An enhanced boosted regression tree model for the prediction of PM10 concentration level using SVM_BRT with QR loss function coupling approach / Wan Nur Shaziayani Wan Mohd Rosly

• Main Author: Wan Mohd Rosly, Wan Nur Shaziayani
• Format: Thesis
• Language: English
• Published: 2022
• Subjects: Atmospheric temperature
• Online Access: https://ir.uitm.edu.my/id/eprint/66923/1/66923.pdf; https://ir.uitm.edu.my/id/eprint/66923/
• Institution: Universiti Teknologi Mara

Malaysia experiences transboundary haze episodes in which the air contains particulate matter (PM) that is harmful to human health and the environment. Therefore, the main prediction model used in this study is Boosted Regression Trees (BRT) to predict three days ahead of PM10 concentration. However, the main problem with the common BRT technique is that it is not suitable for use in predicting extreme values of PM10 concentration levels. Besides, the problem with BRT is that overfitting can occur if the number of trees is not suitable and also because of the complexity of the model, which is caused by the unsuitable number of predictor variables used in the model. Therefore, the aim of this study is to enhance the BRT model with Quantile Regression (QR) and Support Vector Machine (SVM) weight. This study used maximum daily monitoring records from 2002 to 2017 in Alor Setar, Klang, and Kuching which were analysed using four models: a boosted regression tree (BRT) model, a BRT with QR loss function model and a hybrid model between SVM and BRT with and without QR loss function. In order to get the best prediction model and to avoid over-fitting, the number of trees (nt) was optimized by using independent test set (TEST), cross validation (CV) and out of bag estimation (OOB). Then, to solve the extreme value issue in BRT, this study used the QR loss function rather than the Ordinary Least Square (OLS) loss function, since QR is more resistant to outliers. Meanwhile, the model then evaluated and the best method for predicting PM10 concentration was selected based on the lowest error and highest accuracy values.