Identifying potential drug-disease associations through computational approach like drug repositioning plays a vital role in accelerating drug discovery. However, most existing methods has not adequately addressed the cold-start problem frequently encountered in drug repositioning, where the confirmed drug-disease association data are highly sparse. To bridge this gap, we design a novel model based on residual graph normalized convolutional network to predict drug-disease associations, called RGNCNDDA. To alleviate the cold-start problem in drug-disease association data, RGNCNDDA introduces L2-normalization to prevent the embeddings of isolated nodes from collapsing to zero, and employs residual connection to mitigate embedding information loss during propagation. Then RGNCNDDA adopts collaborative training to integrate the extracted features of drug and disease, and infers the unobserved drug-disease associations based on the integrated features. The comparison results on three benchmark datasets show that RGNCNDDA is superior to the exiting methods. The experiments for isolated nodes further validate the predictive ability of RGNCNDDA. Moreover, the case studies demonstrate that RGNCNDDA is competent to accurately identifying potential drug-disease associations.