Performance Evaluation of Artificial Neural Network Models for Downscaling and Predicting of Climate Variables

Atmosphere–ocean coupled global climate models (GCMs) are the main source to simulate the climate of the earth climate. The computational grid of the GCMs is coarse and so, they are unable to provide reliable information for hydrological modelling. To eliminate such limitations, the downscaling methods are used. The present study is focused on simulating the impact of climate change on the behavior of precipitation and temperature of Sirjan synoptic station in Kerman Province. At first, the capability of artificial neural network to downscaling of climate variables that predicted by CanESM2 is tested. Then, using the most appropriate models, the mean monthly temperature and precipitation amounts forecast for future periods under RCP 4.5 scenario. Results of this study for monthly temperature downscaling indicated that the artificial neural network with 2 hidden layer, 8 neurons, with Tangent and Log sigmoid activation function was the best model, so that RMSE, NS and R2 were 0.387 , 0.973 and 0.917 respectively. Also, for precipitation variable, the structure with 2 hidden layer feed forward perceptron, 8 neurons, Tangent and Log sigmoid activation function and Levenberg-Marquardt algorithm had better performance, so that RMSE, NS and R2 were 2.867, 0.849 and 0.924, respectively. Results indicate that until 2099, amount of monthly mean temperature under RCP 4.5 emission scenario will be increased by 3 (˙C) and the highest increase is predicted for August by 3.9 (˙C) and a lower increase in April by 1.8 (˙C). The results also showed considerable increase of precipitation for June to November and noticeable decrease for March and May months. However, no change occure in annaul scale (inter-annual).