NOx emission prediction is important for efficient boiler production and waste control. An adaptive data-driven modeling method is proposed to predict the boiler NOx emissions dynamically. In this method, a linear combination kernel is presented to improve the prediction accuracy of least-square support vector machine. The parameters of the kernel are optimized adaptively by a particle swarm optimization algorithm. Additionally, an adaptive moving time window strategy is presented to maintain model performance. The computational results based on the practical data illustrate that the proposed kernel and the adaptive moving time window strategy are positive and the proposed prediction method is superior to some previous prediction methods.