Prediction of Maximum Temperatures by Time Series and Artificial Neural Networks (Case Study: Isfahan Station)

Abstract

Due to climate changes, global warming, and the recent drought, forecasting, checking, and analyzing maximum temperatures as one of the foremost imperative climatic parameters allows planners to plan and provide the necessary arrangements. The main reasons for checking the temperature as a parameter influencing nature are agriculture, pests, diseases, melting ice and flooding, evaporation and transpiration, and drought. Today, artificial neural networks are used to predict time series like temperature because of their feature for understanding the random mechanism of fully nonlinear and complex series. This study used data from 1953 to 2005, two methods, and multi-layer perceptron artificial neural networks with the learning algorithm after the error propagation to analyze and check the monthly maximum temperature. This issue used an input layer, five hidden layers of TANSIG, and an output layer of the pure line for artificial neural networks. The mean squared error criterion was also used to assess the results. In the following study, 70% of the total data were used as training data (RMSE = 1.8622 and MSE = 3.4677); in order to avoid the phenomenon of the over-load network, 15% of the data were used for validation data (RMSE = 1.7667 and MSE = 3.1213). The remaining 15 percent has also been used to check and test data. (RMSE = 2.134 and MSE = 4.5538). A comparison of monthly maximum temperature forecast results for 1953 and 2005 with observed data shows good agreement of the model. The overall results indicate that approximately every 64 years will add a degree to the temperature.