A soft sensor model based on compression sensing(CS) and least square support vector machine(LSSVM) was proposed to predict the NOx concentration in flue gas. The mapping matrix of the LSSVM was sparsified by the least support orthogonal matching pursuit algorithm(LS-OMP) in the compression sensing theory

and then used in soft sensor model building. Compared with the conventional LSSVM model

this model reduces the computational cost and improves the real-time performance by sparsifying the mapping matrix. Compared with the conventional sparse-LSSVM model

LSSVM model needs to make new support vectors sparse after modeling. The CS-LSSVM model reduces the computational cost and improves the soft-sensing accuracy through once compression. The model was used in soft sensor of NOx emission from a coal-fired boiler. The simulation results show that

only 50% support vectors is enough to achieve good performance with the model and provides data support for online monitoring of the NOx emission.