[Objective] With the increasing penetration of renewable energy

combined cycle power units are frequently operated for grid peak shaving. Affected by the strong nonlinearity and time-delay characteristics of the heat exchange process in heat recovery steam generators(HRSG)

traditional monitoring methods hardly support early fault detection of HRSG units. Accordingly

this paper proposes a data-driven hybrid model of LSTMTransformer-SHAP to realize dynamic anomaly detection and feedforward optimization for HRSG operational conditions. [Methods] Based on approximately 120 000 groups of operating data collected from an F-class combined cycle unit

the modeling and verification procedures are carried out. Firstly

an LSTM model is adopted to accurately predict turbine power

flue gas heat and steam heat

and obtain time-series deviation features. Secondly

an attentionbased encoder-decoder reconstruction algorithm is applied to identify temporal anomalies

and STL decomposition is used to distinguish different anomaly types via residual analysis. Furthermore

the random forest algorithm combined with the SHAP framework is employed for cross-verification and high-contribution parameter tracing. Finally

targeted optimization strategies for early fault suppression are formulated. [Results] The MAPE values of the LSTM model for flue gas heat and steam heat prediction are 2.0% and 1.2%

respectively

while the MAPE values for turbine power and steam temperature prediction are 1.5% and 0.3%. The random forest diagnosis model achieves a MAPE of only 0.21%. After feedforward optimization of HRSG performance anomalies

under steadystate conditions

the average reheat steam temperature and main steam temperature increase by 2.30 ℃ and 4.68 ℃

and the steam cycle output increases by 2.96 MW. [Conclusion] The proposed model establishes a complete closedloop framework covering prediction

diagnosis

tracing and optimization

which realizes earlier warning than traditional threshold alarm methods. The proposed feedforward optimization strategy can effectively restrain unit performance degradation and improve the operational safety and economy of combined cycle units under peak shaving conditions.