Search

• A comprehensive evaluation on power flow solvability guides the search direction of learning. • Combining GCN and RNN to extract the spatial-temporal information makes the learned information more valuable. • Action space of reinforcement learning is narrowed according to knowledge. • For actual power system, transmission channel off-limit is a key factor for unsolvable problem. In view of the problem of manpower and time consumption caused by unsolvable power flow in current AC / DC hybrid power grid with new energy, a power flow state assessment and solvability recovery adjustment method based on graph convolutional recurrent neural network (GCRNN) and deep double Q-learning network (DDQN) is proposed. First, active power balance degree, reactive power balance degree and channel transmission margin are defined and weighted together to evaluate the solvability of power flow. Then, a power flow state evaluation model based on GCRNN is proposed by combining graph convolution with recurrent neural network, where the input is tangent vector of power flow equations, and the output is comprehensive evaluation index of solvability. Next, the state space, action space and multiple rewards of deep reinforcement learning are formed. By introducing the search direction, learning information extraction and search space constraints, an automatic adjustment method to restore power flow solvability is designed. Finally, the simulation results show that the proposed method performs well both in New England 39-bus standard system and actual power grid of China. Compared with the results obtained by other methods, the superiority of the proposed method is verified. [ABSTRACT FROM AUTHOR]