During the period of the project (from 01/02/2020 to 31/01/2022), the work packages (WPs) have been implemented completely and the corresponding technical objects of the project have been achieved. The work performed and the results obtained so far are summarized as follows:
For WP1, we have investigated the spatial-temporal characteristics of traffic patterns generated by some representative mobile network services. We have analysed the auto-correlations and the spatial correlations of each individual mobile network service. Based on the service-level traffic pattern characteristics, we have designed prediction frameworks to forecast the future traffic loads of different mobile network services using the services' own previous traffic loads.
For WP2, we have investigated the traffic correlations among different mobile network services. We have analysed the conditional entropy of future traffic load distribution of each certain mobile network service when previous traffic loads of the other mobile network services are given. We have proven that the inter-service traffic correlations indeed exist and that the future traffic loads of the instant message services are highly influenced by the historical load distributions of the short message traffic generated in a target cell's adjacent cells. We have investigated how the mobile network subscribers will influence the traffic distributions of different mobile network services.
For WP3, based on the inter-service traffic correlations we found, we have proposed novel service-level traffic predicting frameworks, each of which can provide precise prediction results for the future traffic loads generated by a certain mobile network service based on the historical traffic loads generated by this mobile network service and some other mobile network services. We have investigated whether we can help the prediction model of a target traffic prediction task related to a certain mobile network service improve the prediction accuracy and learning efficiency by utilizing the experience of existing traffic prediction tasks related to diverse mobile network services. We have proposed a deep meta-learning based mobile network traffic prediction framework that provides the proper hyper-parameter and initial weight vector for the prediction model of a new prediction task based on the traffic pattern's characteristics in the frequency domain. Our experimental results demonstrated that with meta-learning, our framework can enormously elevate the performance of prediction models of various mobile network traffic prediction tasks.
We have created a website, which is regularly maintained and updated, to constitute the main public source of information about the research, dissemination, and outreach activities. We have published in high-impact scientific journals and international conferences, such as IEEE Wireless Communications, IEEE Wireless Communications Letters, IEEE GLOBECOM, and IEEE NOMS, to report our achievements about this project. Ranplan has facilitated participation at high-impact industrial events such as Mobile World Congress for industrial dissemination and creating business from the project. For the exploitation of the research results of this project, the service-level traffic prediction frameworks are being integrated into Ranplan’s data analytics tools and marketed by Ranplan. Futhermore, from the new knowledge and results gained within the project, Ranplan, the University of Cambridge, and the Alan Turing Institute have offered consulting and training services in mobile network traffic characterization and prediction to the corresponding network engineers.
