The invention discloses a subspace identification based forecasting method for superheated steam output of a boiler of a firepower power station, which comprises the following steps: firstly, acquiring production data from plant-level DCS (Distributed Control System) configuration software at regular time; secondly, correcting bad points, filtering and denoising; thirdly, calculating input and output variables required by model identification; fourthly, constructing an Hankel matrix by utilizing continuous acquired and processed data and rolling to update; fifthly, approximating the boiler process to a linear state space model; sixthly, using a subspace theory to identify a matrix of the state space model [A, B, C and D]; seventhly, utilizing the latest input energy information as the obtained input variables of the state space model after the identification and calculating output energy of future moments; eighthly, calculating the mean value at the same time by utilizing a historical forecast value and a current forecast result and correcting by utilizing the past forecast deviation; and ninthly, rolling to update the data of identification of the model. The forecasting method is applicable to online forecast of superheated steam output energy in the boiler production process and provides guidance and reference for the optimization control of boiler combustion.