The development of Personalized Emotional Support Conversation (PESC) systems encounters significant challenges due to the scarcity of specialized training datasets, which restricts their ability to generate responses tailored to user-specific traits beyond conventional Emotional Support Conversation (ESC) capabilities. To address this issue, we propose a two-stage synthetic data generation framework. Initially, we performed Structured Character Profile Synthesis, in which DeepSeek-V3 generates JSON-formatted profiles that incorporate sampled traits. Subsequently, in the Multi-Agent Dialogue Simulation stage, Qwen2.5-32B-instruct agents role-play. This role-guided strategy ensures that dialogues are aligned with user characteristics. Through supervised fine-tuning experiments, we demonstrated the utility of the synthetic data, with a 2.68-point G-score increase on the validation set. Our pipeline provides a solution for dataset expansion while maintaining character consistency within dialogues. We further validated the effectiveness of our synthetic data by achieving state-of-the-art performance with a 92.67 G-score on the NLPCC-2025 Shared Task 8 benchmark, the highest among submitted systems, and attained third place overall with a score of 38.99.