Prediction of daily average PM10 concentrations using feedforward neural network in Kocaeli, northwestern Turkiye

This study suggested that artificial neural network models be created to assess daily average Particulate matter (PM10) concentrations for Kocaeli. Six models whose input parameters were meteorological data and particulate matter concentrations were created for four stations in Kocaeli. In addition, Kruskal-Wallis test was utilized to determine the robustness of the model's prediction. Developing the model was repeated for each station, and the accuracy and efficiency of the algorithms were assessed by contrasting the predicted outcomes with actual outcomes. Performance evaluations were made with root mean square error, mean absolute error, mean absolute percentage error, and correlation coefficient. The best root mean square error value in the feedforward neural network model was obtained for S2 station in model 6 for training. The root mean square error and correlation coefficient of model 6 for the test were 12.30 and 0.68, respectively. In addition, it is observed that the method's performance improves when PM10(t) is added to the network's input, including the meteorological variables. The performance of feedforward neural network models varies depending on the data, and no generalization can be made in the parameter selections. The findings of this study show that the prediction success of the concentration of PM10, which is a highly variable and noisy data, with the ANN model was found to be moderate.