NeSDeepNet: A Fusion Framework for Multi-step Forecasting of Near-surface Air Pollutants

Abstract

Air pollution is a global issue that poses significant threats to human health and the environment due to industrial development. Forecasting the concentrations of major pollutants such as NO2 and CO can provide early warnings of harmful substances, minimizing health risks and losses. Recent deep learning models have shown promise in air quality prediction, but they have limitations such as insufficient feature representation, high computational costs, and poor generalization. This paper proposes a near-surface deep network (NeSDeepNet) to overcome these limitations. The NeSDeepNet integrates multiple deep learning models and a shallow model to form a hybrid forecasting system. The proposed framework consists of three modules: a preliminary extraction module, a deep extraction module, and a fusion module. The feature extraction module uses a multi-layer network to extract features from air pollutant and meteorological data, and each of which predicts air pollutants for different forecasting horizons. The fusion module combines the outputs of the deep learning module and the shallow models to produce the final prediction results. The proposed framework is evaluated on a real-world dataset, and the experimental results demonstrate that NeSDeepNet achieves optimal RMSE value of 9.59 for NO2 and 274.0 for CO, MAE value of 2.64 for NO2 and 13.75 for CO, and R2 values 0.89 for NO2 and 0.93 for CO, respectively, outperforming cutting-edge deep learning models. Therefore, NeSDeepNet can be a valuable tool for air quality forecasting and miti-gating the adverse effects of air pollution on human health and the environment. The source code for our proposed NeSDeepNet and comparative models is available on GitHub repository: https://github.com/Prasanjit-Dey/NeSDeepNet.