One main application of electrical resistivity tomography (ERT) is the non-invasive detection of geological or hydrological structures in the shallow subsurface. This paper investigates the capability of time-series clustering to retrieve such features on real time-lapse ERT datasets considering three aspects: (1) the comparison between three clustering algorithms k-means, hierarchical agglomerative clustering (HAC), and Gaussian Mixture Model (GMM), including the question of the optimal choice of cluster number and the identification of resistivity series whose classification is uncertain, (2) the effect of adding a spatial constraint in clustering, and (3) the robustness of the approaches to various representations of resistivity values and the number of time-steps involved in the clustering. The real time-lapse ERT dataset is obtained from dipole-dipole arrays on a 48 electrodes profile installed on the top of the Rochefort cave in Belgium. It consists of resistivity time-series defined over 465 days and associated with 1558 cells of the 2D ERT models derived from a time-lapse inversion. The clustering results are appreciated using clustering validation indices and further confronted with the expert-based structural model of the site. Results show that the three clustering algorithms provide similar spatial patterns on the standardized data and reveal correlated resistivity time-series. Some clusters are, however, spatially split and regroup time-series with a wide range of mean resistivity, suggesting different geological units within these groups. Clustering on the raw resistivity time-series may also appear inconsistent as the averaged resistivity series per cluster are highly correlated, thus missing the hydrological and functional traits of the subsurface elements. Applying a spatial constraint to the clustering of standardized data increases the number of clusters in order to retrieve spatially tied clusters. The grouped series are more homogeneous in terms of mean resistivity due to their spatial proximity, but some inconsistencies may remain due to synchronous hydrological forcing. Applying the clustering to various time-series representations allows us to gain confidence about the redundant spatial patterns. However, the patterns obtained from the clustering of the full standardized dataset cannot be reproduced from continuous sub-samples up to 100 days, but well from less than 20 samples picked randomly over the 465 days. Accordingly, our study highlights the importance of time-variable parameters in the identification of structural facies and hydrofacies with ERT while demonstrating the strength of long-term monitoring.