Temporal Knowledge Graph Reasoning (TKGR) seeks to predict unknown facts by leveraging historical data, playing a crucial role in various applications. However, existing models for TKGR have three main limitations: (1) The relational information embedded in short-term historical data remains insufficiently utilized, and the varying contributions of different historical snapshots are ignored. (2) The mining of long-term historical information is insufficient. (3) Both short- and long-term historical information are not considered together. Therefore, we propose a method that combines improved Relational Embedding and Multi-Hop Sampling for temporal knowledge graph reasoning (REMHS). First, we propose an improved relational embedding learning method that uses an undirected graph to learn relational features. An attention mechanism is introduced to weigh the contributions of different temporal snapshots to the query facts. Second, we present an enhanced multi-hop sampling technique to better mine more long-term historical data. Third, we dynamically fuse the long- and short-term historical information using an attention mechanism. Experimental results across five public datasets demonstrate that our method outperforms all baseline models.