最新刊期
- 摘要：Objective:The high penetration of renewable energy creates power systems with low inertia and weak damping, posing a major challenge to frequency stability after active power disturbances. Existing assessment methods are limited by low accuracy, long computation times, and a lack of training samples due to grid reconfigurations. To overcome these limitations, this paper proposes a data-driven method using multi-dimensional stability indicators. The approach enhances evaluation accuracy and efficiency, addresses data scarcity issues, and supports the secure operation of modern power systems. Method:The research adopted a multi-step systematic approach. Firstly, a simplified multi-machine frequency response model was established. Under different new energy penetration scenarios, the variation trends between unbalanced power and three key frequency parameters (maximum frequency change rate, transient frequency extreme value, quasi-steady-state frequency deviation) were analyzed. Combined with frequency stability criteria, a multi-dimensional power-frequency indicator evaluation system was constructed for judging frequency stability. Secondly, input features were preprocessed: CP−SMOTE technology was used to solve sample imbalance, and the LassoNet model was applied for feature selection to reduce data dimension. A BiGRU−KAN model was then constructed by enhancing the BiGRU module with the KAN attention mechanism for frequency stability evaluation. Accuracy, precision, recall, and F1−score from the classification confusion matrix were used to analyze the model's performance. Finally, an improved Wasserstein distance generative adversarial network (WGAN) with optimized gradient penalty mechanism was designed to generate sufficient similar samples for model training when system topology changes. The IEEE 10-machine 39-bus system was adopted as the test case for simulation verification under multiple scenarios. Results and Discussions:Comprehensive simulation experiments were performed on the IEEE 10-machine 39-bus system to evaluate the performance of the proposed method, yielding key results as follows. In sample preprocessing, a significant imbalance was observed between stable and unstable samples in the initial dataset. After treatment via CP−SMOTE technology, the proportion of stable samples increased from 37.5% to 52.5% of the total samples, essentially balancing it with unstable samples and laying a solid foundation for subsequent model training.For feature selection, the LassoNet model was employed. Analysis of its feature selection process revealed that the model attained optimal performance at an accuracy of 95.14% with 13 selected features, thus identifying the top 13 features with the highest importance scores for subsequent model input.In the classification performance evaluation of the BiGRU−KAN model, high accuracy was rapidly attained in the early training stage and sustained at a high level of 99.7% thereafter. This model significantly outperformed other comparative models (BiLSTM−KAN, GRU−KAN, KAN), demonstrating excellent classification accuracy.Regarding convergence performance, the cross-entropy loss of the BiGRU−KAN model declined sharply in the early training stage and approached zero with a small number of training epochs; specifically, it decreased rapidly and stabilized at approximately 5 training epochs. Compared with other models, it not only demonstrated distinct advantages in objective function optimization but also exhibited faster convergence speed, enabling the model to approach the optimal solution with fewer training epochs. Furthermore, the loss of BiGRU−KAN exhibited smaller fluctuations than that of other models throughout the entire training process and remained stable at a low level post-convergence.Comprehensive indicator analysis based on the classification confusion matrix indicated that BiGRU−KAN outperformed other models in classification performance, reflecting a higher correct prediction rate in classification tasks and a stronger ability to identify system stability or instability.In terms of training efficiency, the average time per training sample processing was compared across models: 5.13 ms for BiGRU−KAN, 6.48 ms for BiLSTM−KAN, 2.55 ms for GRU−KAN, and 2.01 ms for KAN. Although BiGRU−KAN required more time for single-sample processing than GRU−KAN and KAN, it was still more efficient than BiLSTM−KAN. Comprehensive consideration of accuracy, convergence, and speed revealed that while GRU−KAN and KAN had shorter processing times, they exhibited lower accuracy and inferior convergence. In contrast, leveraging its unique bidirectional gated recurrent unit structure, the BiGRU−KAN model could better capture the sequential dependencies in data, and its integration with the KAN attention mechanism further strengthened key feature extraction capabilities, achieving a better balance between accuracy and efficiency.For sample generation under topology changes, an improved Wasserstein distance generative adversarial network (WGAN) with an optimized gradient penalty mechanism was adopted. Comparative analysis of the time required to generate the same number of samples indicated that the unimproved WGAN took 9.29 s, while the improved version required 8.58 s. Combined with the analysis of discriminator and generator losses, the improved WGAN demonstrated significant advantages over the unimproved version in both sample generation quality and speed, facilitating faster and higher-quality sample generation.Furthermore, stability tests under noise interference revealed that although the introduction of noise vectors affected the stability during sample generation and resulted in higher losses compared with the noise-free condition, the discriminator could still accurately distinguish between original and generated samples under high-dimensional noise vectors, and the generator maintained high sample generation quality, thus verifying the robustness of the improved WGAN model. Conclusion:This study proposes a data-driven frequency stability assessment method incorporating multi-dimensional stability indicators, which effectively addresses the issues of low accuracy, high time consumption, and insufficient samples induced by topology changes in new power system frequency stability assessment. The multi-dimensional power-frequency indicator system, CP−SMOTE-based sample balancing, LassoNet-based feature selection, and BiGRU−KAN model synergistically enhance assessment accuracy and efficiency. The improved WGAN resolves the sample insufficiency caused by topology changes and strengthens the model’s generalization capability. Simulations on the IEEE 10-machine 39-bus system verify the method’s significant advantages in accuracy and speed over traditional methods, providing an important basis for accurate frequency stability assessment and rapid frequency modulation strategy formulation. Future research will extend this method to more complex power system models and explore its adaptability under extreme conditions to improve its engineering application value.