Abstract: Under the background of new power system, load fluctuation at the end of the distribution network is complex, and load curve clustering is an effective method to simplify the analysis of load fluctuation characteristics. For the loads at the end of the distribution network, the mainstream single-day load curve clustering ignores the difference of daytime fluctuation, but the multiday load curve is difficult to achieve the ideal clustering effect. In view of this, this paper proposes a new clustering method for multi-day load curve based on the concept of network structure reduction. Firstly, the multilayer associated load time series network of clustered users is generated by horizontal visibility graph method. Secondly, the structural redundancy of the associated multilayer network is evaluated and the similar layers are aggregated. Finally, the interpretability of the multi-day load curve clustering results is analyzed from three aspects: trend effect, fluctuation trend and social attribute. The experimental results show that, compared with the mainstream clustering methods, this method can efficiently deal with the clustering problem of multi-day load curve, and has the advantages of free parameter tuning in the clustering process and strong interpretability of the clustering results.