Forecasting Smog Clouds With Deep Learning

TL;DR

An integrated, hierarchical model architecture inspired by air pollution dynamics and atmospheric science that employs multi-task learning and is benchmarked by unidirectional and fully-connected models is proposed.

Abstract

In this proof-of-concept study, we conduct multivariate timeseries forecasting for the concentrations of nitrogen dioxide (NO2), ozone (O3), and (fine) particulate matter (PM10 & PM2.5) with meteorological covariates between two locations using various deep learning models, with a focus on long short-term memory (LSTM) and gated recurrent unit (GRU) architectures. In particular, we propose an integrated, hierarchical model architecture inspired by air pollution dynamics and atmospheric science that employs multi-task learning and is benchmarked by unidirectional and fully-connected models. Results demonstrate that, above all, the hierarchical GRU proves itself as a competitive and efficient method for forecasting the concentration of smog-related pollutants.

Figures (5)

• Figure 1: Utrecht area with markers indicating the AWS locations.
• Figure 2: Coefficient matrix for the initially considered features. A threshold $r_{th}$ for the absolute Pearson coefficient is set at $r_{th} = 0.15$. When not met, the entry remains white.
• Figure 3: HGRU forecasts for NO2, O3, PM10, and PM2.5 taken for a week from the evaluation set. Black indicates the ground truth and maroon the forecasts. Dashed lines indicate zero.
• Figure 4: Loss plots for all models, showing the training versus validation losses over epochs.
• Figure 5: Training loss plotted of the shared and branched part of the HLSTM. For illustrative purposes, the first epoch is left out from the plot. Both model parts have different complexities (see Section \ref{['sec:model_architecture']}), causing their learning process to be different as well. The branches were more complex, causing its learning process to be less stable, visible by the small "bumps" in its descend.