Académique Documents
Professionnel Documents
Culture Documents
http://www.elsevier.com/locate/biombioe
Universitat Rovira i Virgili, Dept. Eng. Meca`nica, Av. Pasos Catalans 26, 43007 Tarragona, Spain
Universidad Autonoma del Estado de Morelos, Centro de Investigacion en Ingeniera y Ciencias Aplicadas (CIICAp), Av. Universidad No.
1001 Col. Chamilpa, 62209 Cuernavaca, Mexico
b
article info
abstract
Article history:
Artificial neural networks (ANNs) have been applied for modeling biomass gasification
process in fluidized bed reactors. Two architectures of ANNs models are presented; one for
circulating fluidized bed gasifiers (CFB) and the other for bubbling fluidized bed gasifiers
16 November 2012
(BFB). Both models determine the producer gas composition (CO, CO2, H2, CH4) and gas
yield. Published experimental data from other authors has been used to train the ANNs.
The obtained results show that the percentage composition of the main four gas species in
producer gas (CO, CO2, H2, CH4) and producer gas yield for a biomass fluidized bed gasifier
Keywords:
can be successfully predicted by applying neural networks. ANNs models use in the input
Biomass
layer the biomass composition and few operating parameters, two neurons in the hidden
Gasification
layer and the backpropagation algorithm. The results obtained by these ANNs show high
agreement with published experimental data used R2 > 0.98. Furthermore a sensitivity
Simulation
analysis has been applied in each ANN model showing that all studied input variables are
Fluidized bed
important.
2012 Elsevier Ltd. All rights reserved.
1.
Introduction
280
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
Input layer
( i)
i=1
Ash
Hidden layer
Output layer
Weights
IWj,i
Moisture
4e33.4
35e220
476.6e529.9
383.8e435.5
54.3e78.6
0.19e0.64
701e861
LWk,j
C
j=1
ER
k=1
1.72e3.30
3.00e7.30
1.20e4.60
13.94e18.30
6.90e21.40
Output
i=7
Tg
j=2
b1j
b2k
biases
5.5e11.0
62.8e250
458.9e505.4
411.1e471.8
56.4e70.8
0.19e0.47
700e900
0e0.04
1.17e3.42
4.97e26.17
2.40e6.07
9.82e18.60
10e29.47
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
Fig. 3 e Comparison of the experimental results with the results calculated by ANN for CFB gasifiers.
281
282
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
2.
Methods
2.1.
Table 3 e Weights and biases of the ANNs designed for the four major gas species of producer gas (CO, CO2, H2, CH4) and
producer gas yield for ANN model for CFB gasifiers.
CO
IWi,j
3.2006
1.1408
LW1,j
5.4159
0.0722
1.8333
CO2
IWi,j
1.7859
9.8078
LW1,j
4.6685
3.1087
9.1839
CH4
IWi,j
1.1889
1.4276
LW1,j
1.2490
0.5638
0.3493
10.0337
3.7749
0.4085
4.0413
12.4537
5.0279
2.3948
b1j
7.6506
2.1235
3.9112
1.7017
2.6406
2.4613
3.9629
6.3563
H2
IWi,j
1.7403
16.8436
LW1,j
3.0137
Producer gas yield
IWi,j
6.8841
6.4443
5.7169
1.6951
LW1,j
0.5083
0.3425
3.4878
24.2709
0.8185
1.2959
1.9792
2.3434
1.5775
0.9014
3.8072
b2
12.9445
0.9632
0.8495
1.5078
9.5719
b2
10.2800
0.6350
4.0890
2.6040
1.9226
b2
8.5215
2.7315
2.5402
4.0765
6.6673
7.8066
20.5250
b2
7.8781
0.7735
18.7020
3.7339
3.6455
9.9848
19.7080
b2
2.4497
1.4279
6.0075
5.3061
0.1148
b1j
3.6063
0.0732
1.9819
15.3984
4.4029
3.0161
b1j
3.8959
5.2290
0.2984
0.8706
3.3632
3.6059
b1j
3.9094
17.6810
1.3813
3.2875
b1j
14.6688
0.8342
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
The selection of an appropriate set of variables for inclusion as inputs to the model is a crucial step in model development, as the performance of the final model is heavily
dependent on the input variables used.
