Probabilistic forecasting of multiple air pollutants via latent dynamics modelling with deep learning

Fine particulate matter (PM_2.5), ozone (O₃) and nitrogen dioxide (NO₂) each pose significant risks to public health and are among the World Health Organisation (WHO) criteria pollutants. Operational air quality forecasting relies on computationally expensive Chemical Transport Models (CTMs), and recent deep learning methods focus on station-based forecasts, limiting usability to areas with station networks. We present a deep learning framework for probabilistic, gridded ambient air pollution forecasting to address both limitations. 

Our approach employs a latent dynamics architecture. A convolutional variational autoencoder (Conv-VAE) learns compressed latent representations of input channels. A temporal core block captures the dynamical evolution of ambient pollutants in the latent space, and a probabilistic decoder reconstructs forecasts with uncertainty intervals. Probabilistic forecasting allows for more trustworthy predictions, as stakeholders are also presented with relevant confidence. We systematically compare four latent cores: ConvLSTM, Mamba (state-space model), Transformer (attention) and Neural ODEs. This comparison will identify which inductive bias best represents the dynamics of ambient air pollution evolution.  

Experiments utilise a dataset of CAMS European reanalysis and ERA5 reanalysis (by ECMWF), as well as EDGAR emissions inventories over the UK (2015-2022), targeting 24–72 hour forecast horizons. Multi-pollutant settings test the framework's capacity to represent species with distinct atmospheric and chemical interactions in a unified latent representation. We will evaluate forecast skill, uncertainty quantification and computational efficiency of all models. Ongoing work is exploring physics-informed constraints, stochastic latent dynamics, and self-supervised pre-training for improved generalisation.