Hybrid Modeling of Singular Spectrum Analysis and Support Vector Regression for Rainfall Prediction

Abstract

Rainfall can provide many benefits such as for agriculture, water resource management, and electricity. However, it can also cause hydrometeorological disasters such as floods and droughts, so that an accurate rainfall prediction is needed to anticipate the risks and minimize the losses. Rainfall characteristics are diverse, complex, and uncertain. Thus rainfall data are usually nonlinear time series and difficult to predict using traditional methods such as ARIMA. In recent years, machine learning models such as support vector regression (SVR) hybrid model has been developed to improve prediction accuracy. The SVR hybrid model can be carried out with Singular Spectrum Analysis (SSA) for data pre-processing. The SSA method is used to decompose original time series data into trend, oscillatory, and noise components. The SVR model is then used to predict rainfall based on reconstruction series from SSA without noise components. The grid search algorithm using an optimization method is used to estimate the parameters of the SVR model. This research aims to apply an SSA-SVR hybrid model and compares it to the SVR model using monthly rainfall data at Kemayoran Station from 1980 to 2019. Based on the result showed that the hybrid model yielded more accurate than the single model.