In this study, an extensive literature review was done to
obtain experimental data that could be used to develop the
ANNs models. Due to the different properties and behavior of
different biomasses, and to have more homogeneous data,
only experimental data for wood gasification in atmospheric
pressure and inert bed reactors was considered. Data for circulating fluidized bed ANN model was obtained for air gasification of wood from Li et al. [7] (cypress, hemlock and mixed
sawdust) and van der Drift et al. [8] (mixed wood). Published
experimental data for bubbling fluidized bed reactors was
found in the studies of Narvaez et al. [9] (pine sawdust),
Campoy [10] (pellets), Kaewluan and Pipatmanomai [11]
(rubber wood chips) and Lv et al. [12] (pine sawdust) for air
and airesteam gasification.
In both ANNs models, the data sets containing the information (the values of input and output variables) of different
biomass gasification tests are small. The data sets for CFB and
BFB gasifiers contain the results of 18 and 36 tests, respectively.
Due to the small size of the data sets and after some preliminary validation tests and results from the literature [5,6];
283
2.2.
Fig. 4 e Relative impact (%) of input variables on the different outputs for the four main producer gas components and
producer gas yield of the ANN model for CFB gasifiers.
284
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
in the natural neurons. An ANN is composed of a large number of highly interconnected processing elements (neurons or
nodes) working in unison to solve specific problems. The
neurons are grouped into distinct layers and interconnected
according to a given architecture. Each layer has a weight
matrix, a bias vector and an output vector.
In this study, two ANNs models were developed in the
Matlab environment using the Neural Network Toolbox [15].
Fig. 1 and Fig. 2 illustrate the architecture of the models for
CFB and BFB gasifiers, respectively. Since there is no explicit
rule to determine either the number of neurons in the hidden
layer or the number of hidden layers, the trial and error
method was applied to find the best solution by minimizing
the Root Mean Square Error (RMSE). In this step of training,
a study was carried out to determine the number of neurons in
hidden layer which was considered to one and two neurons
for both ANNs models. The best obtained results (data not
show) were considering two neurons in hidden layer (see Figs.
1 and 2).
The ANNs models proposed in the present study consist in:
3.
3.1.
Proposed ANN model for circulating fluidized bed
gasifiers
Five neural networks with seven inputs, two neurons in the
hidden layer and one output each, was found to be efficient in
predicting producer gas composition as well as gas yield for
CFB gasifiers.
Experimental and simulated values for CO, CO2, H2, CH4,
and gas yield were compared satisfactorily through a linear
regression model ( y a$x b) for each. The obtained
regression coefficients (R2) are presented in Fig. 3. It can be
seen how all R2 values are higher than 0.99 except for the case
of H2 composition that it is 0.98.
According to Verma et al. [16] and El Hamzaoui et al. [17] to
satisfy the statistical test of intercept and slope; the interval
between the highest and lowest values of the intercept must
contain zero and the interval between the highest and lowest
values of the slope must contain one. The proposed ANNs
passed the test with 99.8% of confidence level. This test
guarantees that whole ANN model, containing five ANNs, has
a satisfactory level of confidence.
Table 3 gives the obtained parameters (IWj,i, LW1,j, b1j, b2)
of the best fit for 2 neurons in the hidden layer for each of the
five ANN developed in the CFB model. These parameters were
used in the proposed model to simulate the output values. In
consequence, the proposed ANN model follows Eq. (2):
- CFB gasifier model: five ANNs, one for each output (CO, CO2,
H2, CH4 and gas yield). Each ANN has one input layer with
seven variables (biomass moisture (MC), biomass content of
ash, C, H and O, gasification temperature (Tg) and equivalence ratio (ER)), one hidden layer with two neurons and
one output.
- BFB gasifier model: five ANNs with eight variables in the
input layer (biomass moisture (MC), biomass content of ash,
C, H and O, gasification temperature (Tg), equivalence ratio
(ER) and injected steam ratio (VB)), one hidden layer with
two neurons and one output each ANN.
To test the robustness and predict the ability of the models,
in both ANNs models, the data sets were divided into training
(80%) and validation-test subsets (20%), randomly selected
from the available database. Due to the small size of the
database, validation and test sets were the same.
In all models, a hyperbolic tangent sigmoid function (tansig) was used in the hidden layer and the linear transfer
2
aoutput
13
C7
B
2
C7
B
6
1C7 b2
6LW1;j $B
A5
@
4
P
i7
j1
b1j
1 exp 2$
i1 IWj;i $pi
j2 6
X
(2)
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
Fig. 5 e Comparison of the experimental results with the results calculated by ANN for BFB gasifiers.
