Affiliations: School of Science, Nanchang Institute of Technology, Nanchang, Jiangxi, China; Key Laboratory of Engineering Mathematics and Advanced Computing, Nanchang Institute of Technology, Nanchang, Jiangxi, China
Correspondence:
[*]
Corresponding authors. Yuerong Li, E-mail address: [email protected] and Jinxing Che, E-mail address: [email protected].
Note: [1] These authors contributed equally and should be considered co-corresponding authors.
Abstract: Accurate prediction of carbon price is of great value for production, operation, investment decisions and the establishment of carbon pricing mechanism. However, the large amount of data often limits the application of learning model with good predictive performance in carbon price prediction. Therefore, the development of learning algorithms with low computational complexity has become a research hotspot. Among them, subsampling integration technology is an effective method to reduce the computational complexity. However, lack of data representativeness in subsamples and ignorance of differences among submodels inhibit the prediction performance of the subsampled ensemble model. This project proposes an optimal weight random forest ensemble model with cluster-based subsampling (FCM-OWSRFE) for carbon price forecasting. Firstly, Fuzzy C-means cluster-based subsampling to ensure the data representativeness of subsamples. Secondly, a series of sub-random forest models are built based on subsamples with data representativeness. Finally, an optimal weight ensemble model from these sub-models is derived. To verify the validity of the model, we test FCM-OWSRFE model with the carbon price of Guangzhou Emission Exchange and the carbon price of Hubei Carbon Emission Exchange, respectively. Experimental results show that Fuzzy C-means cluster-based subsampling and the optimal weight scheme can efficiently improve the prediction performance of the subsampled random forest ensemble model.
Keywords: Random forest, Fuzzy C-means clustering, subsampling, optimal weight ensemble
