Hybrid Deep Q-Learning and Genetic Algorithm-Based Intelligent Agent Simulation Framework for Dynamic Distribution Network Optimization

Abstract

In modern power distribution systems, ensuring stability, adaptability, and intelligent decision-making under dynamic load conditions remains a major technical challenge. This paper proposes an Intelligent Agent–based Optimization Simulation Framework (IA-OSF) designed to enhance the reliability and operational efficiency of complex medium- and low-voltage distribution networks. The framework integrates a multi-agent perception–learning–optimization cycle, where each agent represents an autonomous decision node capable of sensing local environmental changes, extracting relevant features, and executing optimized reconfiguration strategies through reinforcement feedback. The proposed IA-OSF uses a hybrid GA–DQL learning core to achieve fast convergence and fault-adaptive control. Experimental evaluation using a dynamic power distribution dataset demonstrates that IA-OSF surpasses existing optimization approaches such as PSO, MARL, and GNN-PPO by achieving superior accuracy (97.8%), precision (96.3%), and reliability indices (Delivery Reliability = 0.971, System Resilience = 0.926). The results confirm that intelligent agent collaboration significantly reduces convergence time and enhances recovery performance after fault events. This study provides a reproducible and extensible foundation for developing autonomous, intelligent, and resilient distribution network control systems, offering a theoretical and practical reference for China’s ongoing smart-grid modernization initiatives. The proposed framework was evaluated in a simulated 48-node distribution network environment constructed using a digital-twin architecture. All experiments were executed on a Python-based platform with GA–DQL hybrid optimization, using a dataset comprising 3,456 time-stamped operational records generated from dynamic load and voltage profiles. Performance metrics were obtained under consistent train–test splits (80/20), five repeated simulation runs, and uniform hyperparameter settings to ensure reproducibility and experimental fairness.