The cement calcination process suffers serious pollutant emission problems, especially nitrogen oxide (NOx) emission. Traditional methods mainly used physical modeling to optimize the operation of a cement calcination process for NOx reduction. However, physical modeling of NOx emission in the rotary kiln is too complicated because of the difficulties with determining model parameters. To address this challenge, the present study proposes the use of data-driven modeling and model-based real-time cement calcination process optimization (RTO) for reducing NOx emission. The Gaussian mixture regression model based on an improved just-in-time learning is introduced for modeling NOx in the kiln tail. Data preprocessing based on multivariate empirical mode decomposition is used, and an improved similarity strategy taking into account time-space information is utilized for local learning relevant sample selection to address problems associated with nonlinear and time-varying characteristics. The RTO problem is solved by a particle swarm optimization algorithm, and optimized set values of decision variables are obtained. Finally, the proposed modeling and optimization approach are applied to practical cement calcination process data. The result shows that the proposed modeling strategy has better performance than the traditional strategy, and the optimization approach performs better than previous conditions in calcination NOx emission reduction.