The invention discloses a deep learning-based power grid failure detection method, which comprises the following steps: (1) common point voltage signals are subjected to multi-resolution empirical mode decomposition; (2) phase space reconstruction is carried out by using the decomposition signals in the first step, each layer of phase space matrix is subjected to singular spectral decomposition toobtain a singular value of each layer of phase space matrix, the singular spectral entropy of each layer of phase space is extracted according to the singular value of each layer, multiple layers ofsingular spectral entropy are combined to form a depth neural network parameter feature vector of the multi-resolution singular spectral entropy; and (3) the feature vector in the second step is directly substituted into the depth neural network model for power grid failure detection. The feature vector extracted by the method has stability for similar signals and the same kind of samples, and plays a good role in distinguishing failure and disturbance states, reducing misjudgment in a disturbance state and avoiding malfunction of a photovoltaic system.