Cryptocurrency-based digital payments are integral to modern financial systems, creating complex transaction networks influenced by user interactions. However, fraud within these networks presents significant risks, underscoring the need for more effective detection mechanisms. Traditional methods, often relying on static graph representations, struggle to address the dynamic nature of fraudulent activities and the problem of incomplete data. This paper proposes an Optimized Temporal Graph Embedding Model (OTGE) for fraud detection, combining Multi- Graph Long Short-Term Memory (MGLSTM) and a two-layer Support Vector Machine (SVM) classifier. The OTGE model tackles three key challenges: temporal dependency modeling, missing information prediction, and capturing complex temporal patterns. First, it employs a reachability network (reach-net) to accurately model temporal dependencies, generating embedding vectors based on node and edge labels. Second, it predicts missing data using domain models, enhancing node features and improving edge classification. Third, the MGLSTM component incorporates attention-based graph neural networks (AGNN) and LSTM units to capture intricate temporal patterns and predict missing information in the evolving network structure. The two-layer SVM improves classification by utilizing node embeddings and predictions from the first layer. Experiments on real-world transaction networks demonstrate that OTGE significantly outperforms baseline methods, achieving superior Area Under Curve (AUC) and Precision scores.