A method to identify wood based on random forests prediction model of terahertz time-domain spectroscopy is proposed. We analyzed the differences of the absorption spectrum of four kinds of the woods (two kinds of real red wood and two kinds of false red wood) in the frequency range of 0.2～1.2 THz. Then the principal component analysis was applied to decrease the dimension of the spectral absorbance data, and the five principal components with top cumulative variance contribution rates were extracted (the total contribution rate is up to 99.65%). The processed spectral data were substituted into the random forest prediction model to identify real red wood and false red wood, and then the recognition rate of the training set and test set were obtained. The experimental results show that the terahertz time-domain spectroscopy combined with random forest prediction model can obtain a higher recognition rate, the recognition rate can reach 91.25%, when comparing with that using the traditional support vector machine prediction model and single decision tree model. The research results show that it is feasible to apply the terahertz time-domain spectroscopy combined with random forest prediction model into the identification of red wood.