Forecasting carbon trading price effectively is of great significance for the formulation of carbon financial market policies that suit to Chinese condition and the risk management of carbon financial market. This paper proposes a multi-scale combined forecast method for carbon price based on manifold learning of unstructured data. Firstly, the unstructured data related to carbon price is extracted using the network search index, and dimensionality reduction is performed based on the isometric mapping manifold learning. Then, the structured data of other influencing factors and the carbon trading price are decomposed into a variable number of intrinsic mode functions(IMFs) using empirical mode decomposition(EMD) respectively. The IMF is reconstructed to get high frequency sequence, low frequency sequence and trend item based on the fine-to-coarse method. Moreover, autoregressive integrated moving average model(ARIMA), Partial least squares(PLS) regression and neural network are used to forecast high-frequency data, low-frequency data and trend items, which are aggregated to get the final forecast result. The results of simulated experiments show that the proposed method can effectively use multi-source information and has high prediction accuracy and good applicability.