Developing neural networks-based prediction model of real-time fuel consumption rate for motorcycles: A case study in Vietnam

ABSTRACT

This paper presents a study on simulation of the on-road fuel consumption for motorcycles using the Neural Networks. Ten representative streets in the inner city of Hanoi were selected to collect the on-road operation characteristics of the test motorcycle, consisting of instantaneous speed and fuel consumption rate. The collected data, including 14,000 data points, was divided into 70% for the training process and 30% for the validating and testing process. An ANN architecture consisting of one layer of suitable input variables, two hidden layers with 20 neurons, and one layer of output fuel consumption rate was proposed. The Levenberg-Marquardt ANN fitting tool and the sigmoid activation function were used in developing model architecture. Three powerful input variables were identified, consisting of instantaneous speed, acceleration of motorcycle, and engine revolution speed. In the model development, the mean absolute percentage errors were 5.10% and 11.10%, and the correlation coefficient R values were greater than 0.8 for the training and testing dataset. The developed model performance was post-evaluated using on-road and laboratory datasets. More than 75.40% of the predicted instantaneous fuel consumption values had a relative deviation less than 20% compared to the on-road measured values. The difference in average fuel consumption (L/100 km) between predicted and measured values was less than 10%. Following the datasets from the laboratory test on the motorcycle chassis dynamometer AVL CD20”, including the World motorcycle test cycle and Hanoi motorcycle driving cycle, the fuel consumption rate determined using the developed model correlated well with the measured ones, with the R² values were 0.67 and 0.72.

KEYWORDS:

• Instantaneous speed
• fuel consumption
• neural networks
• motorcycles
• real-world driving

Abbreviations

Table

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This research is funded by University of Transport and Communications (UTC) under grant number T2022-MT-013TĐ.