Ms Atefeh Bosak, Dr Zahra Hejazizadeh, Dr Akbar Heydari Tashekaboud,
Volume 0, Issue 0 (3-1921)
Abstract
Air pollution has significant impacts on human health, environmental quality, and the sustainable development of cities. This study aimed to evaluate PM10 using meteorological data from the city of Ahvaz through statistical methods and artificial neural networks. Daily meteorological data and air quality control station data for 4485 days (from 2011 to 2023) were obtained from the National Meteorological Organization and the Khuzestan Department of Environment. Initially, the data were processed and refined, and their normality was assessed using the Kolmogorov-Smirnov test. Given the non-normality of the data, Spearman's and Kendall's Tau-b methods were employed to examine their correlations. The time series and statistical information of the data were obtained using Python programming language. Furthermore, to predict future PM10 levels, the Multilayer Perceptron (MLP) neural network method was utilized. The results of these analyses indicated a significant correlation between meteorological variables and PM10. The Spearman and Kendall Tau-b correlations showed that PM10 had a positive and significant correlation with wind speed (0.094 and 0.061) and temperature (0.284 and 0.187) at a 99% confidence level. Conversely, PM10 exhibited a negative and significant correlation with visibility (-0.408 and -0.300), wind direction (-0.048 and -0.034), precipitation (-0.159 and -0.125), and relative humidity (-0.259 and -0.173) at the 99% confidence level. For future PM10 predictions, the MLP neural network was used. The model was of the Sequential type with an input layer consisting of 6 neurons, three hidden layers of Dense type with 16, 32, and 64 neurons, and an output layer with a linear activation function. The mean squared error (MSE) for the training set was 0.0034, and for the validation data, it was 0.0012. For the test set, the obtained validation accuracy was mse_mlp=0.0048 and val_loss=0.0012. The results indicate a significant direct or inverse correlation between meteorological data and PM10. Additionally, the outcomes of the MLP neural network demonstrated that the network provided satisfactory performance and acceptable predictions for PM10 data in Ahvaz.
1 Somayeh Mehrabadi,
Volume 21, Issue 60 (3-2021)
Abstract
The classical methods, also known as hard methods, are based on the accuracy of calculations, while the real world is founded on the inaccuracy of boundaries and the uncertainties, which is more consistent with soft computing methods. Each of these methods has its own strengths and weaknesses, and the hybridization theory was introduced to solve these problems. In the hybridization theory, which is also called intelligent hybrid systems, two or more single intelligent methods are combined to eliminate or rectify the shortcomings and limitations of single methods. In this study, forest degradation was modeled by employing the single-perceptron neural network and hybrid neuro-fuzzy method. For this purpose, the images from Landsat-5 TM sensor in 1999 and Landsat 8 OLI sensor in 2017 were utilized. Then, the degraded and non-degraded forest areas were sampled in 200 locations. Seven factors identified as the most effective factors in forest degradation, including the distance from the features like city, river, village, sea, and road, elevation and slope were measured for the 200 locations. The mean squared error (MSE) was used to evaluate the performance of models, which was 0.0535, 0.0704, and 0.0908 for the perceptron neural network in the Levenberg-Marquardt, Bayesian regularization, and scaled conjugate gradient algorithms, respectively. Also, the MSE value for the neuro-fuzzy model in the optimization and hybrid algorithms was 0.0190 and 0.0102, respectively. The analysis of the results showed the optimal performance of the neuro-fuzzy method both in reducing the error and in generalizing the model. Relying on the uncertainty rule, the neuro-fuzzy model provides the conditions that are closer to reality and have been more successful than the perceptron model at selecting the appropriate data.
