17:15 〜 19:15
[ATT35-P06] Modeling of turbulent pollutants transport in planetary boundary layer expliting large eddy simulation and machine learning
キーワード：Machine learning, deep learning, large eddy simulation, atmospheric boundary layer, atmospheric turbulence, pollutants transport
The atmospheric boundary layer (ABL) plays a crucial role in the transport of air pollutants, which has a direct impact on the ecological state of both the atmosphere and underlying surfaces. Therefore, accurate modeling of the spatial distribution of pollutants in the ABL is essential for assessing air quality on different scales.
Traditionally, it has been assumed that the transport of particulate matter in the ABL can be approximated by the normal distribution function. However, current computational methods are limited by turbulent processes that take place during the interaction between air masses and underlying surfaces, which requires significant computational resources. Semi-empirical methods for predicting pollution transport have shown limited accuracy, particularly in areas with complex topography. The aim of this research is to develop a technique for estimating the first moments of the spatial distribution of contaminants in a turbulent ABL using the large eddy simulation (LES) method for numerical modelling and machine learning techniques at the estimation stage..
As part of this study, the first and second central moments of the impurity distribution along the vertical and horizontal axes were analysed, averaged over time under steady-state ABL conditions. To generate a training dataset, a numerical simulation of contaminant transport using the LES method was performed. Global parameters of the medium, such as surface roughness and temperature gradient, as well as parameters of the pollution sources, such as power and position, were specified for the numerical model. These parameters were included in the training sample. The numerical simulation results were compared to approximations calculated using three different methods: traditional Gaussian distribution theory, classical machine learning algorithms, and artificial neural networks.
A comparative analysis of the results showed that machine learning methods outperformed the traditional semi-empirical approach.
In future work, we plan to expand the methodology to include more atmospheric parameters and a wider range of surface conditions.
Traditionally, it has been assumed that the transport of particulate matter in the ABL can be approximated by the normal distribution function. However, current computational methods are limited by turbulent processes that take place during the interaction between air masses and underlying surfaces, which requires significant computational resources. Semi-empirical methods for predicting pollution transport have shown limited accuracy, particularly in areas with complex topography. The aim of this research is to develop a technique for estimating the first moments of the spatial distribution of contaminants in a turbulent ABL using the large eddy simulation (LES) method for numerical modelling and machine learning techniques at the estimation stage..
As part of this study, the first and second central moments of the impurity distribution along the vertical and horizontal axes were analysed, averaged over time under steady-state ABL conditions. To generate a training dataset, a numerical simulation of contaminant transport using the LES method was performed. Global parameters of the medium, such as surface roughness and temperature gradient, as well as parameters of the pollution sources, such as power and position, were specified for the numerical model. These parameters were included in the training sample. The numerical simulation results were compared to approximations calculated using three different methods: traditional Gaussian distribution theory, classical machine learning algorithms, and artificial neural networks.
A comparative analysis of the results showed that machine learning methods outperformed the traditional semi-empirical approach.
In future work, we plan to expand the methodology to include more atmospheric parameters and a wider range of surface conditions.