<p>Most faults in the medium voltage and low-voltage networks are related to single phase-to-ground faults with currents lower than the relay pick-up currents settings (CLP faults). In some networks such as resistance-grounded networks and resonant grounding systems, it is difficult to detect and classify these types of faults and cause fires and damage to the network and consumers. This paper proposes a multi-layer protection scheme for active distribution networks integrated with AC and DC microgrids using different agents, intelligent electronic devices, and deep neural networks (DNNs) to classify faults including single phase-to-ground faults. In this way, 4 types of layers, namely, the microgrid layer (the first layer), active distribution network (ADN) layer (the second layer), substation layer (the third layer), and system layer (the fourth layer) are utilized to protect the system, comprehensively. If the protective equipment cannot detect a fault in the ADN layer (the second layer), the protection agent in the substation layer can recognize the fault and determine its location. On the other hand, due to the changing topology in the post-fault conditions, a procedure for ADN protection using a multi-layer procedure is proposed. Also, the impact of different DNN structures on fault classification performance is investigated. This paper has used DIgSILENT Power Factory, MATLAB software, and Python programming language (Tensorflow platform and Keras library) for simulation, feature extraction, and DNN training processes. The numerical results demonstrate the high accuracy of the proposed scheme for classifying CLP faults.</p>