Power generation enterprises are one of the important sources of carbon emission, and the refinement of carbon emission forecasts on the power generation side is of positive significance to the formulation of China’s carbon emission policy. In this context, a carbon emission forecast model based on variational modal decomposition（VMD）and temporal convolutional network（TCN）is proposed for the characteristics of irregularity, nonlinearity and temporal sequence of carbon emission on the power generation side. First, VMD is used to smooth the preprocessing of the carbon emission time series data, splitting the raw carbon emission data into several modal components to reduce irregularities and nonlinearities in the data series. Second, considering the performance degradation of existing machine learning algorithms during the network training process, each modal component is predicted separately based on TCN to maximize efficiency in the use of carbon emission time seriesdata. Finally, the forecast results are reconstructed to obtain the final forecast values of carbon emissions. The results show that compared with the traditional four forecast models, the method effectively improves the effectiveness and accuracy of the forecast model by innovatively combining VMD model and TCN.