Multivariate Procedure for Modeling and Prediction of Temperature in Punjab, Pakistan

Kanwal, Bushra; Ashraf, Zeeshan; Mehmood, Tahir; Kanwal, Summrina; Dashtipour, Kia; Gogate, Mandar

Abstract

Climate study often relies upon global climate models (GCM) to project future scenarios of change in climate behavior. This study aims to refine GCM results to fill the gap between local scale surface weather with regional atmospheric predictors. The world is toward the hotter side; the reason is natural variability or activities of humans. Temperature readings all over the globe rose slowly and gradually since the industrial revolution started. In this study, we use partial least square (PLS) for modeling the minimum and maximum daily temperature of Punjab’s 11 stations for the period 1961–2001 using A2 scenario. HadCM3 (Hadley Centre Coupled Model 3) data of 26 variables are used for calibration and validation. After calibration the model is validated. As far as the high-dimensional data is concerned, employing multivariate methods for modeling the actual life phenomena is innate as well as natural. In existence of multicollinearity and identification issues, ordinary least square is unable to successfully model the bond between response variable and explanatory variables. PLS is considered as a better solution to this situation. The multivariate procedure known as PLS was successfully used for identification of influential variables for high-dimensional data. The latest PLS methods for variable selection are based on PLS loading weights, assuming the loading weights are normally distributed, which may not be the case in some situations. Modeling the loading weights with leptokurtic distributions like Laplace probability distribution can improve the mapping. The results are compared with selection of variables through partial least square based on soft-threshold, uninformative variable eradication partial least square, and distribution-based truncation in PLS. To have reliable parameter estimates and performance assessment, Monte Carlo simulation has been used. Finally, through RMSE we observe which method is best among all these partial least square methods.

• Online Publication Date: Feb 21, 2024
• Publication Date: 2024
• Deposit Date: May 21, 2024
• Publisher: Springer
• Pages: 99-124
• Book Title: Decision Making and Security Risk Management for IoT Environments
• DOI: https://doi.org/10.1007/978-3-031-47590-0_6