关键词：new energy power system;frequency stable rapid assessment;BiGRU−KAN;LassoNet;gradient penalty mechanism;WGAN19|7|0更新时间：2026-06-25
- 摘要：Abstract: Significance: Under the background of the construction of the new-type power system, massive data owners have multi-source heterogeneous data resources. However, owing to the concerns of data assets and privacy security, there is often a problem of “data silo” among data owners. In the meantime, as significantly important data, which is closely related to national security and economic development, electric power data is a crucial production factor, which supports the development and construction of the new-type power system. Hence, as a new type of distributed machine learning technology, federated learning can effectively help multiple data owners to carry out data sharing and machine learning modeling under the premise of satisfying privacy protection, data security and government regulations. As a result, it is of great significance for artificial intelligence modeling and data value mining of relevant business scenarios of the new-type power system. Progress: In this context, this paper conducted a systematic review of the status quo of applications of federated learning in the new-type power system during recent years, existing problems and challenges, as well as future research directions. Firstly, the basic concept, classification, and applicable scenarios of federated learning were briefly summarized. According to the distribution of feature and sample spaces of data among different participants, federated learning can generally be divided into horizontal federated learning, vertical federated learning, and federated transfer learning. Horizontal federated learning is applicable to scenarios where the datasets of all participants have the same feature space but the different sample spaces; vertical federated learning is applicable to scenarios where the datasets of all participants have the different feature spaces but the same sample space; federated transfer learning is applicable to scenarios where there is relatively little overlap in both the feature spaces and the sample spaces of all participants. Then, aiming at the problem of secure data sharing faced by data owners under the background of the new-type power system, the status quo of applications of federated learning in each business scenario of the new-type power system was described in detail. To be specific, the application scenarios of horizontal federated learning mainly include renewable energy generation and load forecasting, analysis of electricity consumption behaviors of customers, network security protection of power systems, dispatching and optimization of power networks, and security and stability assessment of power networks; the application scenarios of vertical federated learning mainly include operation and control of power distribution networks/microgrids, and vehicle-to-grid interaction and electric vehicles; the application scenarios of federated transfer learning mainly include load forecasting, fault diagnosis of power equipment, regulation ability assessment of buildings, and power safety protection equipment detection. Currently, the application scenarios of federated learning involve all the basic links of the new-type power system. Also, through the systematic analysis of the status quo of applications, it can be found out that as multiple subjects have emerged in the new-type power system, when selecting the technical type of federated learning, it is necessary to comprehensively consider relevant aspects such as data types of different subjects, computing resources, privacy protection intensity, and scenario requirements, thereby proposing different federated learning solutions. This process is a highly customized application method of federated learning in specific application scenarios of the new-type power system. Furthermore, due to the typical characteristics of the new-type power system, including high real-time requirements for application scenarios, high security requirements for data privacy, as well as high personalization requirements for business models, the current problems and shortcomings of application of federated learning in the new-type power system were deeply analyzed. Specifically, in terms of communication, as communication capability is the key guarantee for the practical and widespread application of federated learning in the new-type power system, it is faced with challenges such as heterogeneous computing resources, large model parameter scales, and a huge number of clients, thus resulting in the decrease of its communication efficiency; in terms of privacy and security, although federated learning employs various encryption methods in the application of the new-type power system to protect data security and privacy, it faces issues such as information leakage and malicious attacks, thereby affecting its credibility and reliability; in terms of modeling, as a large number of various participants of federated learning tend to own non-independent and identically distributed data and traditional federated learning focuses on obtaining a global shared model to achieve the global optimum of the whole training process, it has problems such as the poor