Multi-objective deep reinforcement learning for frequency emergency control with auxiliary voltage deviation mitigation

This paper introduces a Multi-Objective Deep Reinforcement Learning framework for emergency control that addresses frequency–voltage coupling in modern power grids. By jointly achieving frequency security and voltage stability, the proposed scheme solely utilizes active power redispatch and staged load shedding. The problem is formulated as a multi-objective Markov Decision Process and solved using Proximal Policy Optimization, employing a coupling-aware reward function that penalizes frequency violations, steady-state voltage deviations, and control costs. Trained within a high-fidelity dynamic simulation environment, the agent leverages intrinsic active–reactive, P − Q power coupling to regulate voltage through coordinated active power modulation. Case studies on the IEEE 39-bus system demonstrate that the controller maintains the frequency nadir above 59.58 Hz, restores steady-state frequency to 59.91 Hz within 40 s, and limits post-fault voltage deviation to 0.067 p.u. The learned policy exhibits zero-shot generalization to unseen fault locations and load variations from 90% to 110% and outperforms the existing emergency control DRL approach by securing stability with significantly reduced load shedding. Furthermore, robustness tests confirm reliable performance under communication latencies of up to 0.8 s, presenting a practical, resource-efficient solution for integrated emergency stabilization in future grids. • Multi-objective Deep reinforcement learning controller for emergency frequency control with integrated voltage-deviation mitigation. • Smaller frequency drop, reduced steady-state voltage deviation, and faster post-fault frequency recovery under severe events. • Cross-disturbance generalization to unseen faults and operating conditions. • Robust to measurement noise and communication delays, sustaining stable performance.