The invention discloses a weather sensitive load power estimation method based on stacked auto-encoders (SAE). The method comprises the steps of adding a multi-layer full-connection layer to the output end of an SAE model, and building a weather sensitive load power estimation model based on the SAE; and extracting a dimensionality reduction characteristic of a daily load curve by utilizing unsupervised learning of the SAE, and training the full-connection layer by utilizing a weather sensitive load power curve as a labeled sample, so that the mapping from the dimensionality reduction characteristic of the daily load curve to weather sensitive load power is formed in the full-connection layer. The estimation model provided by the method can directly obtain the weather sensitive load powercurve from the daily load curve, so that the method is particularly suitable for the situation that weather data is frequently lost during actual application. The SAE can extract the dimensionality reduction characteristic of the daily load curve in an unsupervised manner, and the number of input neurons of the full-connection layer is greatly reduced, so that the network parameters of the full-connection layer are greatly reduced, and the training difficulty of the model is remarkably lowered.