The day-ahead power forecasting of wind cluster is one of the important foundations for provincial level or above power dispatching and control centers to make power generation plan and promote wind power consumption. Day-ahead wind power forecast （next 0 to 24 hours） is essentially to build mapping model between numerical weather prediction and actual power. Fully mining the deep mapping relationship between them is an important way to improve the accuracy of wind power prediction. In this paper, grid numerical weather prediction is used to establish wind cluster power forecasting models by using residual network, which mines the correlations between three-dimensional grid meteorological distribution and power. Case studies show that the proposed method is better than the existing mature methods in terms of progressiveness and adaptability.