Comparison of Carbon Emission Forecasting in Guangdong Province Based on Multiple Machine Learning Models

Abstract

Guangdong province was selected as the experimental region to construct and compare the prediction ability of carbon emissions with six machine learning models: K-Nearest Neighbor, Back Propagation neural network, Random Forest, multiple linear regression model, XGBoost, and LightGBM. mRMR algorithm was used to select the optimal features as the input for each model to conduct carbon emission prediction experiments. The prediction ability of each model was investigated by calculating and analyzing the prediction accuracy, model running time, and model memory consumption. The results show that Random Forest, XGBoost, and LightGBM models have higher prediction accuracy than other models, and the features of population size and GDP per capita have the highest importance. LightGBM has the advantages of the short model running time and small memory consumption with the best overall performance while having higher model accuracy.