Data-Driven Engine Modeling: A Case Study
Fadaeian, Yeganeh (2025)
2025
Automaatiotekniikan DI-ohjelma - Master's Programme in Automation Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
Hyväksymispäivämäärä
2025-11-17
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025111410647
https://urn.fi/URN:NBN:fi:tuni-2025111410647
Tiivistelmä
Accurate modeling of transient engine behavior is crucial for control and optimization in internal combustion engines. This thesis investigates data-driven approaches for predicting key dynamic variables—engine speed and charge air pressure—under transient operating conditions. The objective is to evaluate machine learning models that can reproduce nonlinear engine dynamics while maintaining suitability for control-oriented applications.
Engine simulation data were generated using the GT-SUITE platform, representing transient operation scenarios with varying control inputs. The data were preprocessed and divided into training, validation, and test sets. Two neural network architectures, a Nonlinear AutoRegressive model with eXogenous inputs (NARX) and a Multilayer Perceptron (MLP), were developed and trained in GT-POST. Model performance was assessed based on prediction accuracy, transient response, and generalization capability.
The results show that both models can approximate the dynamic behavior of the system with good accuracy, while the NARX model demonstrates superior performance in capturing transient responses. The study highlights the potential of data-driven modeling for control-oriented engine applications and emphasizes the importance of data quality, preprocessing, and network structure in achieving reliable predictive models.
Engine simulation data were generated using the GT-SUITE platform, representing transient operation scenarios with varying control inputs. The data were preprocessed and divided into training, validation, and test sets. Two neural network architectures, a Nonlinear AutoRegressive model with eXogenous inputs (NARX) and a Multilayer Perceptron (MLP), were developed and trained in GT-POST. Model performance was assessed based on prediction accuracy, transient response, and generalization capability.
The results show that both models can approximate the dynamic behavior of the system with good accuracy, while the NARX model demonstrates superior performance in capturing transient responses. The study highlights the potential of data-driven modeling for control-oriented engine applications and emphasizes the importance of data quality, preprocessing, and network structure in achieving reliable predictive models.