Virtual Machines (VM) offer data center owners the option to lease computational resources like CPU cycles, Memory, Disk space and Network bandwidth to end-users. An important consideration in this scenario is the optimal usage of the resources (CPU cycles, Memory, Block I/O and Network Bandwidth) of the physical machines that make up the cloud or 'machinefarms'. At any given time, the machines should not be overloaded (to ensure certain QoS requirements are met) and at the same time a minimum number of machines should be running (to conserve energy). The loads on individual VMs residing on these machines is, in fact, not absolutely random. Certain patterns can be found that can help the data center owners arrange the VMs on the physical machines such that both of the above conditions are met (minimum number of machines running without any being overloaded). In this work we propose a framework, PoWER that tries to intelligently predict the behavior of the cluster based on its history and then accordingly distributes VMs in the cluster and turns off unused Physical Machines, thus saving energy. Central to our framework are concepts of Chaos Theory that make our framework indifferent to the type of loads and inherent cycles in them as opposed to other current prediction algorithms. We also test this framework on our testbed cluster and analyze its performance. We demonstrate that PoWER performs better than another FFT-based time series method in predicting VM loads and freeing resources on Physical Machines for our test loads.