A SPATIO-TEMPORAL NEURAL NETWORK APPROACH FOR PREDICTING TRAFFIC ACCIDENT HOTSPOT AREAS: A CASE STUDY IN BANDUNG CITY

• Main Author: Wiwit Nugroho, Hermawan
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79485
• Institution: Institut Teknologi Bandung

Traffic accidents are the most significant cause of human death after disease, with the majority of victims being of productive age. In order to decrease the number of traffic accidents, it is crucial to identify where and when they occur most frequently. Previous research stated that the Support Vector Regression (SVR), Random Forest, and Decision Tree methods were quite good in predicting traffic accident cases. However, this method actually still has limitations in capturing spatial dimensions in datasets which have an important role in the case of predicting traffic accidents. The Spatio-Temporal Neural Network (STNN) approach is a reasonably new yet superior method for modeling space-time data. However, this method needs to be improved to deal with the limitations of spatial datasets related to traffic accidents in Bandung. In this research, the STNN method was compared with traditional Machine Learning to find out the most suitable method for producing predictions of traffic accidents in the city of Bandung. This research also proposes to use the Getis Ord Gi* statistical approach in data processing to improve STNN prediction performance. The research results show that STNN consistently has superior performance compared to traditional Machine Learning for predicting traffic accident hotspots in Bandung City. Using the Getis Ord Gi* statistical approach in data processing can also improve STNN performance by reducing the RMSE value by 0.00624 on the Diskominfo dataset and by 0.00386 the RMSE value on the Satlantas dataset.