The invention provides a power meteorological numerical weather forecasting method and device based on multi-space-time scales. Firstly, the collected multi-source heterogeneous power meteorological observation data are fused. Then the mesoscale numerical weather forecast model is obtained. Finally, probability forecast, climate forecast, short-term and impending forecast and refinement forecast are carried out for electric power meteorological elements. Based on the fused electric power meteorological data, based on the mesoscale weather forecast model, an in-depth learning method, a method for real-time four-dimensional data assimilation and a fast cycle update method, a horizontal mesh resolution optimization method, a perturbation method of mode physical parameters, a continuous integration method are adopted and can realize the short-term, short-term, fine-grained, probability forecast and climate forecast, and can support multi-simulation function of new energy resources assessment, short-term and ultra-short-term power forecast, lightning, rainstorm, gale and other disasters early warning of electric power and meteorological numerical forecast, effectively improve the accuracy of meteorological forecast data, and comprehensively improve the support level of meteorological service power grid.