International Journal of Computer Applications (0975 – 8887) Volume 84 – No 9, December 2013 7 Rainfall Prediction using Neural Net based Frequency Analysis Approach Seema Mahajan L. J. Institute of Engineering & Technology, S.G.Road Ahmedabad-382210, Gujarat, India Himanshu Mazumdar Professor E.C and Head R&D Center D.D.University, Nadiad-387001, Gujarat, India ABSTRACT Rainfall prediction is very complex hydrologic process and is important as it holds the key to any countries’ economy. Proposed model presents a new approach for yearly rainfall prediction of 30 Indian subdivisions. Yearly rainfall data of the Indian subdivision is available from IITM, Pune. The combination of Fast Fourier Transform (FFT) and Feed Forward Neural Network (FFNN) is applied for next one year rainfall prediction. Fast Fourier transform with filtering is performed on interpolated rainfall data to separate periodic components. These periodic components and delayed periodic components are given as input and desired output respectively to an FFNN for training. While testing the output of FFNN, inverse FFT gives the predicted rainfall value by amount of training input-output delay. This model is tested with 140 year’s Indian subdivisions rainfall data. The experimental results of 30 subdivisions show that next one year rainfall prediction accuracy is above 92%. Keywords Rainfall prediction, Fast Fourier Transform, Feed Forward Neural Network 1. INTRODUCTION India’s agriculture production is highly dependent on its precipitation behavior of the monsoon rainfall. Monsoon is the main source of water. Average rainfall prediction is a prime important factor for crop planning. India receives 80% of rainfall during June to September. Monsoon follows cyclic behavior and has seasonality. As cycle period cannot be accurately predicted, this adds complexity of rainfall prediction. Many researchers worked on new strategies for rainfall forecasting. “Guhathakurta (2005)” constructed one year rainfall prediction model based on back propagation neural network for 14 districts of Kerala and also for Kerala as a whole subdivision. His results shows the root mean square error (RMSE) of area weighted values of district forecast is 6.26% and RMSE of Kerala as a whole subdivision is 13.29%. Rainfall prediction at up scaling of regions gives improved result than downscaling. “Somvanshi et al.” implemented mean annual rainfall prediction models over Hyderabad region of Andhra Pradesh, India. They compared two models based on ANN and autoregressive integrated moving average (ARIMA). RMSE of ANN and ARIMA model is 145.14 and 262.57 respectively. ANN model found superior to trace the nonlinearity of rainfall data. “Sarah et al.” developed a rainfall prediction model with the parameters like southern oscillation index (SOI), North Atlantic oscillation (NAO), sea level pressure (SLP), sea surface temperature (SST) and past month rainfall in Karoon basin. They have used artificial neural network along with fuzzy logic and wavelet functions. RMSE for two years, annual and six months is 6.22, 7.11 and 13.15 respectively. “Iyengar et al.” proposed rainfall forecasting model based on the rainfall data itself which is decomposable into six empirical time series. They used combination of ANN and simple regression technique to handle non linear and linear portion of rainfall respectively. The model efficiently predicts the rainfall with 75% to 80% of the interannual variability. In past years many researchers have introduced several new concepts and ideas in rainfall forecasting, but still rainfall prediction remains a challenge for researchers. 2. METHODOLOGY There is regional variation in rainfall distribution of India and separate rainfall prediction model for each subdivision is desired. Global climatic factors like, distance from sea, latitude the Himalayan Mountain, distribution of land and water, surface pressure and wind, upper air circulation and western cyclones are affecting Indian sub-divisional rainfall. Rainfall is complex functions of these parameters. Researchers are correlating these independent parameters to predict rainfall. These parameters are embedded in rainfall time series. Rainfall prediction in the proposed model uses effect of these parameters as part of rainfall time series. Monthly rainfall data of all subdivisions (regions) of India shown in Figure. 1 from 1871 to 2010 is available from IITM, Pune. We downloaded these data to train and validate the model. Yearly rainfall values are derived from the corresponding monthly rainfall values of each subdivision. An interactive user friendly application is developed using C # .Net tool to implement the proposed model. Figure 2 shows pictorial representation of the model. The proposed model is performed in three steps of Data Interpolation, Fast Fourier Transform and Feed Forward Artificial Neural Network. Each step is described below. 2.1 Data Interpolation Though 140 years rainfall data is available, 140 discrete points are not sufficient in time scale to develop a forecast model. Number of samples within the function and sample space are increased through data interpolation to get desired size of data set. Interpolation is performed using C# Drawcurve method. This method draws a cardinal spline passing through each point in the given array. This method draws a cardinal spline passing through each point of the given array with default tension value. 5601 data points are extracted from generated curve. 5601 is the empirical value. Each interpolated sample is equivalent to (140×365)/5600, 9.125 days. Each year is equivalent to 5600/140, 40 samples. Gujarat region interpolated data is shown in Figure 3.