Deep Learning for Power System Restoration

 Author:  Alexander Moses Danielsson                                                                                                     Year: 2018                                                                                                                                           Supervisors:  Ragnar Kristjánsson and Samuel Perkin

Abstract

Power systems experience outages due to uncontrollable circumstances despite efforts to reduce the frequency of such events. In an attempt to minimize the negative impact of outages, a deep feed forward neural network (FFNN) was trained to perform optimal actions during power system restoration. First, as a prerequisite, a Restoration Model (RM) was developed to simulate system restoration. The RM was designed to handle any topological degradation of a system and enable interactive exploration of the actions required to bring the system back to an ideal state. Moreover, a cost function was developed in order to evaluate the quality of a given sequence of actions in the context of restoration cost. Using the RM as a simulation environment and the cost function as an evaluation measure, a dataset of optimal power system state-to-action pairs was created using a genetic algorithm (GA) by optimizing restoration action sequences on the Icelandic transmission system. The FFNN was trained via supervised learning using the created data, achieving a 75% test accuracy on optimal decisions on the Icelandic system. The FFNN agent was further tested in a comparison to the GA and operators of the Icelandic system. Results show that the FFNN is 3 orders of magnitude faster than the GA at developing a restoration plan, and performs comparably to the human operators on a simple test restoration scenario. This thesis demonstrates the feasibility and potential of using deep learning for power system restoration and control.