different input variables. As shown in Figure. 2, the pre 124x100mm (300 x 300 DPI)

ACS Paragon Plus Environment

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure.3. Variation of the average and minimum objective value (RMSE) of the validation set for each generation Figure 3 recorded the average 124x100mm (300 x 300 DPI)

ACS Paragon Plus Environment

Page 26 of 30

Page 27 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Figure.4. The performance of the ANN models. a. Typical ANN; b. Optimized ANN As shown in Figure. 4, the MS 126x103mm (300 x 300 DPI)

ACS Paragon Plus Environment

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure.4. The performance of the ANN models. a. Typical ANN; b. Optimized ANN As shown in Figure. 4 130x110mm (300 x 300 DPI)

ACS Paragon Plus Environment

Page 28 of 30

Page 29 of 30

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Energy & Fuels

Figure.5. The validation set RMSEs for 10 groups of computation. As shown in Figure.5, when the 128x107mm (300 x 300 DPI)

ACS Paragon Plus Environment

Energy & Fuels

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Figure.6. Regression graphs of the GCV analysis model that combining 50 trained optimized ANN models. The regression graphs (Figure. 125x102mm (300 x 300 DPI)

ACS Paragon Plus Environment

Page 30 of 30