PERFORMANCE ANALYSIS OF SUPPORT VECTOR REGRESSION (SVR) TO PREDICT CONGESTION

• Main Author: Irzavika, Nindy
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/41360
• Institution: Institut Teknologi Bandung

Traffic congestion is unavoidable, especially in big cities in Indonesia. Traffic prediction is the solution to reduce congestion. In this study prediction of traffic conditions is carried out using the Support Vector Regression (SVR) method. SVR is a Support Vector Machine (SVM) used for regression. The use of SVR in this research is because GPS data is non-linear data and time-series data. The steps taken to predict congestion are preprocessing using map matching, determining congestion parameters, predicting congestion using SVR, evaluating the performance of the SVR model using k-fold cross-validation, and determining the level of congestion using Fuzzy Comprehensive Evaluation (FCE). Map matching is used because there is GPS data that does not represent the actual position and not on the road. This research aims to determine the SVR performance for prediction of congestion. The data are taxi trip data obtained from one of the taxi companies in Bandung and Bandung city road data obtained from OpenStreetMap. Scenarios of testing performed are testing the performance of the SVR model by calculating MAF values and RMSE values, testing the performance of the SVR model by applying k-fold cross-validation, testing the accuracy of the SVR model, and testing by comparing the results of predictions with actual conditions. MAE and RMSE are metrics for measuring the accuracy of a variable. MAE measures the average error in a prediction model without considering its direction. RMSE measures the average error between predictions and actual observations. vi Based on experiments conducted, SVR kernels that have the best accuracy are RBF kernels and polynomial kernels, but RBF2 kernels and linear kernels have the smallest MAE and RMSE error values compared to other kernels. Therefore, the test results in this study indicate that the SVR model has a good performance for prediction of congestion, this is indicated by the high accuracy value of the model and the small error value and prediction results that are the same as the actual conditions.