One main asset of Cloud-Radio Access Network (C-RAN) lies in its centralized architecture that allows network operators to serve dynamic flows of mobile traffic with efficient utilization of baseband resources and lesser operation costs than the distributed RAN architecture. For this very reason, the implementation of online resource allocation algorithms in the BaseBand Unit (BBU) pool for handling loads of multiple Remote Radio Heads (RRHs) is one of the most motivating challenges in C-RAN. Those centralized algorithms must be able to handle efficiently interference between users, as well as to dynamically select RRHs that can be turned on/off based on traffic variation. By doing so, the total RRHs transmission power can be minimized and the number of active BBUs within the cloud can also be reduced. In this paper, the issues of dynamic wireless resource allocation, transmission power minimization and BBU-RRH assignment in downlink C-RAN are addressed in one framework. We have previously attempted to address these problems by proposing a approach based on the branch-and-cut algorithm to solve small instances of the problem to optimality. However, due to the combinatorial complexity of the problem, finding optimal solutions for a large-scale network may take a fair amount of time and will not be suitable for online optimization. Towards this end, we propose a novel two-stage approach to address these issues for a large-scale problem. The first stage is a new proposal that addresses the problems of dynamic resource allocation and power minimization in C-RAN using a simulated annealing approach with a specific neighborhood search program. The BBU-RRH assignment is handled in the second stage using a multiple knapsack formulation. Through extensive event-based simulations, our proposal achieves significant reduction in time complexity and yields near optimal performance compared to state-of-the-art methods.