Time series clustering is a very effective approach in discovering valuable information in various systems such as finance, embedded bio-sensor and genome. However, focusing on the efficiency and scalability of these algorithms to deal with time series data has come at the expense of losing the usability and effectiveness of clustering. In this paper a new multi-step approach is proposed to improve the accuracy of clustering of time series data. In the first step, time series data are clustered approximately. Then, in the second step, the built clusters are split into sub-clusters. Finally, sub-clusters are merged in the third step. In contrast to existing approaches, this method can generate accurate clusters based on similarity in shape in very large time series datasets. The accuracy of the proposed method is evaluated using various published datasets in different domains.