With the rapid advancement of the Internet of Things and 5G technologies, edge computing has emerged as a vital paradigm for supporting real-time processing and low latency applications. At the core of edge computing lies Virtual Network Embedding (VNE), whose resource allocation efficiency directly influences both service quality and revenue generation in edge environments. Recent research has identified reinforcement learning (RL) as a promising approach to enhance the VNE process. However, most existing RL-based methods overlook the potential of attention mechanisms, which can help agents better understand complex network topologies and thereby improve embedding performance. To address this gap, we propose a novel algorithm, D2ARL (Dual Dynamic Attention-driven Reinforcement Learning), which adaptively captures the dynamic characteristics of both physical and virtual networks through a dynamic attention mechanism. D2ARL is built on a sequence-to sequence (seq2seq) architecture, leveraging dynamic attention in the encoder to capture local features of the networks, and in the decoder to extract globally fused features. Experimental results show that D2ARL outperforms state-of-the-art methods across various network environments. It achieves higher overall benefits and demonstrates stable performance under diverse conditions. Notably, D2ARL improves long-term average revenue by 14.5% compared to the best-performing existing method.