Proposes a data -driven “LSTM-Transformer-SHAP” hybrid model framework for dynamic operational anomaly detection, diagnosis, and performance optimization of combined gas -steam cycle unit heat recovery steam generators. First, a Long Short -Term Memory (LSTM) network model is used to accurately predict turbine power and steam heat separately; then, a transformer adaptive reconstruction algorithm is applied to identify anomalies in the steam cycle operational state, and STL decomposition residuals clarify the type of anomaly; next, SHAP values are used to locate high-contribution parameters; finally, case analysis is conducted to estimate macro uncertainty and formulate feedforward optimization strategies. The results show that the LSTM model predicts flue gas h eat and steam heat with mean absolute percentage errors (MAPE) of 2.0% and 1.2% respectively; the LSTM model predicts turbine power and steam temperature with MAPEs of 1.5% and 0.3% respectively; SHAP attribution analysis yields a MAPE of 0.21% for heat re covery steam generator steam temperature. After optimizing the performance trend anomalies of the heat recovery steam generator , the reheat steam temperature and high -pressure main steam temperature can be increased by an average of 2.30° C and 4.68° C respe ctively under steady -state conditions, and steam cycle output can be increased by an average of 2.96 MW under steady-state conditions.