Aiming at the existing clustering algorithms′ lack of non-convex cluster identification ability and parameter sensitivity in extracting typical load curves, a typical load curve extraction method based on the improved density peak clustering (DPC) algorithm is proposed. Firstly, an adaptive cluster centre selection method based on local density and relative distance is proposed to solve the subjective uncertainty problem of artificially selecting cluster centres in the traditional DPC algorithm. Secondly, two new parameters of cluster cross-density and cluster boundary density are defined, and an initial cluster correction strategy is proposed to effectively solve the problem of assigning cascading errors to non-cluster centre points. Comparison experiments with six 2D datasets, four multidimensional datasets and one actual REFIT electrical load measurement dataset show that the proposed improved DPC algorithm outperforms the traditional DPC, K-means and DBSCAN algorithms in three evaluation indexes, namely, accuracy (ACC), adjustment of rand index (ARI) and Fowlkes-Mallows Index (FMI), where they are better than the traditional DPC, K-means and DBSCAN algorithms. Among them, ACC, ARI and FMI are improved by 25.40 %, 46.92% and 21.83 % on average. The results show that the typical load curve extracted by the proposed improved DPC algorithm is more representative, which can provide more accurate data support for the optimal regulation of power system flexibility resources.