On the integration of physics-based and data-driven models for the prediction of gas exchange processes on a modern diesel engine

Abstract

The need for precise control of complex air handling systems on modern engines has driven research into model-based methods. While model-based control can provide improved performance over prior map-based methods, they require the creation of an accurate model. Physics-based models can be precise, but can also be computationally expensive and require extensive calibration. To address this limitation, this work explores the integration of data-driven models into an overall physics-based framework and applies this approach to the gas exchange processes of a diesel engine with a variable geometry turbocharger and exhaust gas recirculation. One of the most complex parts of this gas exchange loop is the turbocharger. Data-driven methods are used to capture the turbocharger performance and are also applied to the intake manifold, while the simpler features are captured with more traditional physics-based models. This combined modeling approach is able to capture the temperature and pressure dynamics with varying error levels depending on measurement availability and the inter-dependency of the submodels, with the turbocharger neural network model achieving a Normalized Mean Square Error (NMSE) of 5e-5 and the overall engine model achieving a NMSE of 4.5e-3. The work illustrates that the integration of data-driven models can improve overall model accuracy and may be able to reduce the number of sensors needed on the system. The contributions of this work are the development and demonstration of a neural network based turbocharger model and intake air path model, the development of empirical equation-based models for the rest of the engine components along the air path and the demonstration of the integration and interaction of these two types of model to adequately characterize engine operation for control applications.

A Hybrid Physics-Based and Stochastic Neural Network Model Structure for Diesel Engine Combustion Events

“…The temperature, pressure, and oxygen fraction at intake valve closing (T IVC , P IVC , F IVC ), SOI timing, and fuel mass were utilized as model inputs to the ANN model, and the CA50 estimation error was able to be reduced from 5 CAD to 1 CAD. In a previous work, the authors utilized an integrated physics-based and data-driven approach to capture the gas exchange processes of a modern diesel engine, and observed similar improvements in performance [1].…”

“…In this model framework, a standard approach was utilized for the ANN weight parameterization, using a Levenberg-Marquardt (LM) training algorithm. The mathematical implementation of the MLP and LM algorithm is described in the authors' previous work [1]. The LM algorithm was utilized due to its known high computational power, low CPU memory usage, and high performance [26].…”

“…Stricter emissions regulations and renewed interest in reducing the carbon footprint of internal combustion engines have prompted the evolution of the fuel and air pathways of modern diesel engines, in order to maximize efficiency while minimizing exhaust emissions [1,2]. Some of the complex subsystems introduced in modern diesel engines include the turbocharger and high-and low-pressure exhaust gas recirculation loops [1]. High-fidelity models or combustion phasing feedback are essential to ensuring proper combustion management on these devices [3,4].…”

“…The literature demonstrates that leveraging physics to inform ANN models could enhance predictions of engine metrics and, as such, improve real-time combustion phasing control methods without the strict need for in-cylinder sensors. While prior work using either direct ANNs or integrated ANN and physics-based approaches to predict critical engine metrics has shown promise [1,[5][6][7]13,15,16,[18][19][20][21][22][23][24], some challenges with uncertainty and complexity still need to be addressed. Some existing methods assume linear relationships and, as such, are simpler to use for control, but have a more limited ability to extrapolate, and encounter greater inaccuracy due to the nonlinearity of engine processes [7,13,18,22].…”

“…These additions provide modeling simplicity that has not been achieved in the literature, and these methods may enable reduced modeling effort and lower computational power to be required for combustion control. Moreover, as previously mentioned, Wang utilized an ANN to capture SOC for CA50 prediction in an SI engine [5], but the CA50 model needed independent modelling of the in-cylinder pressure as a model input, which can be challenging to capture accurately in the presence of uncertainties [1]. The combustion phasing model developed in this study avoids the need for an in-cylinder pressure model.…”