A stepwise decomposition and multi-label feature selection framework for carbon price forecasting amidst real-time data drift

Increasing global greenhouse gas emissions have far-reaching impacts on natural ecosystems and human society. The carbon emission rights market is an effective means to reduce greenhouse gas emissions, and accurate prediction of carbon price is crucial to promote the transition to a low-carbon economy. However, the “data drift” problem caused by real-time decomposition in a multi-factor environment is rarely considered in existing carbon price prediction studies. Specifically, each data point in the real-time decomposition presents different waveform characteristics, and the combination of the newly obtained decomposition results with the previous results causes significant fluctuations in the waveform, thus giving rise to the “data drift” problem. To address this issue, this paper proposes an innovative stepwise decomposition approach that leverages data drift points as training features. By doing so, it circumvents the integration of outdated and recent decomposition outcomes, thereby efficiently tackling the “data drift” challenge arising from real-time decomposition processes. In addition, this paper proposes a multi-label feature selection technique based on this, which effectively combines feature selection and time lags, and solves the time lags effect between carbon price and features in multi-factor prediction. The experimental results on the Guangdong and Hubei datasets show that the proposed new paradigm hybrid prediction model not only accurately captures the trend of carbon price but also effectively measures the uncertainty range of carbon price, which has significant advantages over other comparative models.