Abstract

In future cities, micrometeorological predictions will be essential to various services such as drone operations. However, the real-time prediction is difficult even by using a super-computer. To reduce the computation cost, super-resolution (SR) techniques can be utilized, which infer high-resolution images from low-resolution ones. The present paper confirms the validity of three-dimensional (3D) SR for micrometeorology prediction in an urban city. A new neural network is proposed to simultaneously super-resolve 3D temperature and velocity fields. The network is trained using the micrometeorology simulations that incorporate the buildings and 3D radiative transfer. The error of the 3D SR is sufficiently small: 0.14 K for temperature and 0.38 m s-1for velocity. The computation time of the 3D SR is negligible, implying the feasibility of real-time predictions for the urban micrometeorology.