Given the complex characteristics of carbon price series, such as nonlinearity and non-smoothness, the prediction is quite challenging. To improve the forecasting accuracy, TTBiGRUA was proposed as a carbon price prediction model, which combines time-varying filter empirical modal decomposition (TVFEMD), sample entropy (SE), bidirectional gated recurrent unit (BiGRU) and autoregressive integrated moving average (ARIMA) . First, the carbon price was decomposed into modal components with different frequencies by TVFEMD; then, the sample entropy was used to evaluate the complexity of each component and the K-means algorithm was used for reconstruction. Secondary TVFEMD decomposition was then applied to the most volatile modal components after reconstruction to further extract features and reduce modal aliasing. The high-frequency and low-frequency components were separated according to the sample entropy. The high-frequency components were predicted by BiGRU, while the low-frequency components were predicted by ARIMA, and the final prediction results were obtained by superimposing the component prediction results. We applied the actual carbon price data from the carbon markets in Guangdong and Hubei, China, and evaluated the predictive validity and robustness of the model by applying five evaluation indicators and the Diebold-Mariano (DM) test. The results show that the predictive accuracy of the proposed model is better than other benchmark comparison models.