Mobile data traffic is steadily growing all over the world, as a consequence of the great proliferation of mobile devices that has characterized the past decades. Based on mobile traffic volumes and mobile subscribers projections for the coming years, internet network infrastructures will be put under heavy load, leading Internet Service Providers (ISPs) and Mobile Network Operators (MNOs) into an increasing effort to improve network dimensioning, resources allocation and load balancing. It has been observed that due to time-varying patterns of cellular traffic network, resources management suffers of high inefficiencies which can benefit from an accurate prediction of traffic volumes. This can be achieved leveraging the enormous amount of data that mobile networks constantly generate, e.g., through Base Stations. On the other hand, extracting valuable knowledge from such a massive amount of data is far from trivial. In this context, researchers demonstrated a strong correlation between traffic peaks in Italian cities mobile networks and the corresponding city team playing a football match, and proposed a Median Week Signature(MWS)-based method to improve traffic peaks forecasting leveraging this information. In this work the aim is to understand whether deep learning based models can offer a valid alternative to build a traffic signature when trained on historic traffic volumes data, hence on a univariate dataset, but also if they are capable of learning and forecast traffic peaks in a more independent way, with a training based on both historic traffic volumes and exogenous information of football games presence, hence on a multivariate dataset. Therefore, we trained and tested the models over a purposely generated synthetic time-series resembling a real mobile traffic trace, adding an artificial football league calendar to it. Comparing the LSTM models based on the univariate dataset to MWS, both adopted as traffic signature methods and leveraging football games dates knowledge, the former improved the MAE by ~8.6% on the latter, and median relative error on peaks estimation by ~46.9%. Instead, comparing the predictions of the two LSTM models based on the univariate and multivariate datasets, the former yielded the best MAE improving by ~10,6% on the latter, while the latter obtained better relative error on peaks estimation with a median ~59.8% smaller than the former.
Il traffico dati mobile è in costante crescita in tutto il mondo, come conseguenza della grande proliferazione di dispositivi mobili che ha caratterizzato gli ultimi decenni. Sulla base delle proiezioni sui volumi di traffico mobile e degli abbonati di telefonia mobile per i prossimi anni, le infrastrutture di rete internet saranno sottoposte a un forte carico, spingendo i fornitori di servizi internet (ISPs) e gli operatori di rete mobile (MNOs) ad un crescente sforzo per migliorare il dimensionamento della rete, l'allocazione delle risorse e il bilanciamento del carico. È stato osservato che a causa dei mutamenti nel tempo dei patterns del traffico cellulare, la gestione delle risorse soffre di elevate inefficienze che possono beneficiare di una accurata previsione dei volumi di traffico. Ciò può essere ottenuto sfruttando l'enorme quantità di dati che le reti mobili generano costantemente, ad esempio attraverso le stazioni base. D'altra parte, estrarre conoscenze preziose da una quantità così massiccia di dati è tutt'altro che banale. In questo contesto, i ricercatori hanno dimostrato una forte correlazione tra i picchi di traffico nelle reti mobili delle città italiane e la corrispondente squadra cittadina che gioca una partita di calcio, proponendo un metodo basato sulla Median Week Signature (MWS) per migliorare la previsione dei picchi di traffico sfruttando tale informazione. In questo lavoro l'obiettivo è capire se i modelli basati sul deep learning possono offrire una valida alternativa per costruire una firma del traffico se addestrati sui dati storici dei volumi di traffico, quindi su un dataset univariato, ma anche se sono in grado di apprendere e prevedere picchi di traffico in modo più autonomo, con un addestramento basato sia sui volumi di traffico passati che su informazioni esogene riguardo la presenza di partite di calcio, quindi su un dataset multivariato. Abbiamo pertanto addestrato e testato i modelli su una serie temporale sintetica appositamente generata, avente similarità con una reale, aggiungendovi un calendario artificiale del campionato di calcio. Confrontando i modelli LSTM basati sul dataset univariato con MWS, entrambi adottati come metodi di firma del traffico e sfruttando la conoscenza delle date delle partite di calcio, il primo ha migliorato il MAE di ~8.6% sul secondo, e l'errore relativo mediano sulla stima dei picchi di ~46,9%. Confrontando invece le previsioni dei due modelli LSTM basati sui dataset univariati e multivariati, il primo ha prodotto il miglior MAE migliorando del ~10,6% sul secondo, mentre quello multivariato ha ottenuto un miglior errore relativo sulla stima dei picchi con una mediana del ~59.8% minore rispetto al primo.
Cellular networks downlink traffic forecasting: a deep learning approach based on football matches
Cecchetti, Francesco
2021/2022
Abstract
Mobile data traffic is steadily growing all over the world, as a consequence of the great proliferation of mobile devices that has characterized the past decades. Based on mobile traffic volumes and mobile subscribers projections for the coming years, internet network infrastructures will be put under heavy load, leading Internet Service Providers (ISPs) and Mobile Network Operators (MNOs) into an increasing effort to improve network dimensioning, resources allocation and load balancing. It has been observed that due to time-varying patterns of cellular traffic network, resources management suffers of high inefficiencies which can benefit from an accurate prediction of traffic volumes. This can be achieved leveraging the enormous amount of data that mobile networks constantly generate, e.g., through Base Stations. On the other hand, extracting valuable knowledge from such a massive amount of data is far from trivial. In this context, researchers demonstrated a strong correlation between traffic peaks in Italian cities mobile networks and the corresponding city team playing a football match, and proposed a Median Week Signature(MWS)-based method to improve traffic peaks forecasting leveraging this information. In this work the aim is to understand whether deep learning based models can offer a valid alternative to build a traffic signature when trained on historic traffic volumes data, hence on a univariate dataset, but also if they are capable of learning and forecast traffic peaks in a more independent way, with a training based on both historic traffic volumes and exogenous information of football games presence, hence on a multivariate dataset. Therefore, we trained and tested the models over a purposely generated synthetic time-series resembling a real mobile traffic trace, adding an artificial football league calendar to it. Comparing the LSTM models based on the univariate dataset to MWS, both adopted as traffic signature methods and leveraging football games dates knowledge, the former improved the MAE by ~8.6% on the latter, and median relative error on peaks estimation by ~46.9%. Instead, comparing the predictions of the two LSTM models based on the univariate and multivariate datasets, the former yielded the best MAE improving by ~10,6% on the latter, while the latter obtained better relative error on peaks estimation with a median ~59.8% smaller than the former.| File | Dimensione | Formato | |
|---|---|---|---|
|
2023_05_Cecchetti.pdf
non accessibile
Dimensione
3.27 MB
Formato
Adobe PDF
|
3.27 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/10589/210900