Nonlinear-Constrained Dynamic Load Scheduling in Urban Microgrids via Deep Deterministic Policy Gradient Optimization

Abstract

With the increasing share of renewable energy, dynamic load scheduling in urban microgrids faces the challenge of high-dimensional nonlinear constraints. Existing studies mainly rely on linear approximations or heuristic rules, which make it difficult to achieve globally optimal scheduling under complex constraints and lack rigorous mathematical convergence analysis. To address this issue, this paper proposes a nonlinear-constrained dynamic scheduling method based on Deep Deterministic Policy Gradient (DDPG). First, the hard constraints are transformed into penalty terms in the objective function through the Lagrangian relaxation method, and their mathematical equivalence to the original problem is proven. Second, an action correction mechanism based on Lyapunov stability is designed to ensure feasible solution generation during the policy iteration process. Finally, a differential-flatness-based dimensionality reduction mechanism is introduced to lower the computational complexity of policy search. Experiments conducted on a microgrid model with real data and a scale comparable to the IEEE 33-bus system show that, compared with traditional Model Predictive Control (MPC) and reinforcement learning baselines (e.g., DDPG), the proposed method reduces scheduling costs by 12.7% (relative to DDPG) and improves training efficiency by 19.3% while satisfying all constraints. This study provides a verifiable mathematical framework for high-dimensional nonlinear constrained optimization problems and an extensible solution for real-time scheduling in microgrids.