Air Quality Prediction using Voronai-Based Spatial Temporal Sequence Similarity with Conjugate Gradient Enabled Sparse Autoencoder Deep Learning

Abstract

Air Quality Prediction (AQP) remains a difficult task because of multidimensional nonlinear spatiotemporal features. To solve this issue, an Improved Sparse Autoencoder with Deep Learning (ISAE-DL) and Enriched ISAE-DL (EISAE-DL) models were developed with the combination of concentric circle-based clustering, Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) followed by the ISAE for AQP. In EISAE-DL, concentric circle-based clustering uses Manhattan distance to efficiently split the locations into four regions using its center and cluster the spatially and temporally similar candidate locations. But it was considered a fixed structure and may struggle to find variations in several data points. Also, it accommodate clusters with regular and circular patterns, whereas irregular and non-circular cluster patterns were not handled. Similarly, the ANN inference was often offended or ignored because of complex meteorological characteristics. Hence, this paper proposes a Voronoi-based spatial-temporal sequence similarity with the Conjugate gradient-enabled SAE-DL (VCSAE-DL) model for effective AQP. First, a Voronoi clustering is performed by creating the Voronoi diagram for analogous candidate location clustering. Then, the resultant clusters of location data along with the PM2.5 and other meteorological data are given to the Improved ANN (IANN), and the target stations are given to the LSTM to capture the spatiotemporal relationship features and temporal features, respectively. Also, CNN is used to extract relationships between terrain and air quality features. These extracted features are fused in the merge layer and transferred to the ISAE for final prediction of air quality. Finally, the test outcomes demonstrate that the VCSAE-DL achieves better prediction performance compared to the existing AQP models.