The accurate measurement of NOx concentration is the basis of accurate ammonia injection in selective catalytic reduction (SCR) system of thermal power unit. Excessive or too little ammonia injection will cause ammonia escape or exceeding the standard of NOx emission and cause environmental pollution. There exist a large measurement lag and errors during the purging process in the measurement of NOx concentration in the SCR system of thermal power units. Therefore, to enable the intelligent control of the SCR system, it is necessary to carry out soft sensing of NOx emissions. This article proposes a soft sensing algorithm for NOx emissions of thermal power units based on an adaptive Gaussian mixture model (GMM) two-step clustering (AGTSC) algorithm and ensemble learning. The AGTSC algorithm is used to softly divide the historical data into working conditions. Temporal pattern attention long short-term memory (TPA-LSTM) is also adopted to construct individual learners for each working condition. Here, we use the parameter regression algorithm to train the combiner based on the output of the individual learner and the membership signal of the working condition as the input of the combiner. The clustering algorithm is tested with three datasets. The results show that this algorithm addresses the shortcomings of GMM including not being applicable to the clustering of nonconvex datasets and manual selection of the number of clusters. The soft sensing algorithm is verified by using the historical data of the SCR system of 1000-MW thermal power units. The results show that the proposed method has higher accuracy and stronger generalization ability than the conventional method, thus providing an effective method for soft sensing of NOx emissions of thermal power units.