With the rapid increase in CO₂ emissions, there is a profound impact on global climate change, seriously hindering the sustainable development of the low-carbon economy. Therefore, it is particularly significant to predict CO₂ emissions. To improve the accuracy and robustness, a new hybrid model (KPCA-CEEMDAN-AWOA-RELM) is proposed when considering multiple factors about historical CO₂ emissions, energy consumption, economic and social. First, factors are determined by Pearson doubly significant test, and the principal component is extracted by kernel principal component analysis (KPCA) to realize nonlinear dimension reduction. Then, the CO₂ emissions sequence is decomposed by the ensemble empirical model with adaptive noise (CEEMDAN) model to reduce noise interference and abate reconstruction error. Finally, CO₂ emissions can be predicted via the extreme learning machine with regularization parameter modification optimized by an improved input weight matrix and the deviation matrix of the adaptive optimization whale algorithm (AWOA-RELM). Taking Hebei and China as examples, it is found that the selected model is better than other comparative models.