A Comprehensive Air Quality Prediction Model Based on Enhanced Sparse Autoencoder and Neural Network Architectures

Abstract

This paper presents the Adaptive Spatio-Temporal Representation Learning-based Air Quality Prediction Network (ASTRA-Net), a novel deep learning model to predict air quality, which combines adaptive representation learning in spatio-temporal using representations with robust features encoding in the presence of anomalies. The proposed model will also make use of a improved sparse autoencoder to suppress noise and extract latent features and hybrid models of Long Short-Term Memory (LSTM) and Artificial Neural Networks (ANN) to effectively capture temporal dynamics and high-frequency variations in environmental data. Similarity measures based on kNN are used to form spatial dependencies and these are: Euclidean Distance (kNN-ED) and Dynamic Time Warping Distance (kNN-DTWD). As shown by the results of the experiment, ASTRA-Net has accuracy of 0.99 and precision of 0.921 after 200 iterations, which is better than the ST-DNN baseline. The sensitivity increases to 0.92, with AUC and MCC of 0.8898 and 0.9414 respectively. The model minimises the mean error rate by 52.67 per cent in 500 training epochs.