Conventional small-signal stability assessment methods based on graph neural network face limitations in capturing the cross-scale dynamic features of power systems. Message-passing models easily suffer from over-smoothing when stacked deeply to capture long-range dependencies, while global self-attention models, despite directly modeling long-range relationships, tend to lose local details and incur high computational costs. Addressing these issues, this paper proposes a Dual-Branch Synergistic Contrastive Graph Neural Network (DB-CoGNN). The architecture designs a parallel dual-branch structure: a local encoder preserves crucial topological inductive bias to precisely capture neighborhood dynamics, while a global encoder employs a sparse multi-hop mechanism to effectively capture long-range dependencies while reducing computational complexity. These two branches dynamically synergize through a gated cross-attention mechanism, facilitating interaction between local and global information. Furthermore, an innovative multi-level contrastive learning framework acts as a self-supervised regularizer, enforcing physical consistency between the dual representations and significantly enhancing the model’s generalization ability to unseen operating conditions and topologies. Validations on the Texas 2000-bus system demonstrate that DB-CoGNN exhibits significant performance and robustness advantages over classical GNN models.

The issue of ultra-low-frequency oscillations (ULFOs) presents a critical stability challenge in hydro-dominant power systems, arising primarily from small-signal frequency instability caused by the inherent water hammer effect of hydropower units. The increasing integration of large-scale renewables introduces fluctuating system operating modes, further complicating the task of ensuring frequency stability in power systems. Addressing this challenge, this paper derives distributed stability conditions for frequency small-signal stability using the Nyquist stability criterion. A distributed control design algorithm is proposed to guide the local design of governors in multi-machine systems, ensuring frequency stability even under changing operating conditions. This approach prevents frequency small-signal instability and improves the system’s dynamic response. Validations on a 14-hydropower-unit system and the real-world Yunnan power grid demonstrate the method’s effectiveness in suppressing ULFOs under disturbances and operational variability. The results also highlight the advantages of the distributed control approach in robustness, maintaining stability, and achieving faster frequency regulation compared to existing methods. The findings significantly contribute to the study of frequency small-signal stability in modern power systems integrated with hydropower and renewable energy sources, providing a practical framework for improving system performance under variable operating conditions and paving the way for more resilient and adaptive power grids. Note to Practitioners—This work is motivated by the ultra-low-frequency oscillations in hydropower-dominated power systems, which pose a serious threat to grid stability. The integration of large-scale renewable energy further complicates the issue by introducing highly variable and uncertain operating conditions. A representative case occurred in 2016, when ultra-low-frequency oscillations (0.053 Hz) were observed in the Yunnan power grid following the disconnection of AC lines between Yunnan and the main grid. Conventional measures such as AGC switching and output curtailment failed to resolve the issue. Subsequent studies attributed the oscillations to the combined effects of the water hammer phenomenon and improper PID parameters in hydropower governors. However, existing methods often struggle to adapt to changing operating conditions, require massive data processing, or lead to slow system responses, which limit their applicability in real-time and large-scale systems. This work addresses that gap by proposing a distributed control design for hydro governors, which enables local, plug-and-play optimization of each unit’s parameters while ensuring overall system frequency stability. The approach is grounded in a mathematical condition derived from the Nyquist stability criterion, enabling practitioners to verify whether a locally tuned controller preserves system-wide stability, regardless of which units are online. This distributed design not only reduces coordination and computational overhead but also improves frequency response and damping performance compared to existing methods. Practically, this means operators can adapt individual governor settings without relying on global reconfiguration, making the solution suitable for real-world systems with fluctuating operating modes. While validated on both a simulation test and the actual Yunnan provincial grid in China, the method assumes consistent system-level inertia and damping coefficients over a control cycle, which may need to be estimated or updated periodically in practice. Future work could extend this approach to wind- and solar-dominated grids or integrate adaptive inertia estimation to enhance robustness further.

The phase out of conventional synchronous generators (SGs) and the vigorous development of renewable energy sources (RESs) are indisputably leading to a significant reduction in the inertia of power grids, blowing a hole in frequency security and stability. To address this issue, various fast-acting resources such as battery energy systems (BESS) are being discussed worldwide. Accurate quantifying the relative effectiveness of these resources in arresting frequency decline to conventional methods is, therefore, of great significance to securely operate low-inertia power systems (LIPS). To do so, an analytical model of the equivalent frequency-containment performance ratio (EFCPR) is proposed for heterogeneous resources having different response characteristics. Furthermore, two Sigmoid-function-based approaches are also proposed to extend the EFCPR model to aggregated resources, and a more general scenario taking delivery time instant of instantaneous response BESS into account. Numerical results on the Texas test case and the Great Britain power grids with real operation data (from September 2023 to December 2024) collected from the National Energy System Operator (NESO) website validate the EFCPR model and penetrate many of the parameters' impacts.

