Predictability of artificial neural network (ANN) in performance prediction of a retrofitted CNG engine

• Main Authors: Jahirul, M.I., Saidur, Rahman, Masjuki, Haji Hassan
• Format: Article
• Published: University of Malaya 2010
• Subjects: TA Engineering (General). Civil engineering (General); TJ Mechanical engineering and machinery
• Online Access: http://eprints.um.edu.my/6762/; http://ejum.fsktm.um.edu.my/article/1003.pdf

Compressed natural gas (CNG) is a potential alternative of liquid petroleum fuel in automotive application. The combustion process of CNG in engine is a complex thermodynamic process and highly sensitive with operating conditions. Additionally, the experimental investigations of engine performance are time consuming and quite expensive. Present study utilized artificial neural networks (ANN) modeling technique to evaluate the performance of a retrofitted automotive CNG engine. Back propagation (BP) neural network with single hidden-layer and logistic sigmoid transfer function was used to optimize prediction model performance. The neural networks toolbox of MatLab 7 was used to train and test the prediction models. Engine speed (rpm), throttle position () and operation time (min) were used as the input layers, while engine thermal efficiency (η, ), brake power (bp, kW), break specific fuel consumption (bsfc, kg/kWh) and exhaust temperature (Tex, °C) were used in output layers. For each performance parameter two prediction models, trained with 12 and 24 set of experimental data, were developed in order to investigate the prediction ability of ANN in different number of training samples. After successful model development, CNG performance parameters were simulated with new set of input parameter. Simulation results then compared with experimental results and prediction performance of ANN were evaluated statistically. The results of this study show that ANN is an appropriate modeling technique to estimate performance of the engine used in the experiments. Moreover the prediction ability of ANN models was significantly improved with increasing number of training sample.