To further optimize and enhance the regulation capability of thermal power units, this study employs factor analysis to investigate relevant influencing factors. Building upon this foundation, it integrates LSTM networks and Bayesian optimization theory to propose a PA-mBO-LSTM prediction model for thermal power unit response to AGC regulation. This model is based on feature extraction and multi-level deep learning. The findings indicate that “rapid response capability,” “control stability,” “equipment health,” and “cooperative optimization” are the four primary factors influencing thermal power unit regulation capability, with corresponding importance values of 0.343, 0.325, 0.248, and 0.196, respectively, exhibiting a decreasing order of influence. Validation using actual operational data from an 800MW unit demonstrates that the proposed model’s prediction deviation remains within 15MW. Its prediction accuracy significantly outperforms models without feature extraction and single LSTM models, proving the model’s capability to precisely evaluate thermal power units’ response to AGC regulation under deep peak shaving conditions.
