Abstract: Under the deep peak regulation background, thermal power units encounter such problems as slow load variation rate and inability to respond promptly to peak regulation requirements. To improve the unit's variable load rate, an operational optimization method for coal-fired power units based on a hybrid random forest-particle swarm optimization algorithm incorporating an attention mechanism (AM-RFR-PSO) was proposed. Firstly, correlation analysis was used to select the operation parameters related to unit output power as the model feature variables. Then attention mechanism was introduced to improve the random forest algorithm, and the improved algorithm was used to model the output power prediction of thermal power unit and compared with other algorithms. Finally, particle swarm algorithm was used to optimize the main feature variables to improve the load variation rate of the thermal power unit. The results show that compared with other algorithms, the AM-RFR prediction model has the smallest root mean square error, average absolute error, and the highest regression coefficient, which proves that the prediction model has high prediction accuracy, and the PSO-optimized operating parameters can improve the unit's variable load rate by close to 5%, which can effectively improve the unit's response speed.