Dimensionality reduction and feature representation are the key technique and important methods to address the issue of dimensionality curse for time series. Meanwhile, they are a basis task in the field of time series data mining. Therefore, a novel method of dimensionality reduction and feature representation is proposed. An orthogonal polynomial regression model is used to obtain a feature sequence from an original time series. Furthermore, singular value decomposition combining with the fitting results of the feature sequence to time series is used to reduce the dimensionality of feature sequence and obtain another feature sequence with lower dimension to retain most of the information. The results of numerical experiments demonstrate that the novel method can obtain a good effect of clustering and classification in time series data mining under the space with lower dimensionality.