Rainfall Forecasting using Support Vector Regression Machines

Abstract

Heavy rainfall as a consequence of climate change have immensely impacted the ecology, the economy, and the lives of many. With the variety of available predictive tools, it is imperative that performance analysis of rainfall forecasting models is properly conducted as a measure for disaster preparedness and mitigation. Support Vector Regression Machine (SVRM) was utilized in predicting the rainfall of a city in a tropical country using a 4-year and 17-month rainfall dataset captured from an automated rain gauge (ARG) in Southern Philippines, involving parameter cost and gamma identification to determine the relationship between past and present values, determining optimal cost and gamma parameters to improve prediction accuracy, and forecasting model evaluation. The SVRM model that utilized Radial Basis Function (RBF) kernel function having the parameters of c=100; g=1; e=0.1; p=0.001 and the lag variable which used 12-hour report with lags up to 672-timesteps (i-672) demonstrated a Mean Square Error (MSE) of 3.461315. With close to accurate forecast between the predicted values and the actual rainfall values, the results of this study showed that SVRM has the potential to be a viable rainfall forecasting model given the proper data preparation, model kernel function selection, model parameter value selection and lag variable selection.