Input and Structure Choices of Neural Networks on the Fuel Flow Rate Prediction in the Transient Operation Condition

• Published in: SAE International journal of engines Vol. 5; no. 4; pp. 1535 - 1550
• Main Authors: Deng, Jiamei, Stobart, Richard, Ordys, Andrzej
• Format: Journal Article
• Language: English
• Published: Warrendale SAE International 01-12-2012; SAE International, a Pennsylvania Not-for Profit
• Subjects: Consumption; Correlations; Diesel engines; discontinuous measurement; Energy efficiency; Engines; Flow velocity; Fuel consumption; Fuel economy; Fuel flow; fuel meters; Fuels; Gravimetry; input choice; Internal combustion engines; Linear regression; Modeling; Neural networks; Neurons; Principal components analysis; Real time; real time application; Steady state; Time measurement; Torque; Transient operation
• ISSN: 1946-3936; 1946-3944; 1946-3944
• DOI: 10.4271/2011-01-2458

Measurement accuracy and repeatability for fuel rate is the key to successfully improve fuel economy of diesel engines as fuel economy could only be achieve by precisely controlling air/fuel ratio and monitor real-time fuel consumption. The volumetric and gravimetric measurement principles are well-known methods to measure the fuel consumption of internal combustion engines. However, the fuel flow rate measured by these methods is not suitable for either real-time control or real-time measurement purposes. The problem concerning discontinuous data of fuel flow rate measured by using an AVL 733s fuel meter was solved for the steady state scenario by using neural networks. It is easier to choose inputs of the neural networks for the steady state scenario because the inputs could be chosen as the particular inputs which excited the system in the application. But for transient scenario, such as the NRTC cycle, it will be difficult to choose inputs as there are no excited inputs that could be chosen. This paper attempts to solve the problem of the input choice for the transient state.
This paper illustrates input choices to catch fast and slow dynamics using both phenomenon models and principal component analysis (PCA) and demonstrates a comprehensive and detailed black-box modelling technique for engine applications in the transient condition. A non road transient cycle (NRTC) is used for system identification. Different neural network structures are compared among feed-forward NN, feed-forward NN with delays and non-linear autoregressive model with exogenous inputs (NLARX). The NLARX is the best structure for fuel flow rate prediction in the transient operation of the engine. It is shown that the input choice is reasonable and could be proved by both theory and simulation results. This paper shows that using chosen the inputs, NLARX could predict fuel flow rate accurately with Rsquare above 0.99.