model training effect and the insufficient model generalization ability, therefore leading to the decrease of its model performance and the difficulty in meeting the personalized needs of different participants. Conclusions and Prospects: Finally, in view of the challenges and issues faced by federated learning, three future research directions of federated learning were prospected, including hierarchical federated learning based on model compression and asynchronous update mechanisms, federated learning based on blockchain and personalized differential privacy, as well as personalized federated learning based on client selection, in order to provide the constructive reference for data value mining of the new-type power system under the goal of “double carbon”. Although federated learning has great potential in promoting the digital transformation of the new-type power system, its future implementation and deployment still requires effective solutions to practical problems such as the stability of the electric power communication network, the establishment of the universal electric power federated learning platform, and the technical standards and norms of electric power federated learning, in order to achieve its widespread practical application.关键词：new-type power system;data silo;federated learning;privacy protection;data value mining11|7|0更新时间：2026-06-25
- 摘要：Objective: In the context of new power system development, wind power forecasting faces dual challenges of accuracy and generalization capability. The inherent volatility and intermittency of wind power output, along with other complex uncertainties, are difficult to effectively characterize through traditional modeling approaches. Meanwhile, when introducing large language models (LLMs) into the field of wind power forecasting, significant adaptability bottlenecks also exist. To address these issues, a short term wind power forecasting method based on pre-trained large language models was proposed to enhance the accuracy of short-term wind power forecast in this study. Furthermore, systematic attribution of model prediction results was conducted by using the SHAP (SHapley Additive exPlanations) algorithm, aiming to improve the credibility and practicality of the forecasting model in power system operation and decision support. Methods: First, mainstream LLMs architectures were analyzed, and the superiority of Decoder-Only architecture models was demonstrated through mathematical derivation to determine the base model framework. Subsequently, the basic neural network structure of the Qwen series large language models and their input-output transformation processes were examined in depth, laying the foundation for the subsequent formation of a targeted wind power forecasting architecture. In this process, any input representation is transformed into word vectors through an embedding layer. These word vectors then undergo transformation via multiple stacked Decoder layers with residual connections. Each Decoder consists of RMSNorm (Root Mean Square Layer Normalization), an attention module, and a Multi-Layer Perceptron (MLP) module. Differing from traditional attention computation, the attention module of Qwen incorporates Rotary Position Embedding (RoPE) to express sequential relationships among tokens. Finally, the transformed intermediate states are mapped through a linear layer to obtain the final output representation. Second, a wind power forecasting architecture based on pre-trained large language models was proposed for short-term wind power time series forecasting tasks. Given the significant differences between wind power numerical sequences and linguistic sequences, the original input representation layer of Qwen required redesign. In this architecture, the input representation layer was replaced with a fully connected layer to project time series data into the required dimensionality of the specific pre-trained model. Additionally, Reversible Instance Normalization (RevIN) was employed to independently normalize each variable channel, preserving the integrity of variable features. Partitioning the sequence into sub-blocks via patching enhanced local temporal information. Following the concept of feature transfer, the attention and MLP modules within the Decoder layers of the pre-trained Qwen model were retained, with their parameters frozen during training as these modules store the primary pre-trained knowledge. Lastly, the output representation layer was redesigned to project hidden states into the data dimensionality required for the wind power forecasting task. This architecture achieves deterministic point forecasting for day-ahead wind power output through a joint frozen-fine-tuning training strategy. Third, to address the poor interpretability of deep learning models, a feature contribution analysis method based on SHAP (SHapley Additive exPlanations) values was adopted. Derived from game theory’s Shapley value, the SHAP method quantifies the marginal contribution of each input feature by considering all possible feature subsets, satisfying axioms of local accuracy, feature symmetry, and zero impact. The SHAP value of each feature was calculated via Monte Carlo sampling, and the method was applied to quantify the impact of key meteorological features on wind power forecasting, revealing the underlying contribution mechanisms. Results and Discussions: Two years of data from two wind farms (Oregon and Nevada) provided by the National Renewable Energy Laboratory (NREL) were used for validation, with an 80%–20% train-test split and a raw time resolution of 5 minutes. Autoformer, Dlinear, and PatchTST were selected as benchmarks, and the proposed method was implemented using the Qwen2-0.5B model. Experiments on the number of retained Decoder layers showed that the model achieved optimal performance with 6 Decoders (Qwen(6)), as increasing layers beyond 8 led to overfitting. In terms of training efficiency, Qwen(6) outperformed benchmarks with a training time of 0.147 seconds per step and convergence in 10 epochs—surpassing PatchTST (0.552s/step, 15 epochs) and Autoformer (0.271s/step, 12 epochs). In 24, 96, and 144 time steps forecasting tasks, Qwen(6) achieved the lowest MSE across both wind farms. For example, at 24 time steps, its MSE prediction error was lowered over 11.53% compared to benchmarks. In few-shot experiments using only 25% of the training data, Qwen(6) exhibited superior generalization: its MSE increased by 15.9%, while PatchTST’s error rose by 25.8%. Comparison of pre-fine-tuning and post-fine-tuning performance showed that the untuned Qwen model had an MSE over three times higher than the fine-tuned version, verifying the effectiveness of the proposed architecture in addressing LLM’s prediction limitations. SHAP analysis revealed that the 100m wind speed feature had the strongest impact on predictions, with an average absolute SHAP value of 5.1129 and a contribution ratio exceeding 78%. 