Air Quality Prediction in Smart Environment using Lightweight Residual Network: Sustainable Light-AirNet Approach

Abstract

In densely populated places, air pollution prediction is crucial since it directly affects human health and the local governance. The main objective of this work is to analyze the spatial and temporal patterns of the concentration of the main air pollutants in Bangalore, India. In this paper, a lightweight residual network with an attention mechanism is created using a collection of residual concatenation blocks layered with recursive residual blocks. This aids in the adaptive extraction of useful features, the learning of more expressive spatial context information, and the efficient transfer of information through gradient flow in the network. A unique attention mechanism, known as the Two-Fold Attention Module, has been created with the purpose of enhancing the model’s ability to represent information. The Light-AirNet model was designed to provide hourly forecasts by using past pollution data and three measured weather variables were collected from weather stations. Light-AirNet is compared with existing approaches in terms of different metrics and it was found that it achieves 24.5% of root-mean-square error, 21.5% of mean square error, 12.59% of mean absolute error, and 97.45% of prediction accuracy.