The invention relates to a stacking auto-encoder method which can be applied to ultra-supercritical unit modeling. Under the condition that relatively complete data is selected in the unit, the methodsolves the problem that a traditional system identification method is difficult to establish an accurate analysis model under a big data background. Meanwhile, a tedious data preprocessing process isavoided; a maximum correlation entropy function is adopted as a cost function; the method comprises the following specific steps: setting an ultra-supercritical thermal power plant; extracting appropriate data from the ultra-supercritical unit data set to serve as an input and output sample set of model training; determining a network structure of the stacked auto-encoder, selecting an input dataset as input of a model, training each auto-encoder by adopting a gradient descent method, finely adjusting the whole network of the stacked auto-encoder by adopting the gradient descent method by utilizing input and output data, and carrying out inverse normalization on the data, so as to generate final output of the stacked auto-encoder.