Instruction tuning has become a widely adopted approach for aligning large multimodal models (LMMs) with human intent. It enables multi-task joint training through unified data formats. However, as new vision-language tasks constantly emerge, exhaustive joint training of all tasks becomes impractical. Continual learning offers a more flexible and resource-efficient alternative, enabling incremental training of LMMs on emerging tasks. This study investigates two fundamental questions when applying continual learning to instruction tuning of LMMs: 1) Do LMMs suffer from catastrophic forgetting during continual instruction tuning? 2) Can existing continual learning methods be effectively applied to continual instruction tuning of LMMs? A comprehensive study was conducted to answer these questions. First, we establish the first benchmark for continual instruction tuning of LMMs and reveal the phenomenon of catastrophic forgetting in this setup. Second, we integrate and adapt traditional continual learning approaches to this setting, demonstrating the effectiveness of these strategies to varying degrees in different scenarios. Third, we explore task-similarity dynamics between pairs of vision-language tasks and propose task-similarity-informed regularization and model expansion methods. Experimental results show that our approach can consistently boost the model's performance.

Jinghan He, Kuan Zhu, Haiyun Guo, Junfeng Fang, Zhenglin Hua, Yuheng Jia, Ming Tang, Tat-Seng Chua, Jinqiao Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Subsynchronous oscillations (SSOs) induced by the integration of high renewable energy penetration have significantly impacted the operation of power systems. This article proposed an enhanced matrix pencil method (MPM) for online monitoring of SSO using synchrophasors. To address the issue of eigenvalues of complex-domain matrix pencil not satisfying the conjugate frequency constraints of the synchrophasors, the complex Hankel matrix of synchrophasors is transferred to the field of real numbers. This improvement can shorten the data window of parameter identification to 200 ms while maintaining the computational efficiency of MPM. In addition, to ensure the identification accuracy by fully utilizing the information of complex-domain synchrophasors, the feasibility of constructing a new real-domain Hankel matrix with the combination of the separate real and imaginary parts of the complex Hankel matrix is proved. Compared with the existing MPM-based methods, the proposed enhanced MPM achieves better accuracy for parameter identification of SSOs while significantly reducing the computational burden in practical applications.

Continual learning (CL) is crucial for language models to dynamically adapt to the evolving real-world demands.To mitigate the catastrophic forgetting problem in CL, data replay has been proven a simple and effective strategy, and the subsequent data-replay-based distillation can further enhance the performance.However, existing methods fail to fully exploit the knowledge embedded in models from previous tasks, resulting in the need for a relatively large number of replay samples to achieve good results.In this work, we first explore and emphasize the importance of attention weights in knowledge retention, and then propose a SElective attEntion-guided Knowledge Retention method (SEEKR) for data-efficient replay-based continual learning of large language models (LLMs).Specifically, SEEKR performs attention distillation on the selected attention heads for finer-grained knowledge retention, where the proposed forgettabilitybased and task-sensitivity-based measures are used to identify the most valuable attention heads.Experimental results on two continual learning benchmarks for LLMs demonstrate the superiority of SEEKR over the existing methods on both performance and efficiency.Explicitly, SEEKR achieves comparable or even better performance with only 1/10 of the replayed data used by other methods, and reduces the proportion of replayed data to 1%.The code is available at https: //github.com/jinghan1he/SEEKR.

In this article, a novel, learning-based method for accurate location of faults in multiterminal direct current (MTdc) networks is proposed. By assessing the dc circuit breaker currents during the fault clearance process, a pattern recognition approach is adopted from which the fault location is estimated. The implementation of the algorithm is allocated into three main stages, where similarity coefficients and weighted averaging functions (incorporating exponential kernels) are utilized. For the proposed algorithm, only a short-time window of data (equal to 6 ms) is required. The performance of the proposed method is assessed through detailed transient simulation using verified MATLAB/Simulink models. Training patterns have been retrieved by applying a series of different faults within an MTdc network. Simulation and experimental results revealed that the proposed scheme, first, can reliably determine the type of fault, second, can accurately estimate the fault location (including the cases of highly resistive faults), and, third, is practically feasible.