PREDICTION OF USER MOBILITY FOR HANDOVERS IN 5G NETWORKS IN JAKARTA USING THE LONG SHORT-TERM MEMORY METHOD

• Main Author: Timoteo, Adriel
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85100
• Institution: Institut Teknologi Bandung

Since the initial launch of 5G networks in 2018, the technology has rapidly expanded and now covers about 40% of the world's population. 5G offers higher bandwidth, faster connectivity, and lower latency compared to 4G. To achieve this high bandwidth, 5G operates at higher frequencies, which, although they carry more data, have smaller coverage areas and are easily obstructed. This results in 5G users experiencing more handovers compared to 4G, making handover performance crucial in a 5G network. One way to improve handover performance and efficiency is by predicting user movement using algorithms such as Markov Chain, Hidden Markov Model, and machine learning methods like Support Vector Machine (SVM), XGBoost, Deep Neural Network (DNN), and Long Short-Term Memory (LSTM). Developments in LSTM algorithms have consistently shown that they are excellent for predicting time-related data. Therefore, the LSTM algorithm can be highly efficient for predicting user movement for handover purposes. The goal of this research is to create an LSTM model for multi-step-ahead prediction, which will take 60 seconds of historical data and predict 10 seconds ahead (x = 60, y = 10). This model aims to achieve an average distance between actual and predicted data of less than 20 meters and latency of less than 100 ms. To develop this model, subsystems such as Plotly and TensorBoard will be used to display the map results and model training metrics. Meanwhile, the dataset used for model training is the Grab-Posisi dataset, and the machine learning framework used is TensorFlow, both of which are part of the prediction subsystem. This combination of subsystems is chosen for its simplicity in implementation and relatively high efficiency. Six model variations are built and trained: models with 16 units in 1 layer, 32 units in 1 layer, 64 units in 1 layer, 128 units in 1 layer, 32 units in 2 layers (32 + 32), iv and 64 units in 2 layers (64 + 64). All variations are trained for 1000 epochs, but there is also a variation with 16 units trained for 3106 epochs. Metrics such as MAE, MSE, RMSE, and average difference in meters (avg_m_diff) of each model are measured to find the best performing model. Additionally, the model results will be mapped to visualize the predicted user trajectory and compared to the actual user trajectory. The tests revealed that the model with 16 units trained for 3106 epochs (model_16_lr001_v2) is the most accurate, achieving an accuracy of 16.99 meters with a latency of 31 ms. This research demonstrates that the LSTM method can accurately and quickly predict user movement for handover management in a 5G network. In the future, the accuracy of the model can be improved by increasing the training epochs for each model variation. Models with 32 and 64 neurons are believed to have the same or greater potential compared to the model with 16 neurons if trained for more epochs.