The invention relates to a black-start partitioning method, in particular, a semi-supervised spectrum clustering-based black-start partitioning method. An existing partitioning method strongly relies on the initial value of an algorithm and is difficult to obtain a global optimal solution. The method provided by the technical schemes of the invention comprises the following steps that: 1) a value is assigned to the weight of a power system line; 2) an adjacency matrix W and a diagonal matrix D are calculated; 3) a unit grouping model is established, and the unit grouping model is solved, so that unit grouping information is obtained, and the adjacency matrix W is modified; 4) a normalized Laplacian matrix is obtained; 5) eigenvectors corresponding to the first k smallest eigenvalues of the Lrw are calculated, and the eigenvectors form a column matrix H; 6) clustering calculation is performed; and 7) whether the active power generation capacity and load of each region are basically balanced is judged, when the imbalance of the active power generation capacity and load of a region exceeds a set value, nodes on boundaries are re-divided. The semi-supervised spectrum clustering-based black-start partitioning method provided by the technical schemes of the invention can recognize sample spaces of arbitrary shapes, can converge at a global optimal solution and is suitable for partitioning problems.