Digital payments, particularly in the form of cryptocurrencies, have become an essential part of modern financial systems. These transactions create complex and extensive networks of transaction histories. However, the rise of fraudulent transactions poses a significant threat and can lead to substantial financial losses. Current fraud detection methods largely rely on static graph representations, which are inadequate for capturing the dynamic nature of fraudulent behaviors or for dealing with missing information in transaction networks. This paper introduces the Advanced Temporal Graph Embedding (ATGE) framework, a novel and effective approach that combines Multi-Graph Gated Recurrent Units (MGGRU), Attentional Graph Neural Networks (AGNN), and a Multi-Layer Perceptron (MLP) classifier to address three key challenges. The first challenge is Temporal Dependency Modeling. To tackle this issue, we designed a core network (core-net) that extracts critical node features while directly incorporating edge labels into the embedding vectors. The second challenge is Missing Information Prediction. To improve feature completeness and ensure reliable edge classification, we implemented domain-specific models to infer any missing data. The third challenge is Complex Temporal Pattern Recognition. We introduced MGGRU, which effectively integrates AGNN and GRU units to capture intricate temporal dependencies while reducing the effects of missing information. Extensive experiments conducted on real-world cryptocurrency transaction datasets demonstrate that ATGE outperforms state-of-the-art methods, achieving higher AUC and Precision scores and significantly improving detection accuracy. This research makes a significant contribution to the fields of temporal graph learning and anomaly detection, with important implications for fraud detection and the security of cryptocurrency networks.