regression method

The prediction of the minimum temperature is one of the main approaches to manage and reduce the risk damage caused by frostbite. For this purpose, using daily statistics of minimum, maximum and dew point temperatures during the period of 2005 to 2009, the minimum temperature of the next day was predicted for four stations (Rasht, Kerman, Shiraz and Hamedan) with different climate types by applying the Linacre regression model and the Multi-Layer Perceptron artificial neural network (MLP) in the whole year, the cold period (from October to late March) and the warm period (from April to late September). For this aim, the Matlab-2015 and IBM SPSS-20 software were used and statistical criteria RMSE, NRMSE and R2 were applied to evaluate the performance of the models. The results of this study, in all three periods, demonstrated that the best estimate of the Linacre regression model was obtained with root mean square error of 1.70 and 2.44°C for the whole year, 2.01 and 2.32°C for the cold period and 1.51 and 2.24°C for the warm period for Rasht and Shiraz stations with the PH-C-W and SA-C-W, respectively. The best results from MLP neural networks with Levenberg-Marquardt algorithm, logic sigmoid transfer function in the hidden layer, the linear transfer function in the output layer and two inputs (dew point and maximum temperature), like the Linacre regression model, were obtained with RMSE of 1.57 and 1.93°C for the whole year and 1.61 and 1.8 for the cold period for Rasht and Shiraz stations, respectively. The RMSE of the best results from MLP neural networks in the warm period was 1.21 and 1.44°C for Rasht and Hamedan stations, respectively. To evaluate the role of relative humidity on model results, this parameter was added as a third input to the multi-layer Perceptron network. The improved RMSE for the whole year was 17.4, 12.9, 49.4 and 18.3 percent and for the cold period of the year was 3.4, 2.86, 9.48 and 15.83 percent at Kerman, Shiraz, Hamedan and Rasht stations, respectively. These values for the warm period were 6, 13.33, 2.86 and 18.63 percent for the above mentioned stations. These improved errors indicate that only at Hamedan station, the cold period of the year produces more improvement in error reduction than the warm period of the year by adding relative humidity as the third input to the neural network model. In other stations, adding relative humidity in the warm year has led to a reduction in the error rate. In general, it can be said that selected MLP networks had better performance than the Linacre regression model in predicting the minimum daily temperature.