Long-term network traffic prediction is essential for mobile network operators (MNOs) to better maintain their networks. Network traffic peaks are more significant than normal traffic since they have a direct impact on the network's stability. These irregular peaks can disrupt operations, making it even more important for MNOs to precisely predict such events to reduce any problems and ensure uninterrupted service. In addition, a growing number of people are using online services, and the introduction of high-speed networks like as 5G and 6G has led to a significant rise in traffic and data generation, which has put the speed and accuracy of conventional prediction methods to the question. Several strategies for predicting network traffic have recently been developed, all of which use machine learning. However, because these peaks are caused by external factors that differ greatly from normal traffic patterns, machine learning models continue to struggle to effectively predict them. Furthermore, regulations on privacy restrict access to high-quality and large data sets, complicating the training process for these models. Moreover, privacy restrictions impose limitations on the availability of high-quality and large data sets, therefore complicating the training work for these models. This thesis proposes an AI-based model to address the challenges of long-term network traffic peak forecasting. By using multi-head attention mechanisms and generative adversarial networks (GANs), the model addresses the limitations of traditional methods and offers a more effective approach to predicting traffic peaks. This study also explores the effects of adding event-driven knowledge—specifically, football events— and GAN-generated time-series-based network data to the model to improve its peak forecasting ability, while addressing the mentioned challenges. To evaluate the proposed solution, a real-world dataset has been used along with different training scenarios. Comparing the results against the MWS baselines shows a reduction of MAPE by 3.8%, RMSE by 16.2%, maximum peak error by 27.2%, and mean peak error by 39.1% by adding synthetic data and football events. It thus indicates a potential for better accuracy in the network peak traffic prediction through incorporating football events and synthetic data into the AI-based models.
La previsione a lungo termine del traffico di rete è essenziale per gli operatori di rete mobile (MNO) al fine di mantenere al meglio le loro reti. I picchi di traffico di rete sono più significativi rispetto al traffico normale poiché hanno un impatto diretto sulla stabilità della rete. Questi picchi irregolari possono interrompere le operazioni, rendendo ancora più importante per gli MNO prevedere con precisione tali eventi per ridurre eventuali problemi e garantire un servizio senza interruzioni. Inoltre, un numero sempre crescente di persone sta utilizzando servizi online e l'introduzione di reti ad alta velocità come il 5G e il 6G ha portato a un significativo aumento del traffico e della generazione di dati, mettendo in discussione la velocità e l'accuratezza dei metodi di previsione convenzionali. Recentemente sono state sviluppate diverse strategie per la previsione dei picchi di traffico, tutte basate sull'uso del machine learning. Tuttavia, poiché questi picchi sono causati da fattori esterni che differiscono notevolmente dai normali schemi di traffico, i modelli di machine learning continuano a faticare a prevederli efficacemente. Inoltre, le normative sulla privacy limitano l'accesso a set di dati di alta qualità e di grandi dimensioni, complicando il processo di addestramento di questi modelli. Questa tesi propone un modello basato sull'intelligenza artificiale per affrontare le sfide della previsione a lungo termine dei picchi di traffico di rete. Utilizzando meccanismi di multi-head attention e reti generative avversarie (GAN), il modello affronta le limitazioni dei metodi tradizionali e offre un approccio più efficace per la previsione dei picchi di traffico. Questo studio esplora inoltre gli effetti dell'aggiunta di conoscenze basate su eventi, nello specifico eventi calcistici, e di dati di rete generati da GAN basati su serie temporali, per migliorare la capacità del modello di prevedere i picchi, affrontando al contempo le sfide menzionate. Per valutare la soluzione proposta è stato utilizzato un set di dati reali insieme a diversi scenari di addestramento. Il confronto dei risultati con i valori di riferimento MWS mostra una riduzione del MAPE del 3.8%, dell'RMSE del 16.2%, dell'errore massimo di picco del 27.2% e dell'errore medio di picco del 39.1% grazie all'aggiunta di dati sintetici con eventi calcistici. Ciò indica un potenziale per una maggiore accuratezza nella previsione dei picchi di traffico di rete attraverso l'incorporazione di eventi calcistici e dati sintetici nei modelli basati sull'intelligenza artificiale.
Long-Term Network Peak Traffic Forecasting Using Generative Adversarial Networks and Multi-Head Attention Mechanisms
HAMZENEJADI, SAJAD
2023/2024
Abstract
Long-term network traffic prediction is essential for mobile network operators (MNOs) to better maintain their networks. Network traffic peaks are more significant than normal traffic since they have a direct impact on the network's stability. These irregular peaks can disrupt operations, making it even more important for MNOs to precisely predict such events to reduce any problems and ensure uninterrupted service. In addition, a growing number of people are using online services, and the introduction of high-speed networks like as 5G and 6G has led to a significant rise in traffic and data generation, which has put the speed and accuracy of conventional prediction methods to the question. Several strategies for predicting network traffic have recently been developed, all of which use machine learning. However, because these peaks are caused by external factors that differ greatly from normal traffic patterns, machine learning models continue to struggle to effectively predict them. Furthermore, regulations on privacy restrict access to high-quality and large data sets, complicating the training process for these models. Moreover, privacy restrictions impose limitations on the availability of high-quality and large data sets, therefore complicating the training work for these models. This thesis proposes an AI-based model to address the challenges of long-term network traffic peak forecasting. By using multi-head attention mechanisms and generative adversarial networks (GANs), the model addresses the limitations of traditional methods and offers a more effective approach to predicting traffic peaks. This study also explores the effects of adding event-driven knowledge—specifically, football events— and GAN-generated time-series-based network data to the model to improve its peak forecasting ability, while addressing the mentioned challenges. To evaluate the proposed solution, a real-world dataset has been used along with different training scenarios. Comparing the results against the MWS baselines shows a reduction of MAPE by 3.8%, RMSE by 16.2%, maximum peak error by 27.2%, and mean peak error by 39.1% by adding synthetic data and football events. It thus indicates a potential for better accuracy in the network peak traffic prediction through incorporating football events and synthetic data into the AI-based models.| File | Dimensione | Formato | |
|---|---|---|---|
|
2024_10_Hamzenejadi.pdf
non accessibile
Dimensione
4.37 MB
Formato
Adobe PDF
|
4.37 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/227259