The pervasive adoption of inverter-interfaced resources in 100% renewable power systems fundamentally alters voltage regulation dynamics and renders classical margin - estimation techniques computationally prohibitive. This paper introduces a unified, Jacobian-free Krylov-Arnoldi (JFKA) framework for automated estimation of system static voltage stability (SVS). First, inverter current-limiting behavior is captured by a smooth S-type function, preserving continuous power-flow structure across both grid-forming (GFM) and gridfollowing (GFL) control modes. Building on singularity theory, we derive a Jacobian-free stability indicator that pinpoints the onset of voltage collapse without explicit derivative evaluation. To efficiently solve the resulting large-scale, transcendental power-flow equations, we embed a reduced-order Arnoldi process within a Newton-Krylov solver, yielding rapid convergence and markedly lower memory footprint. Case studies on a modified IEEE-39 bus network with full renewable penetration demonstrate that our method accurately tracks voltage regulation limits under varied load-growth scenarios and automatic mode switches.

The multimodal dynamic stability (MDS) is a critical issue for constructing power systems with 100% renewable energy (PSRE). This paper proposes a multimodal unified control (MUC) method for enhancing the MDS in PSRE. First, an improved Heffron-Phillips model is established to demonstrate the mechanism of multimodal dynamic instability (MDI). Next, a third-order external subsystem model of the grid-forming renewable energy source (GFM-RES) is derived by the Lie derivative. Then, the MUC architecture is designed by utilizing the Lyapunov theory in the third-order external system. Finally, the application of the proposed MUC method is verified by analyzing the pertinent results for the modified IEEE 11-bus system with a 100% renewable energy generation.Note to Practitioners—The construction of PSRE has attracted widespread attention from the academic and engineering communities. Dynamic stability is one of the key issues faced in building PSRE. This work presents a MUC of GFM-RESs for improving the MDS of the PSRE. Practitioners should be able to apply the MUC to GFM-RESs like battery energy storage system, wind turbines and photovoltaic units. The MUC is designed using the Lie derivative and Lyapunov theory. From a practical point of view, the MUC achieves the MDS from a control perspective, without additional investment required.

This paper proposes a novel voltage-adaptive strategy (VAS) considering current limits of renewable energy resources (RESs), to enhance the transient stability of the power system with 100% renewable energy (PSRE). First, taking the current limits into account, a new transient stability mechanism is revealed by deriving the fault critical clearing time (CCT) of a PSRE with two RESs. Next, leveraging the Lie derivative and the Lyapunov theory, a novel adaptive control method is proposed, which is more in line with the output saturation characteristic of RESs. Then, VAS is formulated based on the proposed adaptive control method for enhancing the transient stability of PSRE. Finally, the proposed VAS is verified on a modified IEEE 11-bus system with 100% renewable energy generation.Note to Practitioners—Transient control can be considered one of the challenges in power system field for constructing the PSRE, which has significant implications for reducing carbon emissions. This work presents a VAS of RESs taking the current limits into account for improving the transient characteristics of the PSRE. Practitioners should be able to apply the VAS to RESs represented by wind farms and photovoltaic power stations. The VAS implements the transient control through the Lie derivative and the proposed adaptive control law. From a practical point of view, it should be highly emphasized that the proposed VAS can achieve transient control of the PSRE without increasing any investment and bringing negative impacts.

Training large language models (LLMs) encounters challenges in GPU memory consumption due to the high memory requirements of model states. The widely used Zero Redundancy Optimizer (ZeRO) addresses this issue through strategic sharding but introduces communication challenges at scale. To tackle this problem, we propose Lins, a system designed to optimize ZeRO for scalable LLM training. Lins incorporates three flexible sharding strategies: Full-Replica, Full-Sharding, and Partial-Sharding, and allows each component within the model states (Parameters, Gradients, Optimizer States) to independently choose a sharding strategy as well as the device mesh. We conduct a thorough analysis of communication costs, formulating an optimization problem to discover the optimal sharding strategy. Evaluations demonstrate up to 52% Model FLOPs Utilization (MFU) when training the LLaMA-based model on 1024 GPUs, resulting in a 1.56 times improvement in training throughput compared to newly proposed systems like MiCS and ZeRO++.