DEVELOPMENT OF USER THROUGHPUT-DOWNLINK PREDICTION SYSTEM IN 4G LTE NETWORK USING MACHINE LEARNING METHOD

• Main Author: Hanif Naufal Eka Wiratama, M
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64069
• Institution: Institut Teknologi Bandung

With advances in wireless network technology and adoption of fourth-generation (4G) long term evolution (LTE) networks over the last few decades, the capabilities and services provided to end-users seem to be endless. Of course, smartphone users who take advantage of this high-speed network service expect to have high-quality and consistent connections. However, due to the huge load demands on the cellular network due to skyrocket increase in technology and increasing number of users, users often experience unexpected variations in connection quality. To deal with such variations and maintain consistent connections, we need to predict these variations before they occur. This can be achieved by analyzing network quality parameters and investigating the main factors affecting network performance and network QoS. According to previous research, accurate user throughput can significantly improve bandwidth utilization. Therefore, it is important to predict network throughput based on other network parameters. This is a necessary initial step before work such as predictive resource allocation and implementation of QoS strategies on customers that network providers can undertake. In this final project, the author study two techniques for throughput prediction: predictive modeling with regression and time series forecasting. For predictive modeling with regression, the author apply various machine learning models to the collected data for throughput prediction and achieve the highest prediction performance with the Random Forest model. For time series forecasting, the author uses statistical methods and deep learning architecture. The author's evaluation shows that the time series forecasting technique has a higher throughput prediction performance than the machine learning model. Then the author has analyzed the effect of different data preparation techniques such as PCA and Feature Selection on the accuracy and duration of modeling on the data using the Random Forest algorithm. Data preparation techniques can reduce modeling time with a duration of three times faster than without processing. In addition, to make it easier for users to take advantage of the developed model, the author creates a web-based dashboard app that can be accessed from anywhere on the internet.