Comparison and Analysis of Prediction Models for Locomotive Traction Energy Consumption Based on the Machine Learning

Abstract

Locomotive traction energy consumption is a multivariate coupled nonlinear system closely related to many factors such as locomotive properties, routing, line conditions, and operating methods. In order to accurately calculate and predict locomotive traction energy consumption, three prediction models are constructed based on a large number of measured operation data of the HX $_{\mathrm {D}}1$ locomotive. This research uses two neural network methods of backpropagation and radial basis functions, as well as support vector machines. Among them, the training group uses 600 data; the validation group uses 200 data. The above methods can be compared and analyzed with prediction performance between different neural networks and machine learning algorithm models. The results show that the RBF and BP neural networks can effectively predict locomotive traction energy consumption among the different neural network models. The $\text{R}^{2}$ of the two neural network model test groups is 0.9926 and 0.9885, the MAPE is 2.91% and 7.28%, and the MSE is 0.02% and 0.08%, respectively. Moreover, we have avoided the influence of the randomness of the neural network algorithm through repeated running. It shows that RBF neural network is better than BP neural network in predicting locomotive traction energy consumption, with more powerful approximation performance and higher accuracy. Among the different machine learning algorithms, the $\text{R}^{2}$ of the SVM algorithm model test group is 0.9983, the MAPE is 2.01%, and the MSE is 0.02%, which shows the prediction accuracy and overall performance of the SVM algorithm model are better than the neural network model. Finally, we prove the broad generalization of the SVM algorithm through the application on other lines. The SVM algorithm model can be a powerful tool for calculating and predicting the traction energy consumption of HX $_{\mathrm {D}}1$ locomotives.

Keywords

• Machine learning
• neural network models
• support vector machine models
• traction energy prediction