Considering the uncertain characteristics of power grid operation, a reinforcement learning(RL) framework is designed to optimize power grid operation and maintenance. RL constructs an environment-dependent random behavior model. By using the running state of grid components for prediction and learning, the optimal decision-making behavior to maximize expected profits can be determined in an uncertain environment. Artificial neural network (ANN) tool is used to replace tabular representation of state-behavior value function, and a non-tabular RL algorithm integrated with ANN is designed to improve the adaptability of RL algorithm. Test results in small-scale power grid show that compared with reference Bellman optimality strategy, ANN has better Q-learning approximation ability, and collect valid information from predictive state of components, and can be used to support the optimal operation and maintenance strategy.
 

Power grid operation and maintenance,artificial neural network,reinforcement learning,uncertainty