- 15 views
- 11 downloads
Realtime Machine Learning based In-Cycle Control of Homogeneous Charge Compression Ignition
-
- Author / Creator
- Gordon, David C
-
Homogeneous Charge Compression Ignition (HCCI) has the potential to significantly reduce engine out oxides of nitrogen NOx emissions, while maintaining a high fuel efficiency compared to existing lean-burn spark ignition engines. HCCI is characterized by compression induced autoignition of a lean homogeneous air-fuel mixture. The challenge with HCCI combustion is the high cyclic variation due to the lack of direct ignition control leading to cycles with high emissions. Advances in control strategies and microcontrollers mean that it is now possible to experimentally test these control strategies for HCCI combustion. In this thesis two advanced control strategies will be developed and experimentally implemented on a single cylinder research engine.
To control HCCI combustion it is necessary to combine fast actuators, in-cycle control and model-based cycle-to-cycle control. Using a high-speed ignition system which can add energy to the cylinder in under 10µs, an in-cycle controller was developed and experimentally tested. The proposed controller was able to successfully reduce the standard deviation of combustion phasing and indicated mean effective pressure (IMEP) by up to 34% and 28%, respectively. Based on the success of the in-cycle controller, it was coupled with a cycle-to-cycle controller for the experimental implementation of a multiscale controller. When compared to the single actuation of the in-cycle control strategies, the advantages of MIMO controller showed a significant improvement in the prediction and prevention of misfire cycles. This control approach provided a clear increase in indicated efficiency from 28.09% to 29.42%, improved IMEP from 2.88bar up to 3.02bar and helped to stabilize the operation point and a reduction of standard deviation for IMEP and CA50 by more than 65% was achieved.
Three models have been developed for determining the emissions and performance of HCCI. These models cover a wide range of modeling strategies from physics based kinetics modeling to fully machine learning based black-box models. The first model investigated was a physical kinetics model which is able to provide a breakdown of the chemical species in the cylinder. An offline kinetics model was successfully created and showed that the chemical components of the cylinder could be estimated using a relatively simple kinetics model with 34 species and 36 reactions. To improve the real-time capability of the model, a machine learning (ML) model was then investigated. The first ML model developed is a support vector machine (SVM) model with the goal of determining the effect of different ML approaches and feature set selection on the model quality for HCCI emissions prediction. A linear and a non-linear SVM model were compared to a traditional artificial neural network (ANN) model. This comparison showed for a small data set that SVM based models were more robust to changes in feature selection and better able to avoid local minimums compared the ANN leading to a more consistent model prediction. Finally a transient engine performance and emissions model for HCCI was created using a deep neural network (DNN) containing a long short term memory (LSTM) layer. This model requires significantly more data than the SVM model. However, due to the recurrent neural network the model is able to capture time dependencies in the data. This resulted in an accurate model for transient engine operation with an error less than 5% for all four model outputs. Of the models developed and tested the LSTM model showed the greatest accuracy while preserving a simple model structure to allow for real-time implementation.
Using the developed LSTM based DNN model a nonlinear model predictive controller (NMPC) has been designed and experimentally tested. To successfully implement the NMPC, the open-source package acados which enables the integration of embedded solvers for nonlinear optimal control has been used. The NMPC online optimization was tested on various processors to determine which provided the required realtime turnaround time. Of the six devices tested, all showed the possibility to meet the real-time requirements. Therefore to keep implementation costs low a Raspberry Pi 400 was chosen to test on the engine testbench. The implementation of the acados NMPC on the Raspberry Pi has been experimentally shown to follow an IMEP reference with an RSME of 0.133bar on the HCCI engine. The NMPC was also able to observe constraints and keep the combustion phasing close to the target value.
-
- Graduation date
- Spring 2023
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.