Deep reinforcement learning-based economic dispatch strategy for automation-oriented renewable grid integration

Abstract

Automation-oriented dispatch centers, renewable microgrids, logistics-park charging hubs, and transportation-energy facilities require rapid and reliable scheduling under volatile load, renewable output, electricity prices, and network-security margins. This paper proposes a constraint-aware deep reinforcement learning (DRL) strategy for economic dispatch in renewable-integrated power grids. The problem is formulated as a constrained Markov decision process (MDP) whose state includes load, wind and photovoltaic output, battery energy storage system (BESS) state of charge, price, and security margin. A convolutional neural network (CNN)-assisted temporal encoder and an actor-critic policy produce continuous set-points, while a lightweight safety projection enforces generator, storage, grid-exchange, ramping, and power-balance constraints. Experiments on a modified IEEE 30-bus system use 1,200 training scenarios and 300 unseen test scenarios derived from public wind and solar profiles. Relative to rule-based, optimization-based, and representative DRL baselines, the proposed method reduces average operating cost to $116,780, renewable curtailment to 2.41%, the violation rate to 0.10%, and the ramp index to 0.631, while maintaining a 0.043-s online decision time. These results indicate that the method can support automated dispatch centers and high-power charging facilities requiring fast, feasibility-preserving decisions.