Transfer-Reinforcement-Learning-Based rescheduling of differential power grids considering security constraints.

• Model-based method is combined with model-free method. • Actions are mapped to the devices to control the active and reactive power in parallel. • Domain-adaptation-based transfer learning is proposed for the same power grid with different structures. • Parameter-based transfer learning is constructed for different power grids. The power system rescheduling based on model-free methods has obvious defects in practical application, such as poor scenario transferability, long data training time, and waste of domain knowledge. To overcome the above defects, a transfer-reinforcement-learning-based rescheduling method of differential power grids considering security constraints is proposed. When constructing the Markov decision-making process of security-constrained rescheduling, both the off-limits of line power and node voltage are considered in the reward. The action space of deep reinforcement learning is narrowed by calculating the sensitivities of devices and mapped to control the active and reactive power regulating devices to reschedule active and reactive power simultaneously. According to the change degree of transfer object, the applications of transfer learning are divided into two scenarios. For the security-constrained rescheduling transfer scenario of different structures of the same power grid, a domain-adaption transfer learning method is formed, realizing good data adaptability after structure changes with the original model. Moreover, a policy-based transfer learning method is constructed for the security-constrained rescheduling transfer scenario of different power grids, enhancing the training speed and training effect of target power grid. Two standard systems and two actual power grids are utilized to verify the effectiveness of the method. For the actual power grids, the effects of the two scenarios are improved by 5.8% and 3.9% with transfer learning. Compared with other methods, this method not only has obvious advantages in transferability, but also has a shorter learning process and lower control cost. [ABSTRACT FROM AUTHOR]