Abstract
Integrating renewable energy into power systems introduces significant challenges in balancing generation costs and grid stability, necessitating advanced solutions for the Economic Dispatch Problem (EDP). While classical mathematical and meta-heuristic methods face scalability and computational efficiency limitations, reinforcement learning (RL) offers a promising alternative due to its adaptability to high-dimensional and dynamic environments. This study employs Twin Delayed DDPG (TD3), an enhanced version of Deep Deterministic Policy Gradient (DDPG). TD3 integrates Prioritized Experience Replay (PER) and Noisy Networks (Noisy Nets) for the EDP in a regional microgrid with photovoltaic (PV) generation. PER improves sample efficiency by prioritizing high-error transitions, while Noisy Nets enhance exploration through adaptive parameter noise. Experiments demonstrate that combining these techniques with TD3 achieves a 54.6% reduction in testing operation cost and a 95.3% decrease in cumulative power unbalance compared to the baseline TD3. The improvements are validated across various deterministic and stochastic RL models, with TD3+PER+Noisy Nets outperforming others in cost efficiency and stability. The findings demonstrate the proposed approach's capability to optimize microgrid dispatch while providing a scalable and practical framework for power system control.