Time Series Forecasting for Air Quality with Structured and Unstructured Data Using Artificial Neural Networks

Various machine learning algorithms exist to predict air quality, but they can only analyse structured data gathered from monitoring stations. However, the concentration of certain pollutants, such as PM2.5 and PM10, can be visually significant when there is a marked difference in their levels. Consequently, air quality from meteorological cameras can be estimated and integrated with data from monitoring stations to generate an air quality forecast. This research delves into the prospect of creating a methodology capable of rapidly processing this information and producing precise air quality predictions using time series analytics. This paper presents a study of developing a new model, the “Convolutional Neural Network, Recurrent Neural Network Dual Input Model” (CORD). This model combines the convolutional neural network (CNN) and recurrent neural network (RNN) models that are applied to prediction to create an air pollution-related forecasting function to overcome the monitoring stations’ physical limitations. CORD is a model that allows for dual input types: structured data from air quality data collected with meteorological cameras and images (unstructured data) from monitoring stations. This prototype could be applied to all air quality indices worldwide, and CORD is tested based on the Air Quality Health Index provided by the Hong Kong Observatory, a unique data-analytic framework based on air quality measurement. CORD has a similar result to GRU and slightly smaller mean absolute and root mean square errors than LSTM. Compared with an ANN algorithm, CORD has better accuracy.