A Hybrid Time–Frequency-Domain-Enhanced iTransformer for Temporal Carbon Emission Prediction

Short-term forecasting of carbon dioxide (CO2) emissions supports near-term mitigation planning, but real-world emission series are nonlinear, non-stationary, and contaminated by multi-scale fluctuations. This paper proposes RVWF-iTransformer, which integrates RIME-optimized variational mode decomposition (RIME-VMD), a trainable one-dimensional wavelet convolution block (WTConv1d), and a discrete-cosine-transform-based frequency-enhanced channel attention mechanism (FECAM) into an iTransformer backbone. The model was evaluated on daily national CO2 emissions for China and India using Carbon Monitor and on two public benchmarks (PM2.5 and ETTH2) using a chronological 8/2 (train/test) split (with validation taken from the tail of the training segment for early stopping) and horizons H∈{5,10,15,20}. Predictive results are reported as mean ± standard deviation over 20 independent runs; RVWF-iTransformer yields the lowest errors at longer horizons on China-CO2 and maintains robust performance under additive Gaussian noise with SNR = 20/10/5 dB. These findings suggest that aligning adaptive decomposition and time–frequency representation learning within a single causal pipeline improves forecasting stability for non-stationary environmental time series.