Enhanced short-term carbon emission forecasting via hybrid decomposition-denoising modeling.

Carbon emissions have long been a topic of significant concern both domestically and internationally. The phased establishment of emission reduction targets allows for more precise control over the growth trajectory of greenhouse gases. Therefore, enhancing the accuracy and generalization of carbon emission predictions serves as a crucial foundation for China to achieve its "Dual Carbon" goals. Based on the "Decomposition-Integration" methodology, this paper develops a novel hybrid model named "Decomposition-Denoising-Integration" for short-term carbon emission forecasting. The framework integrates three methodologies: Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Empirical Wavelet Transform (EWT), and Long Short-Term Memory (LSTM) neural networks. First, CEEMDAN decomposes carbon emission time series into multiple intrinsic mode sub-sequences and a residual sequence, enabling multi-scale feature extraction from the raw data. Subsequently, EWT is applied to perform adaptive frequency band segmentation and reconstruction on high-frequency noisy modes, effectively suppressing noise interference while preserving valid information. Building upon these processed components, independent LSTM prediction models are constructed for each modal component to capture temporal dependencies at varying scales via gate-controlled mechanisms. Finally, the predictions of each consequence are integrated through linear superposition to generate a composite prediction results that combine detailed features and trend patterns. Experimental evaluations demonstrate that the proposed framework outperforms seven baseline models, as evidenced by improvements across four key evaluation metrics: Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R