With the continuous expansion of the power grid, it has become more and more difficult to perform unified voltage regulation on the entire grid. This paper proposes a fast reactive power and voltage partition method based on deep learning and an improved K-means clustering algorithm. First, the electrical coupling strength matrix is established to reflect the strength of the electrical coupling relationship between the nodes of the system. Then the sparse autoencoder in deep learning is used to realize feature extraction and dimensionality reduction of the input high-dimensional matrix through training. Finally, the improved K-means clustering algorithm is used to perform cluster analysis on the feature sequence after dimensionality reduction, and the final partition is determined by checking the electrical modularity value. The quality of power grid divisions is evaluated with two evaluation indicators: electrical modularity and reactive power reserve verification. The simulation analysis of IEEE39 and IEEE118 bus systems verifies that the proposed method has high electrical modularity on the basis of ensuring connectivity and sufficient reactive power reserve. This work is supported by the National Key Research and Development Program of China (No. 2018YFB0905105).