Abstract: Developing an efficient prediction model for NO_x generation is of great significance for reducing NO_x emissions and denitrification costs in coal-fired units. The NO_x model is built based on related variables and depends on the design of the model structure, with the parameters of the model structure referred to as hyperparameters. Setting these hyperparameters appropriately can substantially enhance the accuracy and generalization capabilities of NO_x prediction models. This paper presents a NO_x generation prediction model based on tree-structure parzen estimator optimized long short-term memory neural network (TPE-LSTM). Using historical operational data from a 330 MW coal-fired unit, model structural parameters are combined with time series data window length and the number of principal components of NO_x generation related variables, thus creating a new type of hyperparameter. The improved hyperparameters are then optimized to construct NO_x generation prediction model based on long short-term memory (LSTM) neural networks. Comparing the proposed hyperparameter optimized NO_x prediction model with the unoptimized LSTM model and the typical optimization algorithm particle swarm optimization optimized LSTM (PSO-LSTM) model, the prediction results reveal that the TPE-LSTM prediction model demonstrates superior accuracy and generalization abilities.