Loading...

Modeling Gaseous Air Pollutants Concentration in Tehran Using Artificial Neural Network and Land Use Regression

Mirzaee, Mohsen | 2020

566 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 53246 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Mohammad Arhami; Amini, Zahra
  7. Abstract:
  8. In this thesis the hourly concentration of different gaseous air pollutants in Tehran is modeled using Land Use Regression (LUR) and Artificial Neural Network, separately. Both models are provided with the same set of input data; the first step is to find these data. Since traffic affects air pollution, information about traffic conditions is one of the main inputs in air pollution modeling. Therefore, to obtain traffic information, in this thesis, first a novel method is developed to extract and analyze Google Maps traffic data. In this method, image processing is used along with the Geographic Information System (GIS) to count the number of pixels of different traffic colors for each road in Tehran. Hourly weather data is another modeling input; it is taken from the Dark Sky database for 90 locations in Tehran. Other data required for modeling, such as information about land use, population, and hourly air pollutants concentration, are acquired from the relevant organizations. To find the optimal inputs for traffic and land use data, buffers with different radii are created around each air pollution monitoring station. In addition, trial and error is employed to find the optimal architecture of the neural network, because factors like the number of neurons and layers, learning rate, batch size, and the type of activation functions affect the output results. In modeling gaseous air pollutants concentration, the results show the superior performance of neural network over LUR, for the neural network model can detect complex, nonlinear relations between the input data and the concentration of pollutants. In both models, the most accurate results are for the Ozone pollutant (with R2 measure of 0.92 for the neural network model); this has to do, among other things, with considering the Sun’s ultraviolet index as an input in modeling this pollutant. On the other hand, the results for CO-pollutant (with R2 measure of 0.73 for the neural network model) are not as accurate as for other gaseous pollutants, which is probably related to insignificance of changes in the concentration level of CO-pollutant in the period under the study
  9. Keywords:
  10. Neural Network ; Ozone ; Traffic Data ; Air Pollution ; Gas Pollutant ; Geographic Information System (GIS) ; Linear Regression ; Land Use Regression (LUR)

 Digital Object List

 Bookmark

No TOC