Traditional RNN and CNN models have the issue of the time?invariance problem when they are used to make short?term predictions of wind power on longer time scales. This paper proposed a short?term wind power uncertainty prediction method based on the time?variant deep feed?forward neural network architecture (ForecastNet) model. This method has a time?varying network structure to improve multi?step ahead prediction, and has an interlaced output capability to mitigate the gradient disappearance problem. The probability density distribution can be obtained by using mixture density network. This model not only avoids the cumulative error of recursive multi?step prediction in the traditional deep learning model, but also fully considers the correlation of wind power at adjacent moments. In the hidden layer of the model, the actual data of wind power from PJM network in the United States are used to test three kinds of neural network models, namely, fully connected network, convolutional network, convolutional network with attention mechanism. The wind power of the next 12 hours is predicted each time, and the range and probability density of wind power of the next 500 hours are obtained by rolling prediction. The results of the experimental simulations prove the effectiveness of the proposed prediction model.