A physics-constrained intelligent prediction model for gas turbine macro-performance in integrated energy system dispatch

As a key piece of equipment enabling cascade energy utilization in integrated energy systems, the gas turbine exhibits strong nonlinear characteristics during operation. To address the need for accurate prediction of gas-turbine macro-level energy-supply performance in the operation optimization of integrated energy systems, this study proposes a new operation-oriented gas-turbine performance prediction method, in which an intelligent forecasting algorithm is integrated with prior physical-law constraints to accurately characterize the key energy-output behavior of the gas turbine. Firstly, on the basis of the operating mechanism analysis of main gas-turbine modules, namely compressor, combustor and turbine, prior monotonic physical relationships are derived, including those between fuel-gas consumption and electricity generation, as well as between the IGV opening degree and waste-heat flue-gas mass flow rate. Subsequently, a customized regression layer is constructed, such that the physical constraints are innovatively embedded into the model-training process through the loss function; Meanwhile, the optimal combination of critical model parameters is determined by an improved Northern Goshawk Optimization algorithm (INGO) incorporating multiple enhancement strategies. Finally, a novel INGO-CNN-GRU gas-turbine prediction model incorporating prior physical-law constraints is established. Taking a gas turbine operating in a steel enterprise in China as the research subject, the comparative simulation experiments are conducted, by which the scientific validity and reliability of the proposed model are verified. The obtained results indicate that, compared with the other five benchmark models, proposed approach achieves the best predictive performance. For power generation, the MAE values are reduced by 5.79 MW, 0.14 MW, 0.23 MW, 0.37 MW, and 0.26 MW, respectively; for waste-heat flue-gas mass flow rate, the corresponding MAE reductions are 33.69 kg/h, 1 kg/h, 0.74 kg/h, 1.8 kg/h, and 0.18 kg/h, respectively. Further validation under unseen operating conditions demonstrates that the introduction of physical monotonicity constraints can effectively suppress prediction drift under out-of-training-domain conditions, thereby improving the prediction robustness and extrapolation stability of the model. Sensitivity analysis of the operating variables further shows a physically reasonable ranking of variable influences and confirms that the model preserves response directions consistent with the prior mechanisms for the two key relationships, namely fuel-gas consumption–power generation and IGV opening degree–waste-heat flue-gas mass flow rate. It is demonstrated that, under the physical monotonicity constraints, the dependence on data quality can be markedly reduced, the extrapolation capability and interpretability can be improved, and a new modeling paradigm for gas-turbine simulation can thereby be provided.