1. Title, Journal and Authors
Title: Neural network prediction model for fine particulate matter (PM2.5) on the US-Mexico border in El Paso (Texas) and Ciudad Jua"L rez (Chihuahua)
Journal: Environmental Modelling & Software
Authors: J.B. Ordieresa,*, E.P. Vergaraa, R.S. Capuzb, R.E. Salazar
2. Summary
Introduction
In this paper, the authors tried to predict PM2.5 concentration at the border between the United States and Mexico using Multilayer Perceptron (MLP), Square Multilayer Perceptron (SMLP), and Radial Basis Function (RBF). We also compared the performance of the three to identify potential uses or the strengths and weaknesses of each neural network. The reason for studying the concentration of PM2.5 is to find out the source of anthropogenic air pollution.
Method
Data from 2000 and 2001 were used as training data, and data from 2002 were compared with the predicted data. The PM2.5 concentration for the remaining 16 hours was predicted using data of 24 hours/8 hours prior to the start of prediction. Also, the wind speed, direction, humidity and temperature values for 8 hours were used.
The existing ARIMA linear modeling tool used artificial neural networks (ANNs) because air pollutants are not as reliable as the tendency to not behave in a linear manner. In order to carefully appraise the need for complex models like the neural networks, their outcomes have been compared against the ones pertaining to the traditional models. Thus, a persistence model and a linear regression have been assessed.
We used our own Linux-based software tool and R for data preprocessing and conditioning, and several Linux-based NODELIB libraries were used for neural network processing.
Seven nodes in the input layer of the neural network and one node in the output layer were set, and the amount of neurons in the hidden layer was estimated through experimental analysis of the error. The MAE index was used to determine the number of nodes in the hidden layer.
The accuracy of the prediction was expressed by RMSE and R square.
For Multilayer Perceptron (MLP) I used aBPM (Back Propagation with Moment) as the learning rule. The number of nodes in the hidden layer has been tested from 1 to 30.
In the case of RBF, training was performed with various variance values.
Conclusion
When I used SMLP with many 20 hidden nodes to represent the actual model, the work was slow.
As a result of testing the number of nodes in the hidden layer, the stability reached was exceptional except for “like resonant” topologies.
In this study, RBF had better predictive performance and better stability than MLP. In addition, RBF demanded less training effort than MLP.
Add
Data on context and location per hour can be very important, so you need to train your network with a data set suitable for your study.
3. Originality and Creativity
4. Application to research
5. Contact
Yeong Gyu Gu / Intern student
Environmental Systems Engineering Lab.
School of Earth Sciences and Environmental EngineeringGwangju Institute of Science and Technology
1 Oryong-dong Buk-gu Gwangju, 500-712, Korea
Phone : +82-10-6589-6653
E-mail : kududrb1@naver.com