Simulation of biomass gasification with a hybrid neural network model
Bioresource Technology 76 (2001) 77±83
Simulation of biomass gasi®cation with a hybrid neural network model ,1
Bing Guo * , Dingkai Li, Congming Cheng, Zi-an L u, Youting Shen Department of Thermal Engineering, Tsinghua University, Beijing 100084, People's Republic of China Received 30 June 1998; received in revised form 25 July 2000; accepted 25 July 2000
Abstract Gasi®cation of several types of biomass has been conducted in a ¯uidized bed gasi®er at atmospheric pressure with steam as the ¯uidizing medium. In order to obtain the gasi®cation pro®les for each type of biomass, an arti®cial neural network model has been developed to simulate this gasi®cation processes. Model-predicted gas production rates in this biomass gasi®cation processes were consistent with the experimental data. Therefore, the gasi®cation pro®les generated by neural networks are considered to have properly re¯ected the real gasi®cation process of a biomass. Gasi®cation pro®les identi®ed by neural network suggest that gasi®cation behavior of arboreal types of biomass is signi®cantly dierent from that of herbaceous ones. Ó 2000 Published by Elsevier Science Ltd. Keywords: Biomass; Gasi®cation; Neural network; Simulation; Fluidized bed
1. Introduction Biomass gasi®cation in a steam and/or air blown ¯uidized bed reactor is one of the more favorable processes to convert solid biomass materials to gaseous fuel. The advantages of this process include modest equipment demand, ¯exible scale-up of installation, perfect solid-gas contacting and ecient gas±solid and gas phase reactions in the bed, relative high yield and medium heat value of product gas (for steam ¯uidized bed). It is necessary to simulate the above process for scaleup and industrial control in applications of the process. Yet, it is dicult to develop a simulation model for this process. This is due to the complexity of biomass gasi®cation in a ¯uidized bed reactor, which involves gas± solid two phase ¯ow, heat and mass transfer, pyrolysis of biomass material, cracking and subsequent steam reforming of tar vapor arising from the pyrolysis, heterogeneous gas±solid reactions and homogeneous gas± phase reactions. In addition, behavior of the gasi®cation process is dependent on many factors, e.g., type of * Corresponding author. Present address: Department of Mechanical and Aeronautical Engineering, University of California, Davis, CA 95616, USA. Tel.: +1-530-752-6214; fax: +1-530-752-4158. E-mail addresses: [email protected], [email protected] (B. Guo). 1 Tel.: +86-10-62781741; fax: +86-10-62770209.
biomass material, operating temperature and pressure, residence time of solid and gas in the gasi®er, and biomass and steam and/or air feed rates. Although considerable research in this ®eld has been done (Corella et al., 1989; Aznar et al., 1989; Reinoso et al., 1995), actual studies involved in process modeling are minimal (Bilodeau et al., 1993). Mathematical models of biomass gasi®cation in a ¯uidized bed, which are employed for performance prediction of the process, are usually mechanism models. However, due to the complexity of the real world processes, many idealized assumptions have to be made in the development of these models. Arti®cial neural networks (ANN) have been extensively used in the ®eld of pattern recognition, signal processing, function approximation and process simulation (Tsen et al., 1996; Leib et al., 1996; Guo et al., 1997). Sometimes a hybrid neural network (HNN) model is synthesized for process modeling (Psichogios and Ungar, 1992). This modeling approach usually combines a partial ®rst principles model, which describes certain characteristics of the process being simulated and involves a multilayer feedforward neural network (MFNN), that serves as an estimator of unmeasured process parameters that are dicult to model from ®rst principles. It is well known that MFNN is a universal function approximator, which has the ability to approximate any continuous function to an arbitrary
0960-8524/01/$ - see front matter Ó 2000 Published by Elsevier Science Ltd. PII: S 0 9 6 0 - 8 5 2 4 ( 0 0 ) 0 0 1 0 6 - 1
78
B. Guo et al. / Bioresource Technology 76 (2001) 77±83
precision even without a priori knowledge on structure of the function to be approximated (Hornik, 1991). This characteristic of HNN provides a simple and feasible way to explore the underlying biomass gasi®cation process in a ¯uidized bed gasi®er. In this study, a HNN model has been developed to simulate the biomass gasi®cation in a steam ¯uidized bed gasi®er. A series of gasi®cation runs have been conducted on a bench scale facility, with four types of biomass as feed stock. The data obtained from these experiments were used to train the HNN model. The resulting model predictions and gasi®cation pro®les for the dierent types of biomass revealed by the neural network were investigated in detail.
