Abstract: Clustering analysis is the basic method for the load characteristic classification and synthesis. Aiming at the shortcomings in clustering quality and robustness of existing clustering methods applied to online load modeling based on the power grid big data platform, this paper proposes an improved piecewise linear representation（IPLR） method for daily load curve dimension reduction.Based on the advantages of IPLR for adaptive dimension reduction and reconstruction of data sets, combined with the characteristics of dynamic time warping（DTW） distance which is suitable for measuring the similarity between time series of unequal dimension, a daily load curve clustering method combining IPLR and DTW distance is constructed. Firstly, according to the variation of adjacent and interval sampling points of load curves, the characteristic points of load curves are extracted, and the curves are reconstructed by adaptive dimension reduction. Then, the DTW distance is taken as the similarity measurement index,and the clustering analysis of dimension reduction data is carried out by using Canopy based K-means（CK-means） algorithm. The method proposed is applied to the classification and synthesis of the daily load curves of typical consumers in a provincial power grid of China. The results show that the proposed dimension reduction method matches with the similarity measurement method, has good comprehensive performance, and is suitable for the analysis of the industrial composition ratio of synthesis load in substations.