20190724_Forecasting PM10 levels using ANN and MLR: A case study for Sakarya City

1. Title, Journal and Authors

Title: Forecasting PM10 levels using ANN and MLR: A case study for Sakarya City

Journal: Global Nest

Authors: Ceylan Z.1,2,* and Bulkan S.1

1Marmara University, Faculty of Engineering, Industrial Engineering Department, 34772, Istanbul, Turkey

2Ondokuz Mayis University, Faculty of Engineering, Industrial Engineering Department, 55139, Samsun, Turkey


2. Summary

  MLR and ANN were compared to investigate the relationship between daily PM 10 levels and meteorological variables data for two years period, to predict one day daily PM 10 concentrations. As input data, there were six variables such as temperature, atmospheric pressure, wind speed, relative humidity, dew point, and visibility.


  MLR was based on linear squares and ANN was based on MLPNN. A Levenberg-Marquardt algorithm was used to update the weights and the number of epoch applied was 1000. Tangent sigmoid and purelin function were used for activation function and LEARNGD and LEARNGDM were used for learning function. A set of actual data consisting of 602 data sets obtained from the database were used for training(70%), validation(15%) and testing(15%). Matlab toolbox was used to calculate. RMSE, MAE, and R2 were used to evaluate the performance of ANN.


  As a result of MLR analysis, humidity was excluded since humidity had no significantly linear regression (p-value > 0.05). The R2 value was 0.3239, because the data set had complex and non-linear characteristics.

  A total of 36 models were performed to obtain the optimal result. As a result, MLP 6-12-1(6 inputs, 12 neurons, 1 output) with pair Tansig-Tansig and LEARNGDM had the best performance(RMSE was 15.700, MAE was 9.047 and R2 was 0.840). Accordingly, ANN was regarded as an adequate prediction model.


3. Application to research

- To adjust this method, I think some components are needed. In MLR, analyzing the coefficient of performance R2, VIF was needed. And to compare the weight of input variables, AIC was needed.


4. Contact

Dae Seong Jeong / Intern student


Phone : +82-10-2003-7860

E-mail : jeongds92@gist.ac.kr

첨부 (0)