A Two-Stage PPO-RLMPA Framework for Dynamic Economic Dispatch with Renewable Energy and Storage Integration

Area of Science:

• Power Systems Engineering
• Artificial Intelligence
• Optimization Algorithms

Background:

• The Dynamic Economic Dispatch (DED) problem is crucial for cost-efficient power system operation.
• Non-convexities arise from valve-point loading, ramp-rate coupling, and integrating intermittent renewables (wind, PV, PSH).
• Existing methods like metaheuristics and deep reinforcement learning have limitations in computational cost and solution feasibility.

Purpose of the Study:

• To propose a novel Two-Stage PPO-RLMPA framework for solving the non-convex DED problem.
• To improve computational efficiency and guarantee the feasibility of dispatch schedules.
• To leverage biomimetic strategies for enhanced economic dispatch optimization.

Main Methods:

• A Proximal Policy Optimization (PPO) agent is trained on a Markov Decision Process (MDP) formulation of DED.
• A deterministic Safety Layer ensures policy actions remain within physical constraints (capacity, ramp-rate, power balance).
• A Marine Predators Algorithm (MPA) refines the PPO dispatch, incorporating biomimetic foraging and online adaptation via a Deep Q-Network.

Main Results:

• The Two-Stage PPO-RLMPA framework achieved best costs of USD 368,763 and USD 737,348 on benchmark systems.
• This represents cost reductions of approximately 1.1% and 4.4% compared to the CFCEP baseline.
• The framework demonstrated zero post-repair constraint violations across all independent runs.

Conclusions:

• The proposed Two-Stage PPO-RLMPA framework effectively addresses the challenges of non-convex DED problems.
• It offers a robust and efficient solution by combining data-driven learning with biomimetic optimization.
• The method ensures reliable and economically superior power system dispatch schedules.