Research On CNN-LSTM Fuel Consumption Prediction Method Integrating Driving Behavior Characteristics

The rapid increase in car ownership has brought convenience to people’s lives,but it also has caused problems such as energy crisis and environmental pollution.The study based on the impact of driving behavior on energy consumption is an important foundation for achieving energy conservation and emission reduction.Existing research has not considered Jerk(the first derivative of acceleration),which is a factor that better characterizes driving behavior characteristics,resulting in weak correlation between fuel consumption models and driving behavior,leading to low model prediction accuracy and other issues.This study is based on domestic and international research,analyzing the impact of driving behavior on fuel consumption,and proposing a car fuel consumption prediction model that integrates driving behavior features.The main tasks and achievements completed are as follows:(1)The impact of driving behavior on fuel consumption was analyzed based on hybrid clustering method.Cluster analysis was conducted on the driving behavior characteristics of“speed+acceleration”and“speed+acceleration+Jerk”based on K-means.And optimize the clustering results by combining Density-Based Spatial Clustering of Applications with Noise(Density-Based Spatial Clustering of Applications with Noise,DBSCAN).The experiment shows that compared to the clustering algorithm without considering Jerk,the accuracy of the clustering algorithm with Jerk is improved by 5.03%;This clustering algorithm can classify the fuel consumption levels of driving behavior into three categories:low fuel consumption,medium fuel consumption,and high fuel consumption.(2)Based on CNN-LSTM fuel consumption prediction model(Fuel Net-J)that integrates driving behavior features has been proposed.The best hyperparameter is selected through comparative experiments,and then the best fuel consumption prediction model is obtained.The experiment shows that the optimal training set size,training batch size,sampling rate,number of iterations,and number of hidden nodes are 20000,200,2,50,and 100 respectively;When"speed≤20 km/h and input parameter is speed"or when"speed>20 km/h and input parameter is speed,acceleration,Jerk",the Fuel Net-J model has the best predictive performance;When speed>20 km/h,the predictive performance of the model is significantly improved by introducing Jerk into the input parameters.Among them,at a vehicle speed of 80km/h,the RMSE(Root Mean Squard Error,RMSE)decreases by 35.29%,At a speed of 40km/h,R~2(Coefficient of Determination,R~2)increased by 10.55%,At a speed of 50 km/h,RE(Relative Error,RE)decreased by 31.71%.(3)A multi-dimensional applicability testing study was conducted on the Fuel Net-J model based on actual vehicle data.Comparative experiments were conducted based on different speeds,vehicle models,models,and application scenarios.The experimental results show that the model has good predictive performance at each speed and vehicle type;Under different speed conditions,Fuel Net-J has the best predictive performance compared to VSP,VT Micro,GRNN,RNN,and GRU.Where,RMSE decreased by 13.4%-20.5%,R~2increased by 31.0%-44.2%,and RE decreased by 27.0%-52.1%;The model can be used for predicting good performance in campus,city,highway,and road intersection scenarios,as well as for diagnosing fuel consumption anomalies:when the error between the predicted and actual fuel consumption values begins to increase sharply,it indicates abnormal fuel consumption and warns the vehicle of potential oil leakage faults.