Neural Network Design of Control-Oriented Autoignition Model for Spark Assisted Compression Ignition Engines
- 提供方法
- 版元よりダウンロードリンクを連絡
- 形態
- 価格
- 一般価格(税込)：￥6,600 会員価格(税込)：￥5,280
- 文献・情報種別
- SAE Paper
No.2021-24-0030
- 掲載ページ
- 1-13(Total 13 p)
- 発行年月
- 2021年 9月
- 出版社
- SAE International
- 言語
- 英語
- イベント
- International Conference on Engines and Vehicles 2021
書誌事項
| 著者（英） | 1) Dennis Robertson, 2) Robert Prucka |
|---|---|
| 勤務先（英） | 1) Clemson University, 2) Clemson University |
| 抄録（英） | Substantial fuel economy improvements for light-duty automotive engines demand novel combustion strategies. Low temperature combustion (LTC) demonstrates potential for significant fuel efficiency improvement; however, control complexity is an impediment for real-world transient operation. Spark-assisted compression ignition (SACI) is an LTC strategy that applies a deflagration flame to generate sufficient energy to trigger autoignition in the remaining charge. Operating a practical engine with SACI combustion is a key modeling and control challenge. Current models are not sufficient for control-oriented work such as calibration optimization, transient control strategy development, and real-time control. This work describes the process and results of developing a fast-running control-oriented model for the autoignition phase of SACI combustion. A data-driven model is selected, specifically artificial neural networks (ANNs). The models are trained to an experimentally-validated one-dimensional engine reference model. A hybrid approach is utilized, where a classification ANN determines if autoignition occurs, and a regression ANN computes the crank angle of autoignition. The regression ANN is a two-layer network leveraging Bayesian regularization which produced a root mean square error of 2.5 CAD compared to the reference model. This error is consistent with cycle-to-cycle CA50 variation, sensitivity of autoignition phasing to ringing intensity, and model accuracy from production-intent SACI combustion models. The execution speed of this model derives from the computation of autoignition phasing by bulk cylinder state and engine geometry with no crank-angle resolved data. The model predicts feasible autoignition phasing well outside of the training region and can identify the negative temperature coefficient region if present for the fuel. The results indicate ANNs have suitable accuracy for use as a control-oriented autoignition model. 翻訳 |
