Prior efforts have shown that deep reinforcement learning (DRL) may provide a new method for controlling networked power systems. Though successful, prior approaches have not yet demonstrated their behavior on systems of realistic scale. This effort examined multiple theoretical and technical approaches to allow a DRL model to operate over a system of 2,000 buses or more. We find that allowing the DRL models to run training episodes in parallel provides near limitless efficiency gains, allowing us to train successful agents to behave on our Kuramoto transmission model of up to 4,000 buses. We further show that we can expand our PowerWorld DRL implementation to systems of up to 25 buses but struggle to go beyond this limit due to PowerWorld’s inability to run multiple instances at once. Finally, we examine a multi-agent approach and find that it performs as well if not better than our existing centralized approach.
In this research, artificial intelligence and machine learning (ML) methods are used to search an uncertain parameter space more efficiently for the most important inputs with respect to response sensitivities. These methods are applied to the Extremely Low Probability of Rupture (xLPR) probabilistic fracture mechanics code used at the U.S. Nuclear Regulatory Commission (NRC) in support of nuclear regulatory research. This report documents two separate but related sub-tasks: (1) ranking important uncertain input features with respect to target outputs, determined by convergence in confidence intervals for increasing sample sizes using simple random sampling; and (2) implementation of a reduced-order surrogate model for fast, approximate sample generation. Unoptimized readily available off-the-shelf ML models were used in both sub-tasks.
Uncertainty in severe accident evolution and outcome is driven by event bifurcations that represent distinctive challenges to defensive layers and tend to promote the emergence of discrete classes of core damage and accident risk. This discrete set of "attractor" states arise from the complex networks of competing physical phenomena and conditional event cascades occurring as the overall system degrades – a process that yields increasing degrees of freedom and accident progression pathways. Characterization of these event spaces has proven elusive to more traditional data interrogation methods, but proves tractable by application of more advanced data collection and machine learning approaches. Through application of these approaches we demonstrate a conceptual framework that enables real-time/robust, risk-informed decision-making support to improve accident mitigation and encourage “graceful exits” during low probability, extreme events limiting accident consequences. In this analysis, we simulated over 8,000 short-term station blackout (STSBO) accidents with the state-of-the-art integral severe accident code, MELCOR, and demonstrate the potential for ML approaches to predict simulation outcomes. We chose to pair ML tools with interpretable and mechanistic event trees for the considered STSBO accident space to predict the likelihood of future event paths along the tree. In addition to the current state of the system, we use information from recent trajectories of temperature, pressure, and other physical features, combining both the current state and past trajectories to forecast future event paths. Finally, we simulate the random injection of variable amounts of water to quantify the efficacy of available actions at reducing risks along the many branches in the event tree. We identify scenarios and windows of opportunity to mitigate risk as well as scenarios in which such actions are unlikely to alter the accident end-state.
In recent years, infections and damage caused by malware have increased at exponential rates. At the same time, machine learning (ML) techniques have shown tremendous promise in many domains, often out performing human efforts by learning from large amounts of data. Results in the open literature suggest that ML is able to provide similar results for malware detection, achieving greater than 99% classifcation accuracy [49]. However, the same detection rates when applied in deployed settings have not been achieved. Malware is distinct from many other domains in which ML has shown success in that (1) it purposefully tries to hide, leading to noisy labels and (2) often its behavior is similar to benign software only differing in intent, among other complicating factors. This report details the reasons for the diffcultly of detecting novel malware by ML methods and offers solutions to improve the detection of novel malware.
