Optimasi Support Vector Regression Menggunakan Genetic Algorithm dan Particle Swarm Optimization Untuk Prediksi Curah Hujan Musim Kemarau

Abstract

Support Vector Regression (SVR) is Support Vector Machine (SVM) that is used for regression case. SVM is a set of classification and regression technique which is development of non-linear algorithms. Regression method has been commonly used for prediction models e.g. for seasonal climate prediction. SVR process requires kernel functions to transform the non-linear inputs into a high dimensional feature space. This research focused on predictive modeling rainfall in the dry season at 15 weather stations in Indramayu district. The basic method used in this study was SVR optimized by a hybrid algorithm GAPSO (Genetic Algorithm and Particle Swarm Optimization). This researchbegan by identifying and formulating problem. It then continued by data preprocessing i.e. collecting the dataand determining training data and test data. Data processing were performed by process data with SVR optimized by using GAPSO, testing, and the last steps wereanalyzing and evaluating the results predicted and observed values. In the collecting phase of the data,Indian Ocean Dipole (IOD) index known as the Dipole Mode Index (DMI) and NINO 3.4 Sea Surface Temperature Anomaly (SSTA) index data were used as predictors (independent variables) while dry season rainfall was used as response variables (dependent variables). The data used in this study were from 1978 to 2008. In SVR process, we used data of 20 years as datatrainingand data of 10 years as data testing. SVR kernel function used in this study was RBF kernel.SVR has cost and kernelfunctionparameter valuethat should be optimized i.e. C and γ, respectively. These values givebig impact on the SVR model. The more the optimal value of C and γ,the better the model. SVR was optimized by using the hybrid technique GAPSO to obtain optimal value. Thistechniqueincorporates concepts from GA and PSO and it creates individuals of new generation not only by crossover and mutation operation in GA, but also through the process of PSO. SVR model was obtained by using 24 populations, 100 times iteration for each GA and PSO, 10 iterations of GAPSO. It was designed by using C andγparameters, ranging from 0.1 to 50 and 0 to 10, respectively. This research obtained a model of SVR with the highest correlation coefficient of 0.87 and NRMSE error value of 10.43 at Tugu station. Cikedung station has the lowest NMRSE error value of 9.01 and the correlation coefficient of 0.78.