Localised Near Horizon Predictive Models of Cellular Load

As mobile technologies continue to mature network providers are experiencing ever increasing demands on network resources. This trend will continue for a range of reasons, from growing subscriber expectations to the network being viewed as an enabling technology for the Internet of Things. However, these changes pose significant challenges to network operators at a time when many are facing stagnant or falling Average Revenue per User (ARPU). To provide increased services with reduced costs, network operators are looking to improvements in technology such as Software Defined Networking (SDN) and Self Organising Networks (SON). Several of these techniques will become key components of future 5G networks. With growing network complexity and reduced revenue to hire staff, many of these advanced management techniques will benefit from detailed predictive models of network load to allow for the preallocation of network parameters and resources. This thesis uses anonymised Call Detail Records (CDR) from Meteor, a mobile network provider in the Republic of Ireland, to model network load and investigate how it can be serviced more efficiently. The Meteor network under investigation has over 1 million customers, which represents approximately a quarter of the state’s 4.6 million inhabitants. The main contributions of this thesis are 1. A novel methodology to predict near horizon traffic loads in practical spatially contiguous coverage regions. 2. A novel application of near horizon localised prediction models to the problem of self-organising green networks. 3. Empirically created foundational models of how the network experiences load. 4. A novel examination of causal influences on network load, spatial relationships, communication distances, load predictability, and load usage. 5. A range of novel algorithms and techniques from novel metrics for measuring load prediction performance to novel algorithms for estimating subscriber areas of interest, CDR feature extraction, CDR data cleaning, load visualistation etc. Results from this thesis show that there is a significant underutilisation of network resources. It is demonstrated that sufficiently accurate predictive models of network load are attainable at useful levels of spatial aggregation. These models are applied to the problem of self-organising green networks and demonstrate that a substantial reduction of network resource underutilisation is possible