A drift-aware dynamic ensemble model with two-stage member selection for carbon price forecasting

• Published in: Energy (Oxford) Vol. 313; p. 133699
• Main Authors: Zeng, Liling, Hu, Huanling, Song, Qingkui, Zhang, Boting, Lin, Ruibin, Zhang, Dabin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 30.12.2024
• Subjects: Carbon emission; Drift detection; Dynamic ensemble; Model average; Model selection; Carbon emission; Dynamic ensemble; Model selection; Drift detection; Model average
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2024.133699

Forecasting carbon prices is a pivotal topic in achieving the targets of carbon neutrality and carbon peaking. However, the complex and time-evolving characteristics inherent in carbon price series render precise forecasting a formidable undertaking. Numerous studies have demonstrated that distinct prediction models exhibit varying capabilities and performances, and ensemble learning offers an efficacious approach to enhance forecasting performance. To address variations in model performance and data distribution, a drift-aware ensemble learning framework is employed to adaptively select and combine models for carbon prices forecasting. First, thirty candidate models are generated by integrating data processing techniques with multiple forecast models to comprehensively capture sample information. Second, an initial selection process of candidate models is dynamically executed utilizing a performance drift detection mechanism. Following each drift detection, a second-stage selection is performed given the significance of diversity in ensemble models. Finally, final predictions are calculated by combining the outputs of selected models via a sliding-window weighted average. Carbon price data from four distinct trading markets in China are employed to validate the efficacy of the drift-aware dynamic ensemble (DDE) framework. The results substantiate that DDE can be a convincing tool for the operation and management of carbon trading markets. •CEA prices predictions demonstrate the efficacy of DDE.•DDE ensemble learning can enhance the forecast performance.•The DDE executes in an informed manner.•Model pruning can avoid the risk of choosing the wrong model.