Heterogeneous GPU clusters play an important role in processing parallel applications and massive data sets in the cloud platform. However, due to the diversity of GPU types, how to effectively allocate various GPU types is a challenge. This paper first analyzes the characteristics of request and allocation for various GPU types based on Alibaba cluster data. Then we propose a method to adaptively select the best model to predict demand of various GPU types, and feature extraction from the best model. Further, we design a model based on Long Short Term Memory (LSTM) to forecast allocation of each GPU type. Finally, the extensive experimental results demonstrate the promising performance compared with multiple baseline methods by using real-world trace data from Alibaba cloud data centers. The trial illustrates that the demand prediction accuracy of the adaptive selection method reaches 87%, while the proposed prediction allocation model yields better performances with root mean square error and mean absolute error of 1.78 and 0.85, respectively.