PREDICTION OF AVERAGE TEMPERATURES USING ARTIFICIAL NEURAL NETWORK METHODS: THE CASE OF GAZIANTEP PROVINCE, TURKEY

Abstract

The purpose of this study is to develop a model for monthly prediction of average temperatures in Gaziantep based on geographical variables and meteorological data from 1960 to 2014, provided by the Turkish State Meteorological Service. The input variables used in the study are year, month, latitude, longitude, altitude, minimum and maximum temperatures, while the output variables are average temperatures. In order to predict average monthly temperatures, three different artificial neural network models were developed, namely the Multi layer Perceptron (MLP), Cascade ANN, Elman ANN and multiple linear regression (MLR) method, and the accuracy of the predicted average monthly temperature values were compared with the measured values. The determination coefficient (R-2) value was calculated as 99.90 percent for the Multi layer Perceptron analysis, 99.88 percent for the Cascade and Elman ANN analyses and 99.65 percent for the MLR analysis. Additionally, the mean absolute percentage error (MAPE) values were found to be 2.380 percent for the Multilayer Perceptron, 2.391 percent for the Cascade ANN, 2.586 percent for the Elman ANN and 3.978 percent for the MLR. Similarly, the Mean Squared Error (MSE) value - as a further measure of accuracy was calculated to be 0.081 for the Multilayer Perceptron, 0.094 for the Cascade ANN, 0.116 for the Elman ANN and 0.272 for the MLR. It was determined that the monthly average temperature predictions derived from the Multilayer Perceptron were more accurate than those derived from the other analysis, and it was observed that the predictions derived from the ANN were observed to be closer to the actual monthly average temperature values than those derived from the MLR. The developed model is recommended for use in the prediction of average temperatures.