285
286
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
00
1
1
C
B
B
C
Pj2 BB IWj;i C
C
C$ LW1;j C
j1 BBPi7
A
@@
A
i1 IWj;i
Ii
1
19
>
>
>
=
B
C
C
Pi7 Pj2 B
BB IWj;i C
C
LW
$
B
C
B
C
P
1;j
i1
j1
i7
>
>
@
A
@
A
>
>
i1 IWj;i
>
>
;
:
8
>
>
>
<
00
(3)
3.2.
Proposed ANN model for bubbling fluidized bed
gasifiers
In this model, the same procedure than that applied for
CFB gasifiers has been followed. The topology of the five
ANNs integrated in the model is the same than in the previous case. However, here, eight input variables are considered instead of seven because the model also accounts for
airesteam gasification and not only for air gasification like in
CFB gasifiers.
The obtained regression coefficients (R2) when comparing
experimental and simulated values for CO, CO2, H2, CH4,
and gas yield are presented in Fig. 5. All R2 values are higher
than 0.99 except for the case of CO2 composition that it is
0.98.
The limits for the statistical test of intercept and slope were
calculated. In all cases, the slope contained one and the
intercept contained zero. Consequently, the proposed ANNs
also passed the test with 99.8% of confidence level.
Table 4 e Weights and biases of the ANNs for the four major producer gas species (CO, CO2, H2, CH4) and producer gas yield
for the ANN model for BFB gasifiers.
CO
IWi,j
0.9005
4.0218
LW1,j
33.7782
22.8979
2.0805
CO2
IWi,j
8.6144
0.4782
LW1,j
3.5726
1.1591
3.9688
CH4
IWi,j
27.6038
56.8348
LW1,j
0.4665
30.0594
245.3845
H2
IWi,j
2.6766
1.0173
LW1,j
13.8413
3.3581
0.0697
31.8927
10.9772
0.3383
0.6249
10.2693
1.9391
b1j
15.6788
3.6788
13.9051
1.0988
0.5125
0.6812
b2
12.3524
1.2177
0.1740
1.6145
0.5222
9.1504
5.2829
4.1321
1.2131
b1j
4.4389
5.3372
0.7413
18.4774
12.6004
3.4298
b2
13.4535
1.6067
6.4298
4.8547
7.5909
31.5068
194.6359
31.9344
29.1672
b1j
28.9205
79.8145
49.1297
243.0979
85.5683
158.8235
b2
4.2972
10.8387
82.2433
1.0029
103.3151
1.7070
3.1264
0.7123
1.8738
b1j
1.3738
0.7616
1.0042
0.1026
1.4738
1.6956
b2
13.6191
0.0854
5.1339
2.3963
6.0746
4.4783
2.1819
23.2472
5.8447
b1j
22.8709
6.0126
19.3959
6.7403
10.3177
4.6368
b2
1.7517
4.3555
4.3425
12.2481
1.4914
39.6833
2.6414
1.0988
8.0323
1.2019
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
2
aoutput
13
j2 6
C7
B
X
2
C7
B
6
1C7 b2
6LW1;j $B
A5
@
4
P
i8
j1
b1j
1 exp 2$
i1 IWj;i $pi
287
(4)
Fig. 6 e Relative impact (%) of input variables on the different outputs for the four main producer gas components and
producer gas yield of the ANN model for BFB gasifiers.
288
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
these ANNs show high agreement with published experimental data used: very good correlations (R2 > 0.98) in almost
all cases and small RMSEs. However, it is necessary to have in
mind that ANN models are limited to a specified range of
operating conditions for which they have been trained. For
this reason, a larger experimental database would be desirable
to get improved models.
4.
Conclusions
Acknowledgments
The authors would like to thank the European Commission for
the financial support received as part of the European Project
Polycity (Energy networks in sustainable communities) (TREN/
05FP6EN/S07.43964/51381)
Nomenclature
ANN
BFB
b1, b2
CFB
ER
FC
IW, LW
MC
VM
H
I
O
C
p
p
R2
RMSE
Tg
VB
Subscripts
i
number of neurons in the input layer
j
number of neurons in the hidden layer
k
number of neurons in the output layer
references
b i o m a s s a n d b i o e n e r g y 4 9 ( 2 0 1 3 ) 2 7 9 e2 8 9
289