A deep learning approach to model daily particular matter of Ankara: key features and forecasting

Akbal Y., Ünlü K.

International Journal of Environmental Science and Technology, vol.19, no.7, pp.5911-5927, 2022 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 19 Issue: 7
  • Publication Date: 2022
  • Doi Number: 10.1007/s13762-021-03730-3
  • Journal Name: International Journal of Environmental Science and Technology
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), Biotechnology Research Abstracts, CAB Abstracts, Compendex, Environment Index, Geobase, INSPEC, Pollution Abstracts, Veterinary Science Database
  • Page Numbers: pp.5911-5927
  • Keywords: Convolution neural networks, Feed-forward neural networks, Gated recurrent neural networks, Hybrid neural networks, Long short-term memory neural networks, Particulate matter
  • TED University Affiliated: No


In this study, three different goals are pursued. Firstly, it is aimed to model particulate matter (PM) of Ankara, the capital of Turkey, by utilizing hybrid deep learning methodology. To do this, five different methodologies are proposed in which four of them are hybrid methods. Three different evaluation criteria as coefficient of determination (R2), mean absolute error (MAE) and mean squared error (MSE) are used to compare the proposed methods. In the test set, the hybrid model which consists of feed-forward neural networks, convolution neural network and long short-term neural networks has the best performance with R2 of 0.81, MSE of 73.07 and MAE of 5.6. Secondly, PM levels are categorized to form a prediction challenge in accordance with the World Health Organization standards. The particulate matter level is divided into two categories as being low or not, being moderate or not and being dangerous or not, it is shown that the proposed hybrid model which has the highest performance on forecasting, also worked perfectly in the classification task with accuracy of 94%. Finally, the effect of different pollutants and meteorological variables on the prediction of PM is investigated by employing ensemble machine learning methodology of random forest regression, extra tree regression and multiple linear regression. According to the results of the analysis, it is shown that the most important predictor variables of PM are its own lagged values, other pollutants, earth skin temperature and the wind speed.