Entity alignment aims to identify entities from two heterogeneous knowledge graphs that refer to the same real-world object. Concurrently, research on temporal knowledge graphs has been gaining attention. Most existing entity alignment methods for temporal knowledge graphs embed relationships and temporal information into a unified low-dimensional vector space. Some approaches use attention mechanisms to aggregate relational and temporal data for the entities. However, the datasets commonly used exhibit high similarity, which reduces the difficulty of entity alignment. In response, this paper proposes a novel approach called RL-TEA tailored for datasets with low similarity, which incorporates iterative completion and Graph Neural Networks (GNNs). Our approach dynamically completes the knowledge graph while aligning entities by selecting high-confidence entity pairs, using them to augment the training set. This iterative strategy not only improves the training data but also fills the missing parts of the knowledge graph, thus enhancing the overall alignment accuracy. Experiments on four low-similarity datasets have demonstrated the effectiveness of our model.