Data Pre-Processing Effects on the Artificial Neural Network Performance to Predict Monthly Rainfall (Case Study: Abadeh County)

Since many time series are not normal, it is required to normalize data by transformation functions prior to any analysis and modeling. In this study, the next month rainfall of Abadeh County station was predicted using the average monthly rainfall, minimum and maximum temperatures and minimum and maximum humidity as inputs of MLP network, both normally and raw, at period 1976 to 2013. After scrutiny the existence or nonexistence of missing and outlier data, meteorological data were normalized using three normalization minimum-maximum, rank normalization and z- score. After obtaining the best network structure using try and error for each method, the minimum-maximum method with a three-layer network structure and 13 number of hidden layers of neurons chose as the best method with R=0.92 and MSE=0.13 compared to other methods. Also comparing the performance of the network in using raw and Pre-Processed data showed that Pre-Processing the data improved greatly network performance. The results of the sensitivity analysis showed the maximum sensitivity of model to remove maximum humidity parameter, and the second the maximum temperature had the greatest impact on precipitation forecast. Also comparing the performance of the network with the different numbers of inputs indicated that network with two inputs including minimum temperature and minimum humidity had good results (MSE = 0.13) compare with five inputs.