Being a direct computational aeroacoustics method, Lattice Boltzmann method (LBM) has great potential and broad application perspective in the field of numerical simulation of aerodynamic noise due to its low dispersion and low dissipation. A series of numerical algorithms and the related improvements based on the standard LBM method are proposed and developed in this paper to adapt to the airfoil noise calculation with complex grid at middle‐high Reynolds number. First, a new mathematical‐boundary‐recognition algorithm based on Green's formula is proposed to deal with complex curved geometric models, which is validated by three‐element airfoil 30P30N benchmark. Then, in order to reduce grid redundancy and improve computing efficiency, the grid refinement technique of domain decomposition model (DDM) is adopted and also improved, which is verified by calculating the flow and sound fields around 2D and 3D cylinders at Reynolds number equal to 90,000. Finally, three different LES turbulence models are combined with the standard MRT‐LBM method, where different finite difference schemes are used to solve Reynolds stress tensor which is different from the traditional one. Through the direct acoustic numerical simulation of NACA0012 airfoil at Reynolds number equal to 200,000, the effects of Smagorinsky models and Wall‐adapting local eddy‐viscosity (WALE) model on aerodynamic noise prediction are compared and analyzed. Overall, the proposed methodology is shown to be appropriate for predicting the aerodynamic noise at low Mach number and can successfully simulate the generation and propagation of far field acoustics.