The wind field of a mountain bridge site always shows a strong randomness and uncertainty. As a result, setting up its computational domain in CFD simulation by a simply reference to the existing experience causes a large amount of model errors and huge additional computational overhead. To solve this problem, a method to the selection of CFD computational domain of complex terrains is explored and verified. To start the procedure, a large reference computational domain should be chosen and preliminary solved. Then, its solution should be post-processed to draw the following curves or contour images: range of mean pressure coefficient align with height, static pressure deviation contour of the horizontal plane near the top of the bottom surface,the root-mean-square error of the static pressure of the crosswind planes compared to the corresponding edge surface. According to these curves and contour images, areas parts which have slight contribution on the overall wind field are figured out and abandoned, and the selected computational domain is constitutive of the rest parts of the reference domain.