The emission of carbon dioxide has received much attention in recent years, as it can reflect the effectiveness of those low-carbon measures. To alleviate the nonlinearity and volatility of the annual carbon dioxide emissions, which may affect the forecast accuracy, this paper proposes an efficient decomposition-integration forecasting method to forecast the annual carbon emissions. The empirical mode decomposition (EMD) is used to decompose the original emission series into intrinsic oscillatory modes and a residual with different frequencies. The least squares support vector machine (LSSVM) is optimized using the particle swarm optimization (PSO) algorithm to predict each decomposed part. This paper chooses the real annual carbon emission of 12 countries all over the world to do the case study. The forecasting results indicate the validity of the EMD on improving the accuracy of carbon emission prediction. Furthermore, the comparison results between the EMD-PSO-LSSVM method and other forecasting models show the EMD-PSO-LSSVM can improve the average accuracy of the mean absolute error (MAE) at least 46.46% and at most 90.09%, and can improve the average Pearson correlation coefficient (PCC) at least 10.45% and at most 45.10%.