Detecting overlapping communities is a significant challenge in unsupervised machine learning for network-structured data. While some researchers have applied graph convolutional networks (GCNs) to this problem, most existing GCNs are shallow, limiting their ability to detect communities with larger diameters. Furthermore, the integration of dynamic dilated aggregation into deep GCNs for irregular graphs remains an unresolved issue. To bridge these gaps, we introduce a dynamic residual deep GCN (DynaResGCN), built upon a random dynamic dilated aggregation algorithm within a unified encoder-decoder framework. The encoder utilizes the deep DynaResGCN model, and the decoder uses Bernoulli-Poisson (BP) model. Our framework is tested on three datasets: a topics dataset where ground truth is not available, Facebook graphs with human-labeled ground truth, and large co-authorship networks with less reliable ground truth. Extensive experiments validate that our method is significantly better than many existing methods in detecting overlapping communities in networks.