Research On Power Grid Attack And Protection Methods Based On Reinforcement Learning

Nowadays,large-scale power outages caused by cascading failures occur frequently around the world.It is of great significance to comprehensively analyze and study the cascading failures progress to ensure the safety and reliability of power grids.Based on the simulation data of power grids,this paper studies the process and control strategy of power grid cascading failures by means of network theory and machine learning method.The main contents are as follows:1)To analyze the attack methods of smart power grid considering critical power lines,a sequential attack method based on reinforcement learning is studied and an improved reinforcement learning exploration strategy is proposed.Compared with the existing power grid attack methods analyses,the power grid attack methods analysis considering the critical lines is more consistent with the actual situation of the power grid.Compared with the exploration strategies used by traditional reinforcement learning methods,the improved exploration strategy can improve the exploration efficiency of reinforcement learning and further accelerate the search for the optimal attack sequence.Furthermore,the critical lines are defined based on the power betweenness,and the variation of cascading failures scale when the vulnerable lines are protected is analyzed.2)Aiming at the problem of failure diffusion of power grid after initial attack,a line switching off strategy based on reinforcement learning is proposed.Considering the complex influence of cascading failures on the power grid,a comprehensive evaluation index of power grid loss combining topology and electrical characteristics is presented.The line switching off method can reduce the scale of cascading effectively by using less cost of line switching off to avoid a large-scale overload.Most of the existing line switching off studies use a pure topological model for simulation and only consider the line switching method of specific failures,which don't work well with real power grids.The reinforcement learning based line switching method can not only search the optimal switching off sets for specific failures offline,but also give the line switching off sets for random failures online with the help of a trained reinforcement learning model.3)To solve the problem of grid vulnerability to cascading failures using the traditional line capacity allocation method,an improved capacity allocation method is proposed.The capacity of the line in the grid is usually set as a multiple of its initial load in traditional capacity allocation methods.Lines with low initial loads usually have lower capacity and are prone to overload,which can lead to cascading failures.The proposed method distributes a portion of the allocable capacity equally among all lines and the other portion according to the initial load of the line.Compared with the traditional distribution methods,this method can reduce the vulnerability of power grid to line failures.Based on the multi-angle study of the process of power grid cascading failures,this paper presents a new idea of power grid attack methods and specific methods of defense against cascading failures.The analysis process and the results can be used as reference for power grid design and operation personnel,and have certain engineering application value.