A deep-neural-network accelerated precursor-based method for atmospheric boundary layers

Abstract

The use of the precursor simulation method to generate inflowing atmospheric boundary layer turbulence is confronted with the problem of excessive computational load due to the excessive number of grids in the reference flow field. For this reason, this paper proposes an inflow model via a spatial and temporal resolution enhancement method based on deep neural networks. It can deduce the high-resolution computational domain inlet used in formal calculations based on the low-resolution reference flow field, thus achieving a more efficient generation of atmospheric boundary layer turbulence. This method includes two key modules: the Temporal Resolution Enhancement Module (TREM) and the Spatial Resolution Enhancement Module (SREM). By evaluating the effects of autoencoders with different compression ratios and time series prediction models, the optimal performance of TREM has been studied to achieve the best effect of temporal resolution enhancement. Meanwhile, the SREM was constructed by using a high performance autoencoder and combined with the TREM to form the spatial and temporal resolution enhancement model. By simulating the turbulent flow field of an atmospheric boundary layer wind tunnel, the results show that after using the spatial and temporal resolution enhancement method, the turbulent data has been improved in statistical characteristics such as the mean wind speed, turbulence intensity, power spectrum of fluctuating wind speed and spanwise spatial energy spectrum, as well as in the flow field structure, approaching the results of large eddy simulations (LES) with high spatial and temporal resolution. Compared with the traditional precursor simulation method, the generation speed is approximately 12 times faster.