Data-driven methods for real-time dynamic stability assessment and control

Sammanfattning: Electric power systems are becoming increasingly complex to operate; a trend driven by an increased demand for electricity, large-scale integration of renewable energy resources, and new system components with power electronic interfaces. In this thesis, a new real-time monitoring and control tool that can support system operators to allow more efficient utilization of the transmission grid has been developed. The developed tool is comprised of four methods aimed to handle the following complementary tasks in power system operation: 1) preventive monitoring, 2) preventive control, 3) emergency monitoring, and 4) emergency control. The methods are based on recent advances in machine learning and deep reinforcement learning to allow real-time assessment and optimized control, while taking into account the dynamic stability of a power system. The developed method for preventive monitoring is proposed to be used to ensure a secure operation by providing real-time estimates of a power system’s dynamic security margins. The method is based on a two-step approach, where neural networks are first used to estimate the security margin, which then is followed by a validation of the estimates using a search algorithm and actual time-domain simulations. The two-step approach is proposed to mitigate any inconsistency issues associated with neural networks under new or unseen operating conditions. The method is shown to reduce the total computation time of the security margin by approximately 70 % for the given test system. Whenever the security margins are below a certain threshold, another developed method, aimed at preventive control, is used to determine the optimal control actions that can restore the security margins to a level above a pre-defined threshold. This method is based on deep reinforcement learning and uses a hybrid control scheme that is capable of simultaneously adjusting both discrete and continuous action variables. The results show that the developed method quickly learns an effective control policy to ensure a sufficient security margin for a range of different system scenarios. In case of severe disturbances and when the preventive methods have not been sufficient to guarantee a stable operation, system operators are required to rely on emergency monitoring and control methods. In the thesis, a method for emergency monitoring is developed that can quickly detect the onset of instability and predict whether the present system state is stable or if it will evolve into an alert or an emergency state in the near future. As time progresses and if new events occur in the system, the network can update the assessment continuously. The results from case studies show good performance and the network can accurately, within only a few seconds after a disturbance, predict voltage instability in almost all test cases. Finally, a method for emergency control is developed, which is based on deep reinforcement learning and is aimed to mitigate long-term voltage instability in real-time. Once trained, the method can continuously assess the system stability and suggest fast and efficient control actions to system operators in case of voltage instability. The control is trained to use load curtailment supplied from demand response and energy storage systems as an efficient and flexible alternative to stabilize the system. The results show that the developed method learns an effective control policy that can stabilize the system quickly while also minimizing the amount of required load curtailment.