Two-Stage Super-Resolution Simulation Method of Three-Dimensional Street-Scale Atmospheric Flows for Real-Time Urban Micrometeorology Prediction

TL;DR

This paper introduces a two-stage CNN-based super-resolution method for urban atmospheric flow simulation, significantly reducing computation time while maintaining high accuracy for real-time micrometeorology predictions.

Contribution

It presents a novel two-stage CNN approach that separates large-scale and small-scale flow inference, enabling efficient high-resolution urban atmospheric simulations.

Findings

Reduced errors by approximately 50% compared to low-resolution simulations.

Lowered GPU memory usage to 12% during training.

Cut total prediction time to 6.83 minutes, 3.32% of high-resolution simulation time.

Abstract

A two-stage super-resolution simulation method is proposed for street-scale air temperature and wind velocity, which considerably reduces computation time while maintaining accuracy. The first stage employs a convolutional neural network (CNN) to correct large-scale flows above buildings in the input low-resolution simulation results. The second stage uses another CNN to reconstruct small-scale flows between buildings from the output of the first stage, resulting in high-resolution inferences. The CNNs are trained using high-resolution simulation data for the second stage and their coarse-grained version for the first stage as the ground truth, where the high-resolution simulations are conducted independently of the low-resolution simulations used as input. This learning approach separates the spatial scales of inference in each stage. The effectiveness of the proposed method was evaluated using micrometeorological simulations in an actual urban area around Tokyo Station in Japan. The super-resolution simulation successfully inferred high-resolution atmospheric flows, reducing errors by approximately 50% compared to the low-resolution simulations. Furthermore, the two-stage approach enabled localized high-resolution inferences, reducing GPU memory usage to as low as 12% during training. The total wall-clock time for 60-min predictions was reduced to 6.83 min, which was 3.32% of the high-resolution simulation time.