[Objective] Nitrogen oxide(NOx) emissions are a crucial environmental indicator for gas turbine combustors

and their accurate and rapid prediction plays a vital role in combustion optimization and low-emission control. To address the high computational cost of traditional numerical simulations and the limited generalization capability of single data-driven models

as well as the respective shortcomings of neural networks and ensemble learning methods in tabular data processing and small-sample generalization found in existing research

this study proposes a stacked ensemble learning framework based on heterogeneous base learners and a TabNet meta-learner

aiming to integrate the advantages of multiple models to enhance prediction accuracy and cross-condition adaptability. [Methods] The framework employs four heterogeneous base learners—XGBoost

Support Vector Machine(SVM)

Random Forest(RF)

and Multi-Layer Perceptron(MLP)—which represent four distinct learning paradigms

namely Boosting

kernel methods

Bagging

and neural networks

to generate initial predictions from perspectives of different bias-variance trade-offs and feature representations. Subsequently

TabNet

a model specifically designed for tabular data

is introduced as the meta-learner to leverage its sequential attention mechanism for instance-wise feature selection

dynamically integrating the base learner outputs. During the model training process

a Bayesian optimization strategy is adopted for efficient hyperparameter tuning of the key base learners

utilizing the root mean square error from 5-fold cross-validation as the objective function to avoid local optima. [Results] Evaluated on a preprocessed experimental dataset from a micro-mixing combustor

the proposed stacked ensemble model demonstrates significantly better prediction performance than all individual base learners

achieving a root mean square error(RMSE) and mean absolute error(MAE) as low as 0.784 0 mg/m³ and 0.173 0 mg/m³

respectively. Cross-condition validation

using dynamic data for training and steady-state data for testing

shows that the model exhibits good generalization capability

with an RMSE of 3.647 1 mg/m3 and an MAE of 1.426 2 mg/m3

indicating strong generalization performance. Furthermore

interpretability analysis based on SHapley Additive exPlanations(SHAP) indicates that the primary and secondary fuel flow rates are the most critical features affecting NOx prediction. This finding aligns closely with the physical mechanism wherein the temperature in the primary combustion zone

influenced by staged fuel combustion

governs thermal NOx formation

thereby enhancing the physical plausibility of the model predictions. [Conclusion] The proposed stacked ensemble learning framework can effectively improve the prediction accuracy and generalization capability for NOx emissions from combustors. Its architecture

which integrates heterogeneous learners with an attention-based meta-learner

provides a high-precision and interpretable data-driven solution for combustion emission prediction. The identification of key features is consistent with the physical mechanism of staged fuel combustion

demonstrating promising potential for engineering applications.