We consider the large-scale power plant maintenance scheduling and production planning problem submitted by EDF to the 2010 Euro/Roadef Challenge. Electricity demand must be met by the combined production of nuclear power plants and thermic plants. The latter can operate continuously while the former have to be shut down regularly for refuelling and maintenance, and cannot produce during outage periods. In this model, we introduce uncertainty about the duration of maintenance operations: we assume that the durations of outages can be extended beyond their normal durations. The objective is to minimize the worst-case total power production cost. In the context of a rolling horizon approach, we investigate the impact of different hypotheses on the uncertainty set on the cost of the solutions and on the difficulty to obtain good solutions. Our solver uses an extended formulation of the problem, which is solved by column generation.