Temporal knowledge graphs (TKGs) extend traditional knowledge graphs by incorporating temporal information to represent time-sensitive facts as quadruples (subject, relation, object, timestamp), enabling the modeling of dynamic real-world relationships that evolve over time. However, TKGs from different real-world sources are often incomplete and contain complementary information. Therefore, Temporal Entity Alignment (TEA) techniques are needed to integrate knowledge from multiple TKGs by identifying equivalent entities across different temporal knowledge graphs, thereby supporting the consolidation of knowledge from multiple sources. Although there has been extensive research on temporal entity alignment, existing approaches suffer from significant limitations: some methods fail to fully exploit rich semantic information and contextual background, while others that employ large language models incur prohibitively high computational costs. To address these challenges, we propose T3A-LLM, a two-stage triple-information alignment with large language model (LLM), a novel two-stage framework for TKG alignment that efficiently fuses structural, temporal, and semantic information. The first stage uses dual-feature encoding (relation, time) and graph matching for preliminary alignment to a top-n candidate set, reducing the search space and addressing LLM computational overhead. The second stage applies an LLM-Score mechanism for fine-grained semantic reasoning on the candidates, specifically designed to capture deep semantic relationships that traditional structural methods cannot handle. The final scores are obtained by fusing graph-based and LLM-based similarities. Experiments show that T3A-LLM significantly outperforms baselines, with ablation studies confirming the necessity of each component.