| As new energy power generation technologies emerge,power systems continue to grow in scale and complexity,presenting significant challenges to their safe and stable operation.Emergency control serves as the ultimate safeguard for power system security and stability.However,traditional emergency control methods rely on accurate physical models and parameters,which can be difficult to adapt to modern power grids.Reinforcement learning offers a solution by eliminating dependence on precise physical models and enabling the discovery of alternative control strategies.In large-scale and complex power systems,reinforcement learning often requires continuous power constants as state inputs.Traditional reinforcement learning methods struggle to handle continuous state inputs,necessitating the integration of deep neural networks.This thesis proposes a deep reinforcement learning algorithm combined with a deep neural network for emergency control in power systems.We establish a comprehensive emergency control framework based on deep reinforcement learning.The main research contributions include:(1)We focus on the use of deep reinforcement learning for emergency control in power systems,specifically in response to the situation when a three-phase short-circuit fault is considered as disturbance that is highly detrimental to the power system.The fast valve control commonly used in thermal power plants is employed as the emergency control method to maintain system stability by restoring the power balance between the generator and the power grid.To address the continuous action space of fast valve control,this thesis adopts the Deep Deterministic Policy Gradient(DDPG)algorithm and proposes an emergency fast valve control strategy based on DDPG to control the thermal power plant’s valve more quickly and accurately.The effectiveness and generality of the proposed algorithm are demonstrated through the simulations of multiple fault scenarios.(2)We focus on the use of brake resistance to control power systems with hydropower stations,considering that the water guide vane of a hydropower station cannot be shut as quickly as the steam gate of a thermal power plant.The deep Q network(DQN)algorithm is employed for the discrete action space of brake resistance,and the Dueling DQN algorithm with faster convergence is adopted.Two different transient demands are taken into account,and a reinforcement learning reward function is designed accordingly.Simulations of both reward function schemes demonstrate that they can satisfy different requirements of transient processes. |
