The present invention discloses a wind power plant power multi-time scale joint prediction method. The method comprises the steps of acquiring the wind speed and the power historical data within the preset time of a whole wind power plant to obtain a historical wind speed time sequence and a historical power time sequence of the wind power plant; utilizing a multiwavelet packet decomposition technique to carry out the wavelet packet decomposition on the historical wind speed time sequence and the historical power time sequence separately to obtain the low frequency band, the intermediate frequency band and the high frequency band components of the historical wind speed time sequence and the historical power time sequence; utilizing an RBF neural network model to carry out the wind speed and power rolling prediction of a short-term wind power plant on the low frequency band and the intermediate frequency band components of the historical wind speed time sequence and the historical power time sequence separately; utilizing a plurality of least-square support vector machines to carry out the wind speed and power rolling prediction of an ultra short term wind power plant on the low frequency band, the intermediate frequency band and the high frequency band components of the historical wind speed time sequence and the historical power time sequence separately; and uploading the obtained wind speed and power rolling prediction data of the short-term wind power plant and the ultra short term wind power plant to an electric power regulation and control center.