Collaborative optimization of multi-park integrated energy systems is essential for enhancing multi-energy coupling and complementarity

reducing carbon emissions

and lowering operational costs. However

this process faces significant challenges

including heterogeneous privacy protection among multiple agents and the need for real-time decision-making. To address these issues

this paper proposes a low-carbon economic dispatch model that balances personalized decision-making for each park with privacy protection. The model employs an offline training and online decision-making approach to meet real-time requirements. At the individual park level

an improved Deep Deterministic Policy Gradient (DDPG) reinforcement learning method is utilized to develop a dispatch model that comprehensively considers electric and thermal loads

time-of-use electricity pricing

and tiered carbon trading. Additionally

embedded vectors are incorporated into the DDPG neural network to enhance the agent's perception and adaptability to the operational characteristics of each park. To ensure data privacy and address the challenges of multi-agent heterogeneity

a personalized federated framework is established for collaborative training across multiple parks within the integrated energy system. In this framework

each park's DDPG model transmits only model parameters

excluding embedded vectors

to a central aggregator

thereby preventing the direct exchange of operational data between parks. Simulation experiments conducted across multiple scenarios and algorithms demonstrate that each park can achieve global optimization using only local model information

thereby facilitating system-wide low-carbon economic dispatch.