After signing the Paris Agreement and piloting carbon trading for many years, China has taken a significant step toward carbon neutrality. Carbon price forecasting is helpful to construct an effective and stable carbon pricing mechanism and provide practical guidance for production, operation, and investment. This paper builds multiple one-step-ahead predictors to analyze and forecast carbon prices, based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Long Short-Term Memory (LSTM) recurrent neural network, with the close prices of Guangzhou Emission Trading Scheme (Guangzhou ETS) from March 14, 2014, to August 31, 2021. After an initial period of unreasonable pricing and market adapting, the carbon price of the Guangzhou ETS began to recover gradually. According to previous literature, this paper summarizes two fundamental CEEMDAN–LSTM frameworks and proposes a hybrid one combined with Variational Modal Decomposition (VMD). With the help of adaptive reducing learning rate Adam optimizer and the early stop mechanism, the forecast turns out stable and reliable results, with a best average coefficient of determination (R2) of 0.982 and Mean Absolute Percentage Error (MAPE) of 0.555%, which shows that Sample Entropy integration and re-decomposition methods are conducive to carbon price forecasting. Validations of four ETS and different timesteps also verify the effectiveness of the hybrid VMD LSTM method, but it still needs to be optimized for practice. • Summarize two fundamental frameworks based on CEEMDAN and LTSM. • Test and improve by the average performance of multiple runs. • Compare re-decomposition methods and prove the effectiveness of VMD. • Propose a reliable and stable hybrid framework for carbon price forecasting. • Reproduce easily due to the upload files in supplementary and Github.