The precision of wind power forecast is very important in the selection of wind farm site， and in the integration and operation of power system with increasing penetration of wind power. Compared with static neural networks， two dynamic neural network models， locally recurrent time-delay neural network model and globally recurrent time-delay neural network model， are proposed for the forecasting of a wind farm output in order to simulate the time-series characteristic of the generation series. To demonstrate the effectiveness， the models are applied and tested on a wind farm located in the north of China. Base on numerical meteorological predictions， hourly forecasts up to 24 hours ahead are produced for the wind farm. Simulation results demonstrate that the dynamic neural network models outperform the static ones in the forecast of wind power with time-series characteristic.