Research On Ill Conditioned Power Flow Automatic Adjustment Strategy Based On Deep Reinforcement Learning

Power flow adjustment is one of the most complicated contents in power grid calculation,and the convergence adjustment of power flow depends heavily on expert experience.In this paper,deep reinforcement learning algorithm is used to study the adjustment of ill conditioned power flow,and the convergence of power flow is improved by combining knowledge and experience.The fundamental cause of ill conditioned power flow is the imbalance of active and reactive power distribution,the excessive transmission power of some sections or the voltage level approaching the limit.In this paper,the deep reinforcement learning algorithm is used to model the ill conditioned power flow.By adding knowledge and experience,the action space is reduced,the process of manual adjustment is simulated,and the active power and reactive power are balanced,so that the power flow converges.This paper introduces in detail how to explore the environment and carry out effective learning in the process of power flow adjustment.The main contents include theoretical analysis of ill conditioned power flow,deep reinforcement learning principle,automatic adjustment scheme research and model establishment,algorithm improvement.The feasibility of ill conditioned power flow automatic adjustment strategy is verified by IEEE 118 bus and actual power grid.The specific research contents of this paper are as follows:(1)The deep Q-Learning algorithm of power grid convergence adjustment is established and the ill conditioned power flow is modeled.Combined with knowledge and experience,the state action space and reward mechanism of reinforcement learning are designed.The power imbalance caused by ill conditioned power flow is adjusted.Through sample training,the agent can learn effective strategies,so that the power flow can converge automatically.(2)As the active power balance is global balance and easy to adjust,the reactive power imbalance is preferred.The action of reinforcement learning is connected with the load node.When the agent outputs the action,the capacitor is input to the load node for reactive compensation.In the reward setting,the empirical index to measure the sick degree is introduced.The agent finds the weak point of voltage through training,and improves the voltage level by putting capacitor into the weak area,so as to make the power flow converged.(3)In order to be more in line with the actual engineering situation,the ill condition caused by the unbalance of active and reactive power is considered comprehensively.That is to say,active power balance first and then reactive power balance.In order to improve the training efficiency and success rate,the algorithm is improved on the basis of(2).Considering the priority of successful samples of power flow convergence,the learning efficiency of agent is improved by learning these valuable samples first.In the setting of reinforcement learning elements,knowledge and experience are added to reduce the action space and balance the active power in different regions,so as to make the convergence speed of ill conditioned power flow faster.Finally,the effectiveness of the algorithm is verified by IEEE 118 bus and actual power grid.