A simulation model based on Dymola modelling was developed to investigate the dynamic characteristics of automatic generation control (AGC) for variable-load thermal power units in this study. Specifically, a 300 MW unit from a power plant in northern China was used to verify the model’s validity in steady-state processes and to analyze the behavior of the main thermal parameters under different rates of load changes. The economic performance of the unit under different rates of load changes is also analyzed by combining the economic indexes of “two regulations” in the power grid. Results indicate that as the rate of load changes increases, boiler output, main steam temperature, reheat steam temperature, main steam pressure, and working temperatures of various equipment fluctuate more intensely. Specifically, at a rate of load reduction of 2.0% Pe MW/min, the maximum deviation of the main steam temperature can reach 7.6 °C, with the screen-type superheater experiencing the largest heat exchange. To achieve a balance between safety and economics for the unit, the rate of load raising should not exceed 1.2% Pe MW/min, and the rate of load reduction should not exceed 0.8% Pe MW/min. This paper applies the covariance index and AGC assessment index of the thermal power unit load control system to the established dynamic simulation model to supplement the AGC assessment index in the “two regulations”, and to provide a flexible and reasonable system evaluation result for field operators to refer to, so as to improve the economics of the system on the basis of safety.

Measuring the nitrogen oxides concentration accurately at the inlet of the selective catalytic reduction denitrification system plays an important role in controlling the nitrogen oxides concentration for coal-fired power plants, and a coupling relationship exists between nitrogen oxides concentration and multiple operational variables. Here, a modeling method based on feature fusion and long short-term memory network is proposed to mine the spatial and temporal coupling relationship between input variables for improving the prediction accuracy. First, the collected data were converted to image-like sequences. Then, the high-dimensional features of image-like sequences were fused by a convolutional neural network, and the spatial coupling features among the variables were mined. Finally, the constructed fusion features were input into the long short-term memory network to further explore the time coupling characteristics among the variables and complete the prediction of nitrogen oxides concentration at the inlet of the selective catalytic reduction denitrification system. The simulation results show that the prediction error of nitrogen oxides concentration at the inlet of selective catalytic reduction denitrification system based on CNN-LSTM model is 15.15% lower than that of traditional LSTM model.