inner diameter and 1360 mm in height. It was heated by a cylindrical electrical heating element, capable of delivering up to 12 kW of power. Four thermocouples were mounted on the reactor at the height of 80, 300, 600 and 1100 mm from the distributor, respectively. The steam generator had a rated capacity of 3 kW. This produced saturated steam, which was superheated in and extracted from the outer jacket of the feed pipe inside the reactor, and then introduced into the reactor as the sole ¯uidizing agent. A screw feeder controlled by an inverter supplied the feed stock to the reactor. A purge stream of N2 was introduced into the hopper of the screw feeder to prevent gas back ¯ow from the reactor and subsequent condensation of vapor in the feeder.
2. Methods
2.2. Materials
2.1. Facility
Four types of biomass material were used as gasi®cation stocks in this study: poplar sawdust, pine sawdust, comminuted sugar cane bagasse and cotton stem. The proximate and ultimate analyses and other properties of these stocks are shown in Table 1. Coal ash from a local circulating ¯uidized bed boiler was selected as the inert bed material for the ¯uidized bed reactor. This ash was calcined at 800°C for 2 h prior to each experimental gasi®cation run, in order to
The schematic diagram of the atmospheric pressure steam ¯uidized bed biomass gasi®cation facility used in this work is shown in Fig. 1. The facility consisted of six main components: ¯uidized bed reactor, feedstock feeding unit, steam generator, gas cooler and cleaner, measurement and control system and sampling system. The reactor was made of stainless steel tube, 150 mm in
Fig. 1. Experimental apparatus.
B. Guo et al. / Bioresource Technology 76 (2001) 77±83
79
Table 1 Proximate and ultimate analyses of the biomass stocks Poplar sawdust
Pine sawdust
Bagasse
Cotton stem
Proximate analysis (as received basis) Moisture (wt%) Fixed carbon (wt%) Volatile matter (wt%) Ash (wt%)
10.0 0.2 12.3 0.2 73.8 0.2 3.9 0.2
9.4 0.2 14.1 0.2 75.7 0.2 0.9 0.2
7.1 0.2 11.85 80.5 0.2 0.9 0.2
7.9 0.2 15.5 0.2 72.3 0.2 4.3 0.2
Ultimate analysis (as received basis) C (wt%) H (wt%) O (wt%) N (wt%) S (wt%)
43.1 0.2 5.1 0.2 37.7 0.2 0.2 0.2 0.1 0.2
45.3 0.2 5.4 0.2 39.0 0.2 0.1 0.2 0.0 0.2
46.0 0.2 5.4 0.2 40.3 0.2 0.2 0.2 0.1 0.2
42.8 0.2 5.3 0.2 38.5 0.2 1.0 0.2 0.2 0.2
15.5 0.1 48.9 0.3 16.8 0.3 20.8 0.3 123 1
16.4 0.1 42.4 0.3 11.5 0.3 26.7 0.3 131 1
16.2 0.1 44.1 0.3 25.4 0.3 21.0 0.3 86 1
15.2 0.1 36.7 0.3 18.3 0.3 20.9 0.3 176 1
LHV (MJ/kg) (as received basis) Cellulose (wt%) Hemicellulose (wt%) Lignin (wt%) Bulk density (kg/m3 )
Table 2 Size distribution of the gasi®cation stocks and the inert bed material Gasi®cation stocks Size range (mm)
>2.0 2.0±1.0 1.0±0.71 0.71±0.63 0.63±0.442 0.442±0.355 0.355±0.280 0.280±0.105 0.105±0.098 0.098±0.061 1.0 1.0±0.71 0.71±0.45 0.45±0.28 0.28±0.18 0.18±0.105 0.105±0.065
Simulation of biomass gasi®cation with a hybrid neural network model ,1
Bing Guo * , Dingkai Li, Congming Cheng, Zi-an L u, Youting Shen Department of Thermal Engineering, Tsinghua University, Beijing 100084, People's Republic of China Received 30 June 1998; received in revised form 25 July 2000; accepted 25 July 2000
Abstract Gasi®cation of several types of biomass has been conducted in a ¯uidized bed gasi®er at atmospheric pressure with steam as the ¯uidizing medium. In order to obtain the gasi®cation pro®les for each type of biomass, an arti®cial neural network model has been developed to simulate this gasi®cation processes. Model-predicted gas production rates in this biomass gasi®cation processes were consistent with the experimental data. Therefore, the gasi®cation pro®les generated by neural networks are considered to have properly re¯ected the real gasi®cation process of a biomass. Gasi®cation pro®les identi®ed by neural network suggest that gasi®cation behavior of arboreal types of biomass is signi®cantly dierent from that of herbaceous ones. Ó 2000 Published by Elsevier Science Ltd. Keywords: Biomass; Gasi®cation; Neural network; Simulation; Fluidized bed
