Abstract: Compared with the traditional global reconfiguration technology, multi-level dynamic reconfiguration technology is more suitable for improving the operational economics in large-scale urban distribution networks, however, the subsequent reconfiguration level identification has become a new problem. Therefore, relying on the existing mathematical model of multi-level dynamic reconfiguration, this paper proposes a decision-making method of multi-level dynamic reconfiguration for urban distribution networks based on deep learning algorithm, which can be adopted to directly realize the nonlinear mapping between the input net load data and the optimal reconfiguration decision-making scheme without the level identification process. In this paper, a combined neural network combining the feature space attention mechanism, the time attention mechanism, the convolutional neural network and the gated recurrent unit is established. In view of the unbalanced spatial and temporal distribution of the net load data, the feature space attention mechanism and the convolutional neural network are used to learn the spatial characteristics of the net load data in the entire distribution network to perceive the potential relationship of its high-dimensional feature space. Then gated recurrent unit and time attention mechanism are used to fully mine the time-series characteristics of net load data in a long-time scale, extract its unbalanced characteristics in time distribution, and fully train the combined neural network proposed in this paper so as to fit the existing mathematical model. Finally, the effectiveness of the proposed method is verified by numerical examples of 145 real bus system, IEEE 33 bus system, and PG&E 69 bus system.