Dispersion prediction of pollutants in settlements based on generative adversarial networks.

Aiming at the traditional built environment pollutant study which is difficult to meet the requirements of rapid prediction, a data-driven pollutant dispersion prediction method at the neighbourhood scale is proposed. With reference to the data-driven fluid modelling approach, the research is carried out, based on the CFD simulation to collate and produce the relevant datasets with better adaptability to the real urban environment, build the data-driven pollutant diffusion model, compare the performance of three generative adversarial network Pix2Pix, CycleGAN, and Pix2PixHD models in real-time pollutant monitoring scenarios and further optimize the experiments based on the performance differences of each model. performance differences to further optimise the experiment. The experimental results show that the data-driven pollutant dispersion prediction model based on the generative adversarial network can approximate the CFD simulation results, realise the near real-time prediction of pollutant concentration distributions in settlements, and satisfy the scenarios with high requirements on computation time. The model stability improvement and dataset optimisation have positive gain on the model performance.