Evaluation of the Artificial Neural Network Performance in Estimating Rainfall using Climatic and Geographical Data (Case study: Fars Province)

Precipitation forecasting is essential in maintaining, managing, allocating, and distributing water resources, determining the volume of water facilities, supplying the crops' water requirement, and determining the amount of erosion and sedimentation. This research aimed to investigate the performance of artificial neural networks in predicting monthly rainfall based on climatic and geographic information such as minimum and maximum temperature, minimum and maximum humidity, precipitation, latitude and longitude, and altitude above sea level in 23 stations of the Fars province. The results on levels 5, 10, and 18 of input data showed that the model accuracy in estimating the monthly rainfall increases with the increment in the number of inputs. The most accurate neural network model is in the rank normalization method with one hidden layer, and also, the best network structures are 5-25-1, 10-30-1, and 18-20-1, respectively. The results indicated that the neural network with 18 inputs has the smallest MSE=0.032 and the highest R=0.62. The best normalization method is the ranking method with an optimal neural network of one layer and 5-25-1 structure, the Levenberg-Marquardt training algorithm, and the sigmoid tangent stimulus function. Therefore, by using an artificial neural network (with 18 introduced inputs), it is possible to predict the amount and distribution of monthly rainfall in a wide area with acceptable accuracy. This issue plays a very decisive role in the management and planning of drinking and agricultural water resources; by taking into account these forecasts, future policies can be planned to optimize costs and maximum productivity facilities.