Fuel consumption and exhaust emissions of vehicles seriously affect the sustainable development of society and human health. Precise prediction of fuel consumption is very important for developing an intelligent driving strategy that can effectively reduce the mentioned critical issues. This paper proposes a precise vehicle fuel consumption prediction algorithm based on a LSTM (Long Short-Term Memory) deep network. Firstly, we take the vehicle kinetic parameters during a time series and the corresponding accumulated fuel consumption as the output of the LSTM network. To improve the prediction accuracy and efficiency, we optimized the LSTM network parameters with a control variable method. Secondly, in order to train the LSTM model, we collect a batch of original data from a BYD F3 passenger car under different scenarios, including the GPS data, vehicle speed and acceleration data, and the accumulated vehicle fuel consumption. In addition, we analyzes the correlation between the different combinations of the input parameters and the prediction accuracies via the comparison experiments, and find that the LSTM network achieves the best predictive performance when the input is the combination merely comprising of the speed and acceleration parameters. Finally, we compare the proposed model with three existing models: vehicle specific power (VSP) model, Virginia Tech microscopic (VT-Micro) model, and Generalized Regression Neural Network (GRNN) model. The experimental results show that the root mean squared errors (RMSE) of these four models are respectively 0.539, 1.351, 1.273 and 0.832; and the relative errors (RE) are separately 0.083, 0.573, 0.623 and 0.132. Obviously, the proposed LSTM network greatly outperforms the other 3 methods, which can also be used to evaluate the performance of Eco-driving strategies.

CICTP