Prediction Of Automobile CO₂ Emission Based On SHAP And BS-XGBoost

With the continuous development of the global transportation industry,its energy consumption is also increasing,resulting in a large number of CO₂ emissions.These emissions aggravate the degree of global greenhouse effect,causing the international community on the transportation industry CO₂ emission problem.As the main component of the transportation industry,the CO₂ emissions of the automobile account for a large proportion of the CO₂ emissions of the transportation industry.Therefore,it is very important to study the CO₂ emission in automobile exhaust.However,when building a predictive model,researchers usually pursue the high performance and accuracy of the model,but ignore the interpretability of the model.At the same time,in order to facilitate users to use the constructed automobile CO₂ emission forecast model,an easy-to-operate forecast software is also essential.In view of the above problems,this thesis builds an automobile CO₂ emission prediction model based on machine learning technology,explains and analyzes the prediction results of the model combined with SHAP algorithm,and finally designs the automobile CO₂ emission prediction software based on Streamlit,which helps the relevant departments to formulate more reasonable and scientific policies and measures.The main research arrangements of this thesis are as follows:（1）Data preprocessing and feature analysis before modeling.Before modeling,the experimental environment and experimental data of this thesis are introduced,outlier processing,duplicate value processing and missing value processing are carried out,and the string data in the data is labeled and encoded.Then the features in the data are visually analyzed,and the correlation between the features is analyzed by Spearman correlation coefficient.（2）Automobile CO₂ emission prediction model based on machine learning.Firstly,the preprocessed data are randomly divided,and the model evaluation indexes used in this thesis are introduced.Then,a prediction model of automotive CO₂ emissions is constructed based on six single machine learning algorithms,and ten fold cross-validation and learning curves are used to improve the fairness and reliability of the comparison between different models.The BS-XGBoost model is constructed with the best model XGBoost among the six models combined with Bayesian optimization algorithm,and the idea of Stacking is used to fuse multiple models into a better model.Finally,by comparing six single models,BS-XGBoost model and Stacking model,it is found that the BSXGBoost model has significant advantages in the forecasting accuracy and stability of automotive CO₂ emissions.（3）Interpretability analysis of automobile CO₂ emission prediction model.The BSXGBoost model is used as an example to demonstrate a variety of analysis methods,including single sample SHAP analysis,feature influence thermal map analysis,cluster SHAP analysis and SHAP feature interaction analysis.SHAP algorithm is used to explore the importance of each feature in BS-XGBoost model and how they affect the model output.（4）Design and implementation of automobile CO₂ emission forecasting software based on Streamlit.Using the trained machine learning prediction model,an automobile CO₂ emission prediction software based on Streamlit is designed and deployed on Streamlit Cloud.