The invention belongs to the large power grid stabilization and control field and mainly relates to a power grid fault critical clearing time discrimination method based on improved Softmax regression. The method is characterized by modifying a cost function of standard Softmax regression; using normal distribution taking correct classification as a center to be target probability distribution; taking a KL divergence between probability distribution acquired through model calculating and the target probability distribution as the cost function; and through reducing a normal distribution variance and carrying out continuous iteration, acquiring an optimal model.