Duke Energy

Even while long-context large language models (LLMs) have advanced significantly, the supervised fine-tuning (SFT) model's long-context performance is frequently impacted by the poor quality of the LLM-synthesized data, which results in inherent restrictions. Additionally, LLMs may act in ways that are detrimental and inconsistent with human morals. The RL stage does not entail direct comparisons, even if the reward is learned via comparing various replies. The instability of reinforcement learning (RL) is made worse by this discrepancy between the stages of RL and reward learning. We address this by putting forth a novel framework, ReReward, an RL-based technique that rewards long-context based model answers from six human-valued variables using an off-the-shelf LLM as judge. By aligning human feedback in real time, Pairwise Proximal Policy Optimization (PPPO), which learns to improve from direct comparison, mitigates the shortcomings of long-context SFT models.

Storms and floods are more likely in coastal areas. As coastal systems become more socially and environmentally complex, these threats will worsen. To mitigate such impacts, vulnerable coastal areas must be identified and assessed. Modern and future generations are threatened by climate change. Climate change makes natural disasters more frequent, stronger, and unpredictable. Climate change's expected effects—rising sea levels and more powerful and frequent weather events—will make coastal communities considerably more vulnerable to storms, floods, and erosion. The world's shoreline population is expected to triple from 1.8 to 5.2 billion by the 2080s. Weather causes most natural disasters in Korea. Tropical cyclones and high rainfall have caused most disaster damage in the past decade.

This paper introduces a novel framework for estimating inertia from synchronous generators (SGs) and virtual inertia (VI) from inverter-based resources (IBRs) under both large disturbances and ambient conditions. Generator outages induce large disturbances, while ambient conditions are modeled through dynamic load changes. The estimation process begins with Detrended Fluctuation Analysis (DFA) to accurately detect the onset of a disturbance, after which a modified auto-regressive moving average exogenous input (M-ARMAX) model is employed to estimate each generator's inertia constant. The optimal window size for the M-ARMAX model is determined using a minimal variance algorithm. The primary contributions include the application of the M-ARMAX methodology under diverse system operating condition i.e. ambient or disturbances, precise event detection via DFA, and optimizing window selection for accurate inertia estimation. Validation on the IEEE 39-bus transmission system under generator outage and dynamic load conditions, along with tests on real-life event and ambient data from PMU & SCADA in the US Eastern Interconnection, demonstrates that the proposed framework significantly enhances accuracy and efficiency compared to existing methods.