Comparative study of the artificial neural network with three hyper-parameter optimization methods for the precise LP-EGR estimation using in-cylinder pressure in a turbocharged GDI engine. (25th February 2019)

• Record Type: Journal Article
• Authors: Jo, Yuhyeok; Min, Kyunghan; Jung, Donghyuk; Sunwoo, Myoungho; Han, Manbae
• Is Part Of: Applied thermal engineering. Volume 149(2019)
• Page Start: 1324
• Page End: 1334
• Publication Date: 2019-02-25
• Subjects: 00-01 -- 99-00; Turbocharged GDI engine -- LP-EGR -- EGR estimation -- In-cylinder pressure -- Artificial neural network (ANN) -- Hyper-parameter optimization; Heat engineering -- Periodicals; Heating -- Equipment and supplies -- Periodicals; Periodicals; 621.40205
• DOI: 10.1016/j.applthermaleng.2018.12.139
• Languages: English
• ISSNs: 1359-4311

Abstract

Highlights: Our paper presents an artificial neural network (ANN) to estimate LP-EGR for a turbocharged GDI engine. Experimental data at steady-state engine operations was used to train, validate, and test the proposed ANN. Hyper-parameters of the proposed ANN were optimized by three optimization methods. The proposed three optimization algorithms can efficiently find high performance ANNs. Abstract: The LP-EGR technology has drawn to further improve the fuel efficiency and reduce exhaust emissions in turbocharged GDI engines, so that the accurate model to estimate a LP-EGR rate is a prerequisite. For a precise model of the LP-EGR rate, the transport delay for each cycle and the highly nonlinear flow characteristics should be considered. This paper proposes an artificial neural network (ANN) architecture using multiple combustion parameters calculated from the in-cylinder pressure to precisely estimate the LP-EGR rate without time delay and capture the highly nonlinear flow characteristics. The ANN model was trained with Levenberg-Marquardt back-propagation using steady-state data of 9000 cycles and validated using the cross-validation technique. In this work, three optimization algorithms are introduced and compared to search for the optimal values of the ANN's hyper-parameters: random search (RS), tree-structured Parzen estimator (TPE), and hyper-parameter optimization via radial basis function and dynamic coordinate search (HORD). As a result, all three algorithms effectively improved the efficiency to search multidimensional hyper-parameters, with the validation performance R 2 above 0.98 and a total RMSE of less than 0.76%. This means that all three methods were successful to find the hyper-parameters' values for the precise LP-EGR rate estimation. Among the three algorithms the HORD showed the best performance with stable convergence, the R 2 values of 0.9896 or more, and a total RMSE of 0.63%.