Transforming existing grid-following (GFL) doubly-fed induction generator (DFIG) wind farms into grid-forming (GFM) DFIG ones is essential for enhancing system frequency support. Studying frequency response and parameters tuning in hybrid wind farms with both unit types establishes a foundation for frequency control. Frequency response modeling for GFL units is well established, small-signal models of GFM DFIG units fail to capture frequency-active power dynamics due to multi-variable coupling, hindering unified hybrid wind farm modeling. Furthermore, frequency control parameters tuning relies on inefficient trial-and-error methods. These defects limit the application of hybrid wind farms. This paper proposes a comprehensive frequency response model for hybrid wind farms and an efficient parameter tuning method. By analyzing the frequency support process in GFM DFIG turbines, a frequency-active power response model is derived. Considering system inertia sources, wind farm outputs are transformed and aggregated to build a system frequency response model that includes hybrid DFIG wind farms. A multi-objective function based on this model integrates the maximum frequency variation rate, deviation, and steady-state deviation. The improved genetic particle swarm optimization (GPSO) algorithm is applied for parameters tuning under generator output and stability constraints. Simulation results demonstrate that the proposed model accurately captures frequency and turbine output dynamics, and the tuning method improves system frequency response and stability.
Retrieval-Augmented Generation enhances large language models by incorporating external knowledge, but deploying it in sensitive scenarios risks privacy leakage via malicious prompts. To address this, we propose a multi-agent framework that sanitizes retrieved content through semantic rewriting. By employing three specialized agents for privacy extraction, semantic analysis, and reconstruction, our approach collaboratively removes sensitive identifiers while preserving the semantic core. We evaluate the framework on the ChatDoctor and Wiki-PII datasets across six large language models. Experimental results demonstrate a significant reduction in privacy leakage under targeted attacks. For instance, we reduced targeted information exposure in LLaMA-3-8B from 144 instances in the baseline to just 1. Furthermore, we maintain strong contextual fidelity with a BLEU-1 score of 0.122, outperforming the existing SAGE method's 0.117. Finally, the framework operates as an asynchronous preprocessing module, introducing no additional latency to online inference, as all rewriting is executed as a one-time offline preprocessing step. To promote reproducibility, the source code of this work is publicly available at https://github.com/foursoils/Privacy-Preserving-RAG.
The integration of high-penetration distributed photovoltaics intensifies the spatiotemporal imbalance between sources and loads in distribution networks, making it difficult to achieve safe and economic operation under radial topology constraints. Traditional optimization methods also suffer from limited solution efficiency and scheduling performance under uncertain source-load conditions. To address these issues, this paper proposes a deep reinforcement learning (DRL)-based spatiotemporal cooperative optimal scheduling method for an energy router (ER). First, a spatiotemporal cooperative scheduling framework for the ER is established, in which spatial power mutual support is realized through multi-port power transfer, while temporal energy shifting is achieved by coordinating an internal energy storage system (ESS). Second, an optimization model is formulated with the objectives of minimizing voltage violation deviation and network losses, and the scheduling process of the ER and its internal ESS is modeled as a Markov decision process. Furthermore, a soft actor-critic algorithm based on feature grouping and self-attention, namely FGA-SAC, is proposed to enhance the policy network’s ability to identify key state features. Finally, the proposed method is validated on the IEEE 33-bus system, IEEE 69-bus system, and an extreme load-drop scenario. The results show that the proposed method can effectively ensure voltage security and reduce network losses. Compared with SAC, TD3, DDPG, and traditional optimization methods, it also demonstrates better stability and online decision-making efficiency.
High-voltage switchgear is a critical component in power systems, and its operating condition is directly related to grid security and power supply reliability. With the continuous growth of operational and maintenance (O&M) data and the dynamic evolution of operating conditions, conventional O&M models struggle to rapidly integrate newly acquired knowledge while maintaining decision-making performance. To address these issues, this paper proposes a dynamic incremental updating-based method for constructing an O&M knowledge graph for high-voltage switchgear. First, an improved OneRel model is employed to jointly extract entities and relations from O&M texts to construct the O&M knowledge graph. Second, an anchor updating mechanism is introduced to select and match key nodes in newly added O&M knowledge, enabling rapid embedding of multi-source O&M information. Then, a weighted distillation-based incremental learning method is adopted to continuously update the O&M model, thereby improving its adaptability to complex operating conditions while retaining historical knowledge. Compared with traditional methods, the proposed method reduces the embedding time of newly added O&M information by up to 74.35%, while the average expert evaluation score is improved by 30%. The proposed method can provide technical support for the intelligent operation and maintenance of high-voltage switchgear.
The high penetration of renewable energy and large-scale EV integration poses new challenges to the optimal scheduling of regional integrated energy systems (RIESs) by intensifying source-load uncertainty and affecting operational flexibility and economic performance. To address this issue, a bi-level optimal dispatch model is developed to coordinate the economic interaction between the RIES operator and EV users through price-based source-load interaction. The upper-level RIES operator minimizes its net operating cost, while the lower-level EV users minimize their net charging/discharging cost. Three EV participation modes, namely disorderly charging, orderly charging, and vehicle-to-grid (V2G), are compared to quantify their impacts on load smoothing, multi-energy coordination, and system-user economics. Scenario generation and reduction are used to construct expected renewable profiles, while Monte Carlo sampling provides heterogeneous EV travel inputs for the deterministic day-ahead bi-level dispatch. Under the zero-degradation-cost setting, the V2G strategy reduces the RIES operator net cost, EV user net cost, and combined resource cost by 5.56%, 47.74%, and 10.99%, respectively, compared with disorderly charging. When battery degradation compensation is considered, the EV-side economic benefit decreases but remains positive; for example, the EV user cost reduction decreases to 42.41% at a degradation cost of 0.10 CNY/kWh. The proposed strategy balances the interests of EV users and the RIES operator while improving load smoothing and flexible coordination of multi-energy flows.
