Deep learning methods for improved numerical simulations of pulverised biomass combustion

点击翻译按钮获取中文摘要

CO2 neutral energy provision is one of the major challenges that our society faces. Even in the mid to long term perspective, it is unlikely that one single technology will replace all fossil fuels as primary energy source and several options need to be pursued. Pulverized biomass combustion (PBC) is considered to be an attractive option as it allows to re-use the current infrastructure of coal-firing power plants. The combustion characteristics of biomass and coal are, however, not identical and some combustor and combustion chamber modifications will be needed. Such modifications can be aided by computer simulations and large-eddy simulation (LES) presents itself as a promising tool as it captures the inherently non-linear transient turbulent processes that are critical for the performance of solid fuel combustion systems. LES does not, however, resolve all the complex small scale processes that can be associated with pulverized biomass combustion such as heating, drying, pyrolysis, homogeneous combustion of released volatile gases and heterogeneous char reactions. These processes and their interactions need first to be thoroughly understood and second to be modelled. High fidelity direct numerical simulations shall be used to assess some of the pulverized biomass combustion fundamentals such as ignition and flame characteristics and how they are affected by the surrounding gas phase conditions, particle properties and particle-gas interactions. The simulations will include rather detailed multi-step chemistry that is needed for thorough analysis of all sub-grid scale interactions. It is of practical interest, however, to also obtain a computationally more affordable, yet reasonably accurate modelling approach for PBC. We intend to achieve a significant cost reduction by state-of-the-art flamelet approaches similar to those applied to pulverized coal combustion where the composition space could be parameterized by a set of mixture fractions, progress variable and total enthalpy. DNS using the complete and the reduced composition spaces shall then be used to develop LES sub-grid models with the aid of machine-learning methods. Algorithms that contain deep neural layers shall learn the underlying sub-grid correlations, generalize these correlations and generate synthetic data that are indistinguishable from reality such that they can be used as closures in LES. In a final step, the closures will be tested by stand-alone LES of biomass combustion and comparison with the corresponding DNS data.