1. Introduction Biomass gasi®cation in a steam and/or air blown ¯uidized bed reactor is one of the more favorable processes to convert solid biomass materials to gaseous fuel. The advantages of this process include modest equipment demand, ¯exible scale-up of installation, perfect solid-gas contacting and ecient gas±solid and gas phase reactions in the bed, relative high yield and medium heat value of product gas (for steam ¯uidized bed). It is necessary to simulate the above process for scaleup and industrial control in applications of the process. Yet, it is dicult to develop a simulation model for this process. This is due to the complexity of biomass gasi®cation in a ¯uidized bed reactor, which involves gas± solid two phase ¯ow, heat and mass transfer, pyrolysis of biomass material, cracking and subsequent steam reforming of tar vapor arising from the pyrolysis, heterogeneous gas±solid reactions and homogeneous gas± phase reactions. In addition, behavior of the gasi®cation process is dependent on many factors, e.g., type of * Corresponding author. Present address: Department of Mechanical and Aeronautical Engineering, University of California, Davis, CA 95616, USA. Tel.: +1-530-752-6214; fax: +1-530-752-4158. E-mail addresses: [email protected], [email protected] (B. Guo). 1 Tel.: +86-10-62781741; fax: +86-10-62770209.
biomass material, operating temperature and pressure, residence time of solid and gas in the gasi®er, and biomass and steam and/or air feed rates. Although considerable research in this ®eld has been done (Corella et al., 1989; Aznar et al., 1989; Reinoso et al., 1995), actual studies involved in process modeling are minimal (Bilodeau et al., 1993). Mathematical models of biomass gasi®cation in a ¯uidized bed, which are employed for performance prediction of the process, are usually mechanism models. However, due to the complexity of the real world processes, many idealized assumptions have to be made in the development of these models. Arti®cial neural networks (ANN) have been extensively used in the ®eld of pattern recognition, signal processing, function approximation and process simulation (Tsen et al., 1996; Leib et al., 1996; Guo et al., 1997). Sometimes a hybrid neural network (HNN) model is synthesized for process modeling (Psichogios and Ungar, 1992). This modeling approach usually combines a partial ®rst principles model, which describes certain characteristics of the process being simulated and involves a multilayer feedforward neural network (MFNN), that serves as an estimator of unmeasured process parameters that are dicult to model from ®rst principles. It is well known that MFNN is a universal function approximator, which has the ability to approximate any continuous function to an arbitrary
0960-8524/01/$ - see front matter Ó 2000 Published by Elsevier Science Ltd. PII: S 0 9 6 0 - 8 5 2 4 ( 0 0 ) 0 0 1 0 6 - 1
78
B. Guo et al. / Bioresource Technology 76 (2001) 77±83
precision even without a priori knowledge on structure of the function to be approximated (Hornik, 1991). This characteristic of HNN provides a simple and feasible way to explore the underlying biomass gasi®cation process in a ¯uidized bed gasi®er. In this study, a HNN model has been developed to simulate the biomass gasi®cation in a steam ¯uidized bed gasi®er. A series of gasi®cation runs have been conducted on a bench scale facility, with four types of biomass as feed stock. The data obtained from these experiments were used to train the HNN model. The resulting model predictions and gasi®cation pro®les for the dierent types of biomass revealed by the neural network were investigated in detail.
