Abstract Coal is complex and heterogeneous, with extremely variable properties. As a result, it has proved very difficult to construct generalized physical descriptions of pulverized coal combustion for incorporation into reliable mathematical models suited to industrial applications. There are many processes to be simulated: pyrolysis, char kinetics, particle/turbulence interaction, etc. This paper is concerned with the early stages of pyrolysis, which significantly affect flame stability, NO formation, soot formation, and ultimately, char burn-out. In most of the existing predictive procedures for devolatilization, combustion and emissions are modeled by a single-step global chemical reaction, with the yield of volatile matter presumed to experience mixing-controlled combustion. Several more detailed multi-step coal devolatilization models have recently emerged, having a range of capabilities, e.g., predicting the thermal decomposition of a coal under practical conditions. A common shortcoming of these models is that they require a large set of input data, involving kinetic parameters, gas precursor compositions, and additional parameters describing the coal’s polymeric structure. The input data must be generated from an extensive series of experimental measurements for each coal of interest. Very significant computational expense and application restricted to coals, which have already been studied, are implied. All of these problems are exacerbated when coal blending or co-firing with renewable solid fuels, such as forest and agricultural waste, and sewage sludge, is considered. In this paper, a new approach based on neural networks is proposed; it is capable of handling a range of solid fuels. The model considers heating rate, fuel atomic ratios, and the temperature of the fuel particles to predict the volatiles released by the particles. The “learning” properties of the model implicitly facilitate all the physical conditions of devolatilization experiments, which were used during its training and validation phases. The neural-network model was implemented into an existing 3D CFD combustion code. The predictions for high- and low-NO x burners demonstrate improved prediction of in-flame data for reduced computational effort, one-fifth of that with the standard single-global-reaction devolatilization model. Its devolatilization predictions have also been compared with a detailed devolatilization model (FLASHCHAIN) and were found to be comparable.
[1]
M.H. Hassoun,et al.
Fundamentals of Artificial Neural Networks
,
1996,
Proceedings of the IEEE.
[2]
A. M. Eaton,et al.
Components, formulations, solutions, evaluation, and application of comprehensive combustion models
,
1999
.
[3]
A. Gosman,et al.
Solution of the implicitly discretised reacting flow equations by operator-splitting
,
1986
.
[4]
J. Truelove,et al.
Models for rapid devolatilization of pulverized coal
,
1986
.
[5]
F. Lockwood,et al.
Assessment of kinetic and mixing-controlled models for the prediction of fuel-NO emissions in P.F. burners
,
1994
.
[6]
Igor Aleksander,et al.
Introduction to Neural Computing
,
1990
.
[7]
Jacek M. Zurada,et al.
Introduction to artificial neural systems
,
1992
.
[8]
R. E. Rebis,et al.
Spray Forming Quality Predictions via Neural Networks
,
1993
.
[9]
Alan R. Kerstein,et al.
Chemical model of coal devolatilization using percolation lattice statistics
,
1989
.
[10]
S. Niksa.
Predicting the Devolatilization Behavior of any coal from its ultimate analysis
,
1995
.
[11]
W. Jones,et al.
The prediction of laminarization with a two-equation model of turbulence
,
1972
.