Load Prediction in Communications Network Kongluan Lin QCIS UTS, Sydney, Australia linkongluan@gmail.com Ante Prodan QCIS UTS, Sydney, Australia anteprodan@gmail.com Simeon Simoff CM UWS, Sydney, Australia simoff@uws.edu.au John Debenham QCIS UTS, Sydney, Australia debenham.uts@gmail.com Abstract—The prediction of load in communications networks provides a scientific basis for conserving power that is currently attracting considerable interest. This paper reports the results of experiments to predict communications network load using real network data. The problem considered is simply to predict whether load is "generally" increasing, decreasing or unchanged in the immediate future. Various linear and non-linear predictive models were investigated. We conclude that the most efficient solution to the problem considered is a subtle combination of a linear regression model and a highly reactive model applied to the data aggregated at two levels of granularity. Keywords: load prediction, communications networks I. INTRODUCTION AND RELATED WORK Power conservation in both wireless and wired communications networks is a key issue particularly when low capacity batteries are deployed [6]. Recently there has been a substantial research effort with the aim of conserving power in communications networks, and many approaches have been proposed. Most approaches are based either on maximizing the number of stations that can be put into sleep mode, or on minimizing the power for transmission [5]. The cross-layer power management (CLPM) scheme regulates transmitting power based on the node density at the receiver node [8]. The receiver node obtains its node density information from network layer and then sends it to the transmitter. The transmitting node can then adjust its transmitting power such that only the required amount of power is used for communication, and interference is consequently reduced [7]. Lin et. al. propose a new energy-efficient dynamic power management (DPM) scheme, in which battery status, the probability of event generation and optimal geographical density control (OGDC) are taken into consideration to minimize the number of power-active nodes [4]. The Quorum-based Power Saving (QPS) protocols are used to conserve power in wireless networks, but it remains a challenge to apply it to the Mobile Ad hoc networks (MANETs) as the timers of nodes are usually asynchronous, the incurred delay are expected to be adaptive, and the network topology is asymmetric [8]. An Asynchronous, Adaptive, and Asymmetric (AAA) power management protocol proposed by Wu and Chen are able to fulfill the unique requirements of MANETs [8]. Power management protocols such as IEEE 802.11, PSM, S-MAC, SCP-MAC, and B-MAC conserve power by managing sleep and duty cycles. In a receiver-driven protocol such as PSM or SCP-MAC, the sleep schedule is synchronized, and nodes only exchange messages during their regular wakeup period. However, synchronization cost dominates power consumption when data traffic is low. In a sender-driven protocol, synchronization of the sleep schedule is not needed, but the sender needs to wake up the receiver each time that packets need to be sent, which causes a long delay [3]. Harvesting power from the environment is also suggested to prolong the life of battery-based communications systems [2]. However, different nodes in a network have different power harvesting opportunities, and an efficient scheme for assignment of workload needs to be developed that takes account of the battery status of each node [1]. Power can also be conserved if load can be accurately predicted. If each node has reliable forecasts of load then it may adjust its own power demands. This approach is effective if the load prediction method is simple, fast and requires low computational overheads. This paper addresses power consumption by predicting load using tools from data mining. First, data is collected and transformed. A variety of predictive models for network load were considered and evaluated. Finally, an efficient solution is proposed that is based on a subtle combination of a linear regression and a highly reactive model applied to the data that has been aggregated at two levels of granularity. II. MINING NETWORK LOAD Our overall aim is to identify a suitable architecture for load prediction in communications networks generally. To do this we based our experiments on Ethernet load data obtained from the University of Technology, Sydney (UTS). This data is characteristic of load data in communications networks. Our hypothesis is that the solution to the prediction of UTS Ethernet load will indicate the architecture for load prediction in communications networks. In each zone of UTS, network load is limited by factors including the number of students in it, the time of day and the available Internet bandwidth. The total network capacity is bounded by these factors. An agent is located in each zone. It records the network load in the zone over a period of time, and then analyses it to build a predictive model for that zone. Based on its predictive model, each agent can then forecast its zone’s network load. In this experiment, the predicted network load for each zone is compared with the real, observed network load, and