inner diameter and 1360 mm in height. It was heated by a cylindrical electrical heating element, capable of delivering up to 12 kW of power. Four thermocouples were mounted on the reactor at the height of 80, 300, 600 and 1100 mm from the distributor, respectively. The steam generator had a rated capacity of 3 kW. This produced saturated steam, which was superheated in and extracted from the outer jacket of the feed pipe inside the reactor, and then introduced into the reactor as the sole ¯uidizing agent. A screw feeder controlled by an inverter supplied the feed stock to the reactor. A purge stream of N2 was introduced into the hopper of the screw feeder to prevent gas back ¯ow from the reactor and subsequent condensation of vapor in the feeder.
2. Methods
2.2. Materials
2.1. Facility
Four types of biomass material were used as gasi®cation stocks in this study: poplar sawdust, pine sawdust, comminuted sugar cane bagasse and cotton stem. The proximate and ultimate analyses and other properties of these stocks are shown in Table 1. Coal ash from a local circulating ¯uidized bed boiler was selected as the inert bed material for the ¯uidized bed reactor. This ash was calcined at 800°C for 2 h prior to each experimental gasi®cation run, in order to
The schematic diagram of the atmospheric pressure steam ¯uidized bed biomass gasi®cation facility used in this work is shown in Fig. 1. The facility consisted of six main components: ¯uidized bed reactor, feedstock feeding unit, steam generator, gas cooler and cleaner, measurement and control system and sampling system. The reactor was made of stainless steel tube, 150 mm in
Fig. 1. Experimental apparatus.
B. Guo et al. / Bioresource Technology 76 (2001) 77±83
79
Table 1 Proximate and ultimate analyses of the biomass stocks Poplar sawdust
Pine sawdust
Bagasse
Cotton stem
Proximate analysis (as received basis) Moisture (wt%) Fixed carbon (wt%) Volatile matter (wt%) Ash (wt%)
10.0 0.2 12.3 0.2 73.8 0.2 3.9 0.2
9.4 0.2 14.1 0.2 75.7 0.2 0.9 0.2
7.1 0.2 11.85 80.5 0.2 0.9 0.2
7.9 0.2 15.5 0.2 72.3 0.2 4.3 0.2
Ultimate analysis (as received basis) C (wt%) H (wt%) O (wt%) N (wt%) S (wt%)
43.1 0.2 5.1 0.2 37.7 0.2 0.2 0.2 0.1 0.2
45.3 0.2 5.4 0.2 39.0 0.2 0.1 0.2 0.0 0.2
46.0 0.2 5.4 0.2 40.3 0.2 0.2 0.2 0.1 0.2
42.8 0.2 5.3 0.2 38.5 0.2 1.0 0.2 0.2 0.2
15.5 0.1 48.9 0.3 16.8 0.3 20.8 0.3 123 1
16.4 0.1 42.4 0.3 11.5 0.3 26.7 0.3 131 1
16.2 0.1 44.1 0.3 25.4 0.3 21.0 0.3 86 1
15.2 0.1 36.7 0.3 18.3 0.3 20.9 0.3 176 1
LHV (MJ/kg) (as received basis) Cellulose (wt%) Hemicellulose (wt%) Lignin (wt%) Bulk density (kg/m3 )
Table 2 Size distribution of the gasi®cation stocks and the inert bed material Gasi®cation stocks Size range (mm)
>2.0 2.0±1.0 1.0±0.71 0.71±0.63 0.63±0.442 0.442±0.355 0.355±0.280 0.280±0.105 0.105±0.098 0.098±0.061 1.0 1.0±0.71 0.71±0.45 0.45±0.28 0.28±0.18 0.18±0.105 0.105±0.065
