Pm10 Concentrations Short Term Prediction Using Regression, Artificial Neural Network And Hybrid Models

• Main Author: Japeri, Ahmad Zia Ul-Saufie Mohamad
• Format: Thesis
• Language: English
• Published: 2013
• Subjects: TA1-2040 Engineering (General). Civil engineering (General)
• Online Access: http://eprints.usm.my/45032/1/Ahmad%20Zia%20Ul-Saufie%20Mohamad%20Japeri24.pdf; http://eprints.usm.my/45032/
• Institution: Universiti Sains Malaysia

Particulate matter has significant effect to human health when the concentration level of this substance exceeds Malaysia Ambient Air Quality Guidelines. This research focused on particulate matter with aerodynamic diameter less than 10 , namely PM10. Statistical modellings are required to predict future PM10 concentrations. The aims of this study are to develop and predict future PM10 concentration for next day (D+1), next two-days (D+2) and next three days (D+3) in seven selected monitoring stations in Malaysia which are represented by fourth different types of land uses i.e. industrial (three sites), urban (three sites), a sub-urban site and a reference site. This study used daily average monitoring record from 2001 to 2010. Three main models for predicting PM10 concentration i.e. multiple linear regression, artificial neural network and hybrid models were used. The methods which were used in multiple linear regression were multiple linear regression (MLR), robust regression (RR) and quantile regression (QR), while feedforward backpropagation (FFBP) and general regression neural network (GRNN) were used in artificial neural network. Hybrid models are combination of principal component analysis (PCA) with all five prediction methods i.e. PCA-MLR, PCA-QR, PCA-RR, PCA-FFBP and PCA-GRNN. Results from the regression models show that RR and QR are better than the MLR method and they can act as an alternative method when assumption for MLR is not satisfied. The models for artificial neural network show that FFBP is better than the GRNN. Hybrid models gave better results compared to the single models in term of accuracy and error. Lastly, a new predictive tool for future PM10 concentration was developed using ten models for each site with average accuracy for D+1(0.7930), D+2 (0.6926) and D+3 (0.6410). This application will help local authority to take proper action to reduce PM10 concentration and as early warning system.