An ensemble self-learning framework combined with dynamic model selection and divide-conquer strategies for carbon emissions trading price forecasting

Abstract

A reliable carbon price forecast system is essential for governments to assess “net-zero emission” targets, guiding production, operation, and investment through risk prevention and control measures. Although existing studies report numerous hybrid or ensemble models for carbon price forecasting, there is still considerable room for optimization due to the lack of targeted judgments on series features. This paper proposes a dynamic multi-objective self-learning combination framework based on the model-algorithm space, which adaptively selects the ensemble scheme with the best performance according to the specific laws of the carbon price series features while ensuring the diversity of base models. Furthermore, the developed divide-conquer strategy, which can better quantify signal irregularities, is employed to overcome obstacles caused by the high complexity of some components during data preprocessing. Carbon price series from the European and Shenzhen carbon markets validate the hybrid method's ability to handle different signals. Experimental studies reveal that the proposed carbon price prediction model possesses a reasonable structure and strong interpretability, yielding accurate, robust, and generalized prediction results.