Aiming at the existing statistical bias phenomenon and the limit of structure element selection of traditional mathematical morphological filter, this paper proposes a multilevel structure element, and the variable step length method is used to construct a kind of multilevel generalized self-adaptive morphological filter, improving the filtering accuracy while reducing the restrictions on structure element selection, and the use of the filter is more flexible. At last, the filter is used for the power signal denoising containing a variety of noise, and the simulation result shows the better filtering effect of multilevel generalized self-adaptive morphological filter compared with the traditional filter.