Research On Fuel Consumption Prediction Of Multi-Type Commercial Vehicles Based On Module Graph Convolution And Configuration Transfer

With the improvement of the level of social development,the demand for the transportation industry has also further increased,which promotes the wider application range of commercial vehicles and the increasing number of vehicles,but also leads to a large amount of fuel consumption.Commercial vehicle fuel consumption prediction by analyzing driving data,can provide route planning guidance for drivers and improve fuel utilization,which is of great significance to saving energy.With the variety of commercial vehicles becoming more and more abundant,a fuel consumption prediction method suitable for various types of commercial vehicles is needed to solve the problem of fuel consumption prediction.Commercial vehicle sensors monitor the status of vehicle components at a fixed frequency and transmit the data to a data center for analysis.With the support of abundant data,datadriven fuel consumption prediction methods have become the mainstream of research.These methods use the multivariate time series prediction method in machine learning and deep learning to model the time dependence of commercial vehicle time series and the spatial dependence of fuel consumption and other variables to predict the fuel consumption of commercial vehicles in the future.At the same time,there are some methods based on physical models,based on the domain knowledge of commercial vehicle engines,cooling systems and other modules,to establish vehicle mechanism models to predict fuel consumption.However,the current research has the following problems.Firstly,data-driven methods do not combine the domain knowledge of commercial vehicles,and cannot be adjusted in time according to the configuration and structural changes of commercial vehicles.As a result,it is difficult for these methods to obtain accurate fuel consumption prediction results when faced with multiple types of commercial vehicle operating data.Secondly,the existing vehicle physical models are modeled for fixed modules and configurations.However,due to the large number of commercial vehicle types and the number of modules,building a physical model for each commercial vehicle will greatly increase the cost of modeling.Thirdly,the number of some personalized configuration vehicles is small,the vehicle driving data is not enough to support the training of the model,and the existing transfer learning method cannot use the complex and changeable driving data of commercial vehicles for the research of commercial vehicle fuel consumption prediction.Due to the above problems,this thesis focuses on the fuel consumption prediction of various types of commercial vehicles,aiming to establish an accurate fuel consumption prediction model under the guidance of domain knowledge.The research faces the following challenges.Firstly,there are many types of commercial vehicles and modules.How to uniformly express the domain knowledge of vehicle modules to apply to different vehicle types and module configurations?Secondly,after the unified expression of domain knowledge,how to guide the data-driven deep learning modeling process and perform accurate fuel consumption prediction?Thirdly,how to use the vehicle configuration to transfer fuel consumption knowledge between different vehicle types,so as to realize the fuel consumption prediction of vehicles with a small number of personalized configurations?To address the above challenges,this thesis proposes the Multi-type Commercial Vehicle Fuel Consumption Prediction Method Based on Module Graph Convolution and Configuration Transfer(MGCN-CT).This method first uses the domain knowledge vector formula to extract the domain knowledge of each module of the commercial vehicle,constructs a module graph,and expresses the domain knowledge and driving data of each module uniformly in the form of a graph.Then,the module graph convolutional network is used to model the spatio-temporal dependence of the module graph embedded with domain knowledge,and the domain knowledge is applied to the fuel consumption prediction task.Finally,the vehicle configuration classifier is used to learn the configuration difference weights between types,and is used to construct transfer constraints,realize the fuel consumption prediction of the few-sample personalized configuration vehicles,and provide interpretability analysis.The main work and contributions of this thesis are summarized in the following three points:1.This thesis proposes the graph convolutional neural network based on the module graph(MGCN),which solves the problem of expression and utilization of data combined with domain knowledge.Firstly,a domain knowledge vector formula is proposed to extract the domain knowledge of each module of the commercial vehicle,and a module graph is constructed to express the domain knowledge and data in a unified manner.Among them,multiple sub-graphs represent the domain knowledge and vehicle running status of each module,and the host-graph represents the overall domain knowledge and running status of the vehicle.Then,based on the module graph,the module graph convolutional network is designed to model the spatial dependence of sub-graphs and host-graph embedded with domain knowledge,and the fusion gate is used to extract the fusion features containing commercial vehicle module features and overall features.Finally,time-dependent modeling is performed using fusion features,and accurate fuel consumption prediction results for general-purpose vehicles are obtained under the guidance of domain knowledge.2.This thesis proposes a configuration transfer module based on a commercial vehicle configuration classifier(CT),which solves the problem of insufficient samples of personalized configuration types.The model first uses embedding layer and attention mechanism to extract vehicle configuration information.Then it uses the multi-layer perceptron to build a classifier to realize the configuration classification of general vehicle types and personalized configuration types,and obtain the configuration difference weights generated for the classification task.The configuration difference weight is used as the model constraint of transfer learning,the model obtained from the general vehicle types data training is used as the source domain,and the model obtained from the personalized vehicle data training is used as the target domain.Finally,the fuel consumption results generated by multiple models are aggregated by using the configuration difference weight as a constraint,and the fuel consumption prediction results of personalized configuration vehicles that can be interpreted are obtained.3.In this thesis,extensive experiments are conducted on two real datasets to verify the effectiveness of the MGCN-CT model.First,by comparing with the existing baseline model,the MGCN-CT model can effectively improve the accuracy of fuel consumption prediction of personalized configuration models,and design sufficient ablation experiments to verify the effectiveness of each module in the model for fuel consumption prediction.What’s more,we design experiments to analyze the impact of different parameter settings on the model.Then,it is verified by experiments that MGCN can achieve better results than the existing baseline model when predicting fuel consumption of general-purpose vehicles.Finally,the applicability analysis and visualization analysis of the CT module are carried out.