The increasing integration of renewable energy sources renders conventional grid fault detection and contingency response strategies inadequate. The stochastic nature of wind generation induces current amplitude fluctuations with highly time-varying characteristics, causing conventional overcurrent or directional protection schemes with fixed settings to experience misoperation or failure to operate. While existing learning algorithms can enhance classification accuracy, their reliance on offline training data limits adaptability to real-time, dynamically evolving grid conditions. The performance of parameter tuning (learning rate and discount factor) of traditional methods is insufficient in different scenarios. Addressing the challenge of real-time steady-state fault isolation, this work proposes a novel Deep Reinforcement Learning Proximal Policy Optimization -based algorithm for adaptive fault section isolation in dynamic environments. Key innovations include constructing a high-dimensional state space incorporating wind power output and prediction error to capture transient characteristics, designing a discrete-continuous hybrid action space enabling simultaneous control of circuit breakers and adaptive adjustment of protection settings, and refining the reward function structure to optimize policy learning. Simulations conducted on a real-time digital simulator platform demonstrate that the proposed algorithm achieves a fault isolation accuracy of 98.7% under test scenarios, with wind power penetration levels ranging from 20% to 50%. This represents a 23.5% improvement over conventional methods. Furthermore, the algorithm achieves a response time consistently less than 80 ms. The universality test also verifies that the compensation mechanism has a good wide range of adaptability. The primary contribution of this work lies in providing a setting-free, adaptive protection scheme for distribution networks with high renewable penetration. This approach can significantly enhance grid reliability and bolster autonomous fault recovery capabilities.
Lyu et al. (Sun,) studied this question.