The Belgian frequency containment reserves market has recently been opened to assets with limited energy reservoirs. Such assets must implement an energy management strategy to deal with energy deviations. This paper is a contribution to the definition of such a strategy. A model for predicting grid-frequency deviations is built. Based on yearly series of grid-frequency measurements, seasonal trends in grid frequency are put in light. After having removed these trends, the remaining autocorrelation is explained through an ARMA-GARCH model. The overall model can be used to forecast average frequency deviations over 15-minute time steps, with any confidence interval. It is a tool to help dispatchers decide when to source energy on real-time energy markets.