1 Learning Based Scheme for Islanding and Reconnection Eduardo Cotilla-Sanchez Oregon State University with Carter J. Lassetter, Jinsub Kim This material is based upon work supported by the Department of Energy under Award Number DE-OE0000780
2 Overview and Motivation Larger number of microgrids, what s the role of these clusters on cascading paths. Incorrect reconnection may result in device damage or cascading failures. Protective islanding may yield subnetworks with limited active synchronization ability. Cyber based attacks may compromise the integrity of important system measurements.
3 Modeling a System, the Recipe IEEE RTS-96 test case system with one of its three zones defined as the subnetwork. Poland test case with Zone 5 representing a subnetwork. Both test cases are adapted for dynamic simulation in PSS/E. Set of initial load states. Subnetworks are islanded and stability is tracked upon reconnection.
4 Tracking Key Locations with PMUs Key locations for PMUs are targeted based upon their electrical distance to the Point of Interconnection. Data from these locations hypothesized to provide real-time predictions on the potential stability of reconnection.
5 Machine Learning Secure Reconnections With PMUs placed at key buses, voltage magnitude and voltage angle measurements are available before reconnection. Leverage them as input to Support Vector Machine. Unsafe Reconnection of RTS-96 microgrid The corresponding measurement vector is mapped to a feature domain along with a label based on the stability of subnetwork reconnection (used to train the classifier). Each simulation bears a single label (stable/unstable) in the given feature domain.
6 Classifying Stability With measurements mapped to a feature domain, the two classes may be separated by a hyperplane. Features may be remapped to another dimension for easier separation.
7 Results RTS-96 with limited PMUs showed results for stable (class 1) and unstable (class 0) as follows: Similarly, the Poland case yielded results as shown: Even with limited PMUs, high accuracies could be achieved. % %
8 Automated Mitigation of Cascading Failure in Power Systems Eduardo Cotilla-Sanchez Oregon State University with Jesse Hostetler, and Alan Fern
9 Overview and Motivation Following an initiating event, what actions can we take to minimize negative effects and mitigate cascading failure? Our approach: Formalize control problem as a Markov decision process (MDP) Use standard MDP solver to compute good actions
10 Markov Decision Processes Formalism for sequential decision-making problems: States possible system configurations Actions Probability that doing a in s leads to s Reward How desirable is state s?
11 State Space State variables (not exhaustive!) For each bus: Voltage magnitude Frequency For each generator: Speed Exciter, Governor For each load: MW and Mvar Ø State dimensions in the hundreds or thousands
12 Actions Islanding: Partition grid into zones and isolate zones where failures occur. Load Shedding: Uniformly reduce consumption of all loads in a zone. Figure: [Meier et al., 2014]
13 Rewards Rewards define the objective of the controller Our formulation: R(s) = total active power delivered during single time step Objective: Maximize sum of rewards (i.e., total energy delivered) during the planning period
14 Dynamics We don t have access to, but we can sample from it using a simulator E.g. Cosmic, PSS/E, MATPOWER
15 Online Planning A solution approach for MDPs Fits our requirements: Works with any MDP simulator Scales to large state spaces Algorithm ( policy rollout ): 1. For each action a: A. Simulate doing a in current state s and then letting the system run for a while B. Let Q(s, a) = sum of rewards received 2. Choose action a* that maximizes Q(s, a)
16 Results We evaluated policy rollout on the IEEE 39 case using the Cosmic simulator Avg. total energy (kw s) 1800 1600 1400 1200 1000 800 600 400 200 0 Avg. Total Energy Delivered Rollout No control Avg. power (kw) Avg. Power at Final Time Step 6000 5000 4000 3000 2000 1000 0 Rollout No control
17 Thank you. ecs@oregonstate.edu Enhanced RTS-96 Data: http://web.engr.oregonstate.edu/~cotillaj/ecs/data.html Cosmic Simulator: https://github.com/ecotillasanchez/cosmic