​In this paper, we address the optimal operation of a virtual power plant using stochastic programming. We consider one risk-neutral and two risk-averse formulations that rely on the conditional value at risk. To handle large-scale problems, we implement two decomposition methods with variants using single- and multiple-cuts. We propose the utilization of wind ensembles obtained from the European Centre for Medium Range Weather Forecasts (ECMWF) to quantify the uncertainty of the wind forecast. We present detailed results relative to the computational performance of the risk-averse formulations, the decomposition methods, and risk management and sensitivities analysis as a function of the number of scenarios and risk parameters. The implementation of the two decomposition methods relies on the parallel solution of subproblems, which turns out to be paramount for computational efficiency. The results show that one of the two decomposition methods is the most efficient.