The topology of low-voltage distribution grids（LVDGs）depicts how various electrical components are physically interconnected within the distribution system. Due to the use of distributed energy resources（DERs）, there is an overlooked mutual load dependent characteristics among end users, which brings great challenges to node correlation analysis and topology identification. In this regard, a low-voltage topology identification method focusing on the mutual load dependent characteristics of DERs is proposed. First, a user classification method based on support vector machineis proposed to classify users according to usage of DERs at different times. Then, convolutional recurrent neural network is applied for distributed feature extraction among load data to decrease load dependency. Finally, the sibling pair search algorithm with residuals resistance is proposed to hierarchically identify the topology. Test results in different simulation scenarios and practical LVDGs demonstrate the effectiveness and robustness of the proposed method.