With increasing urbanization and expanded transportation, traffic accidents have become more frequent, causing significant societal harm. While the rise of IoT and mobile communication technologies provides access to diverse traffic data, the complexity of this data requires advanced processing methods. Current deep learning-based traffic accident prediction models often lack sufficient spatial-temporal correlation extraction, struggle with data sparsity, and fail to model interconnected urban traffic tasks effectively. This paper addresses these challenges by: 1) proposing a multi-channel spatial-temporal graph neural network model that integrates time attention mechanisms, GCN, and GRU modules to extract spatial-temporal features, demonstrating superior performance across multiple datasets compared to baseline models; and 2) addressing the challenge of sparse traffic data in target cities through an adversarial domain adaptation approach, leveraging an enhanced multi-channel spatial-temporal graph neural network as a feature extractor and optimizing it with a domain classifier to reduce distribution differences between source and target domains, thereby assisting in target domain prediction.