Regional spatio-temporal forecasting of particulate matter using autoencoder based generative adversarial network

Accurate forecasting of air pollutant PM 2.5 (particulate matter with diameter less than 2.5 µm) is beneficial to society. However, the non-linear spatio-temporal correlations, multi-feasible forecast values and incomplete training data due to stochasticity make it challenging for discriminative deep learning approaches to forecasting PM 2.5 data. In this paper, a generative modeling approach is proposed to overcome the challenges in forecasting PM 2.5 data by considering it as an ill-posed inverse problem. To strengthen its applicability, the proposed approach is theoretically validated. Furthermore, based on the proposed generative modeling, an Autoencoder-based generative adversarial network (GAN) named Air-GAN is developed. Air-GAN combines a convolutional neural network- long short-term memory (CNN-LSTM) based Encoder with a conditional Wasserstein GAN (WGAN) to capture non-linear correlations in the data distribution via inverse mapping from the forecast distribution. The condition vector to conditional WGAN is the novelty in Air-GAN, which employs this inverse learning and allows the WGAN’s Generator to generate accurate forecast estimates from noise distribution. The condition vector is composed of two elements: (1) the category label of the best correlated meteorological parameter with the PM 2.5 data, assigned using an efficient classifier and (2) the output of the CNN-LSTM-based Encoder which is the latent representation of the forecast. The extensive evaluation of Air-GAN for predicting the real-time PM 2.5 data of Delhi demonstrates its superior performance with an average inference error of 5.3 µg/m 3 , which achieves 31.7% improvement over the baseline approaches. The improved performance of Air-GAN demonstrates its efficiency to forecast stochastic PM 2.5 data by generalizing to out-of-distribution data.