Piwi-interacting RNA(piRNA) is widely recognized as closely associated to human complex diseases. Therefore, identifying piRNA-disease associations (PDAs) plays an important role for understanding the underlying genetic mechanisms of complex diseases. Many computational methods have been proposed for identifying PDAs. However, they primarily use traditional graph neural networks for feature extraction. In this paper, a method PDA-GTGCN that uses a group feature transformation graph convolutional network (GCN) to predict PDAs. Initially, a heterogeneous network is firstly constructed based on the similarity and association information of piRNAs and diseases. This heterogeneous network is then fed into a GCN with a layer-wise attention mechanism to extract feature information. Secondly, a group feature transformation module is developed for aligning feature dimensions, fully considering the meaning of each feature dimension for preventing overfitting issues. Finally, the score of each PDA is obtained through cosine similarity calculation and a feature fusion attention mechanism. The AUC of five-fold cross-validation achieves 0.9656 and the ACC achieves 0.9572. Case studies on Head and Neck Squamous Cell Carcinoma, and Renal Cell Carcinoma, further confirm the superior performance of PDA-GTGCN. Therefore, PDA-GTGCN is an effective method for predicting hidden PDAs.