A Domain-Knowledge-Guided Modular Neural Network Framework for Gas Turbine Performance and Emission Prediction

Abstract

Gas turbine performance and emission prediction are essential for condition monitoring, fault detection, predictive maintenance, and regulatory compliance, but conventional black-box models provide limited interpretability and poor subsystem-level traceability. This study proposes a domain-knowledge-guided modular neural network framework that decomposes gas turbine behavior into physically meaningful modules trained individually and integrated into a system-level model. Domain knowledge is embedded through modular decomposition, structured information flow, and physically meaningful inter-module connections, enabling measurement-level interpretation, module-level traceability, and modular updating. The framework is evaluated using two case studies. The framework is general at the modelling-procedure level, but the specific modular decomposition is case-specific and must be redefined according to the engine architecture, available measurements, measurement resolution, and target outputs. A high-bypass turbofan dataset, representative of a Pratt & Whitney PW-4056-like configuration, assesses detailed component-level modelling under a controlled single-realization measurement-noise perturbation. Under the controlled perturbation, the framework achieved a maximum mean absolute percentage error of approximately 2.62%, indicating stable prediction for the tested sensor-noise case. A real industrial predictive emission monitoring dataset evaluates adaptability using a reduced section-level structure based on available ambient, process, and emission measurements. Although initial integration showed cascading error, sequential fine-tuning reduced downstream emission mean absolute error by approximately 10.6% for CO and 39.2% for NOx. The fine-tuned framework matched monolithic Artificial Neural Network accuracy while providing intermediate predictions and subsystem-level error traceability, supporting interpretable gas turbine digital twins.