A performance comparison study on PM2.5 prediction at industrial areas using different training algorithms of feedforward-backpropagation neural network (FBNN). (March 2023)

• Record Type: Journal Article
• Authors: Chinatamby, Pavithra; Jewaratnam, Jegalakshimi
• Is Part Of: Chemosphere. Volume 317(2023)
• Publication Date: 2023-03
• Subjects: Air pollution -- Prediction -- PM2.5 -- Artificial neural network -- Feedforward-backpropagation -- Training algorithms; Pollution -- Periodicals; Pollution -- Physiological effect -- Periodicals; Environmental sciences -- Periodicals; Atmospheric chemistry -- Periodicals; 551.511
• DOI: 10.1016/j.chemosphere.2023.137788
• Languages: English
• ISSNs: 0045-6535

Abstract

Abstract: Presence of particulate matters with aerodynamic diameter of less than 2.5 μm (PM2.5 ) in the atmosphere is fast increasing in Malaysia due to industrialization and urbanization. Prolonged exposure of PM2.5 can cause serious health effects to human. This research is aimed to identify the most reliable model to predict the PM2.5 pollution using multi-layered feedforward-backpropagation neural network (FBNN). Air quality and meteorological data were collected from Department of Environment (DOE) Malaysia. Six different training algorithms consisting of thirteen various training functions were trained and compared. FBNN model with the highest coefficient correlation (R 2 ) and lowest root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) were selected as the best performing model. Levenberg Marquardt (trainlm) is the best performing algorithms compared to other algorithms with R 2 value of 0.9834 and the lowest error values for RMSE (2.3981), MAE (1.7843) and MAPE (0.1063). Highlights: Feedforward-Backpropagation Neural Network (FBNN) is used to predict the PM 2.5 concentration. The performance of different FBNN training algorithms were compared and evaluated. Levenberg-Marquardt algorithm ( trainlm ) is the most reliable prediction model with R 2 of 0.9834 and RMSE of 2.3981. Gradient Descent algorithm ( traingd ) is the most least performing model with R 2 of 0.56 and RMSE of 11.0385.