Urban air pollution forecasts generated from latent space representation
This paper presents an approach to replicate computational fluid dynamics simulations of air pollution using deep learning. The study area is in London, where a tracer aims to replicate a busy traffic junction. Our method, which integrates Principal Components Analysis (PCA) and autoencoders (AE), is a computationally cheaper way to generate a latent space representation of the original unstructured mesh model. Once the PCA is applied on the original model solution, a Fully-Connected AE is trained on the full-rank PCs. This yields a compression of the original data by $10^{6}$. The number of trainable parameters is also reduced using this method. A LSTM-based approach is used on the latent space to produce faster forecasts of the air pollution tracer.
PDF Abstract