Evaluation of selected transfer functions of artificial neural network model for prediction of minimum temperature (Case Study: Sanandaj Station)

Short-term prediction of minimum temperature is important in mitigation chilling and frost injury in agriculture. In current study, the frequency of early autumn and late spring frosts in Sanandaj synoptic station, Iran were worked out. Then, using six variables, i.e. dry and wet bulb temperature, relative humidity, wind speed, wind direction, and cloud cover as the inputs, a multilayer perceptron artificial neural network model (MPL/ANN) based on the Levenberg-Marquardt training algorithm of MATLAB software package was applied for prediction of the minimum temperature for the next 3, 6, 9 and 12 hours ahead. The selected Transfer Functions were hardlims, logsig, polsin, radbas, satlins, softmax, tansig, and tribas. The statistical measures of MAD, MSD, RMSD, and R were used for comparisons. The results showed that in case of late spring frost, the poslin, logsig, tansig, and satlin functions in April with a correlation coefficient greater than 0.8 and error values of 1.17, 1.61, 1.88 and 2.00 (◦C) for the different times steps, respectively are the best options. Similarly, in October, the radbas, poslin, poslin, and tribas functions with a correlation more than 0.7 and error values of 1.60, 1.96, 1.99, and 1.36, were found to be the most suitable ones for prediction of the minimum temperature at 21:30, 00:30, 03:30 and 06:30 local time. Also, among the selected functions, the poslin with the highest frequency has the best performance in predicting nocturnal frosts in Sanandaj. The results confirmed the good performance of the ANN approach in short-term prediction of minimum temperature and frost occurrence in study region.