A novel GPLS-GP algorithm and its application to air temperature prediction

Abstract

In this paper, a novel regression algorithm, the Generalized Partial Least Squares Gaussian Process (GPLS-GP), is developed to improve the prediction performance of regression model. Profiting from the latent variables extraction power of PLS, noise, co-linearity between independent variables and other difficult problems could be overcome successfully. More importantly, by designing generalizing variables rationally and by taking advantages of the nonlinear regression superiority of GP (Gaussian process) to calculate the inner model, the nonlinear relationship of the process could be modeled to the most extreme. The theoretical findings are fully supported by the application performed on the prediction of the mean temperature of Izmir of Turkey. It is shown, in comparison to conventional approaches (GPLS, PLS and GP), the model of GPLS-GP yields superior performance while the Root-Mean-Square-Error (RMSE) of calibration and prediction are both improved notably.