Aiming at the problems of large information loss and feature loss in the similarity design of high-dimensional panel data in clustering, a new panel data clustering method was proposed, which named an adaptive clustering method for panel data based on multi-dimensional feature extraction. This method defined "comprehensive quantity", "absolute quantity", "growth rate", "general trend" and "fluctuation quantity" of samples to extract features, and the five features were weighted to calculate the samples comprehensive distance. On this basis, ward method is used for clustering. This method can greatly reduces the loss of effective information. To verify the effectiveness of the method, cluster empirical analysis was conducted using GDP panel data from 31 regions in China, and the clustering results were compared with those of other clustering models. The experimental results showed that the proposed model was more interpretable and the clustering results were better.