A Gaussian process mixture model-based hard-cut iterative learning algorithm for air quality prediction

Abstract

Air quality is closely related to concentrations of gaseous pollutants, and the prediction of gaseous pollutant concentration plays a decisive role in regulating plant and vehicle emissions. Due to the non-linear and chaotic characteristics of the gas concentration series, traditional models may not easily capture the complex time series pattern. In this study, the Gaussian Process Mixture (GPM) model, which adopts hidden variables posterior hard-cut (HC) iterative learning algorithm, is first applied to the prediction of gaseous pollutant concentration in order to improve prediction performance. This algorithm adopts iterative learning and uses the maximizing a posteriori (MAP) estimation to achieve the optimal grouping of samples which effectively improves the expectation–maximization (EM) learning in GPM. The empirical results of the GPM model reveals improved prediction accuracy in gaseous pollutant concentration prediction, as compared with the kernel regression (K-R), minimax probability machine regression (MPMR), linear regression (L-R) and Gaussian Processes (GP) models. Furthermore, GPM with various learning algorithms, namely the HC algorithm, Leave-one-out Cross Validation (LOOCV), and variational algorithms, respectively, are also examined in this study. The results also show that the GPM with HC learning achieves superior performance compared with other learning algorithms.