The characterization of the spectral energy distribution (SED) of dust emission has become a critical issue in the quest for primordial B-modes. The dust SED is often approximated by a modified black body (MBB) emission law but the extent to which this is accurate is unclear. This paper addresses this question, expanding the dust SED at the power spectrum level. The expansion is performed by means of moments around the MBB law, related to derivatives with respect to the dust spectral index. We present the mathematical formalism and apply it to simulations and Planck total intensity data, from 143 to 857 GHz, because no polarized data are yet available that provide the required sensitivity to perform this analysis. With simulations, we demonstrate the ability of high-order moments to account for spatial variations in MBB parameters. Neglecting these moments leads to poor fits and a bias in the recovered dust spectral index. We identify the main moments that are required to fit the Planck data. The comparison with simulations helps us to disentangle the respective contributions from dust and the cosmic infrared background to the high-order moments, but the simulations give an insufficient description of the actual Planck data. Extending our model to cosmic microwave background B-mode analyses within a simplified framework, we find that ignoring the dust SED distortions, or trying to model them with a single decorrelation parameter, could lead to biases that are larger than the targeted sensitivity for the next generation of CMB B-mode experiments.