100m wind direction also exerted a positive effect, while 40m wind parameters and temperature had negligible impacts, providing a quantitative basis for feature selection and model interpretation. Conclusions: This study explored the application potential of LLMs in time series forecasting and proposed a short-term wind power forecasting method integrated with interpretability analysis. The key innovations are threefold: Firstly, the superiority of the Decoder-Only architecture is verified through mathematical derivation and selecting the Qwen model as the base model; 2) A frozen-fine-tuning strategy that freezes attention and MLP modules while fine-tuning input/output layers and RMSNorm was designed, balancing training efficiency and prediction accuracy; 3) The SHAP was integrated to quantify meteorological feature contributions, enhancing model transparency. Experimental results confirmed that the proposed method outperformed benchmarks in training efficiency, prediction accuracy, and generalization capability, effectively capturing long- and short-term temporal dependencies of wind power output. However, limitations remain, such as insufficient stability in extreme weather and difficulty in calculating multi-step dynamic SHAP contributions. Future work will focus on multi-modal data fusion and lightweight model adaptation to further improve engineering practicality.关键词：large language model;wind power forecasting;interpretability;deep learning16|4|0更新时间：2026-06-25
- 摘要：Objective:Distribution Network Maintenance (DNM) decision-making currently relies heavily on manual scheduling and experiential judgment, which suffers from low data integration efficiency, delayed information processing, and subjective bias. General-purpose large language models (LLMs), despite their strong semantic capabilities, exhibit insufficient domain-specific knowledge and frequent hallucination when applied to specialized maintenance scenarios. Existing retrieval-augmented generation (RAG) approaches, which rely on flat text chunk retrieval, fail to capture the complex hierarchical structures and causal relationships embedded in power maintenance documents, such as equipment fault trees, operational procedure dependencies, and cross-referenced industry standards. This work addresses these limitations by proposing a framework that integrates knowledge graph (KG) retrieval-augmented generation with LLMs for DNM decision support, termed KRAG−LLM−DM. Methods: The KRAG−LLM−DM framework comprises several interconnected stages. Maintenance-related texts, including research literature, industry standards, and operational procedures in PDF format, are first converted to UTF-8 encoded plain text. Documents are segmented into text chunks with configurable length thresholds to facilitate subsequent vectorization and retrieval. For KG construction, an LLM-guided prompt-based extraction mechanism is designed. The prompt template incorporates few-shot learning, chain-of-thought reasoning, parameterized entity type placeholders, descriptive anchoring for entity disambiguation, and structured delimiters to enforce output consistency. The LLM automatically identifies maintenance entities such as transformers and circuit breakers, along with their semantic relationships including causal and operational dependencies, extracting both explicit and implicit knowledge elements from the segmented texts. Relationship strength is quantified using an implicit scoring scheme to enable fine-grained relational modeling. Following entity and relation extraction, the KG is partitioned into communities using the Leiden algorithm, which optimizes modularity through iterative node movement, partition refinement, and network aggregation. For each community, a hierarchical summary report is generated by the LLM using a structured prompt that requires critical entity identification, compliance assessment, technical capability evaluation, impact scoring with explanatory justification, and source citation tracking. The generated community summaries and associated text chunks are vectorized using one of three embedding models (Jina, Anything LLM, or Nomic) and stored in a vector database. During inference, a user query is encoded into the same embedding space, and cosine similarity retrieval identifies the most relevant text chunks and community summaries. The retrieved knowledge fragments are recombined with the original query to form an augmented context, which is then fed into the LLM for response generation. Results and Discussions: The framework was evaluated through multiple experiments addressing knowledge base construction, ablation studies, embedding model comparisons, and decision performance benchmarking. A maintenance knowledge base was constructed from a corpus of power distribution maintenance documents. Chunk recall testing with a query on transformer maintenance procedures demonstrated that retrieved text fragments primarily exhibited similarity scores between 0.50 and 0.52. In the first case study, the LLM decision support framework without RAG was deployed on both online cloud platforms and local offline environments. Temperature coefficient analysis revealed that lower coefficients produced more structured and logically coherent responses suitable for high-reliability maintenance contexts. Comparisons among mainstream LLMs on two maintenance planning queries showed that the proposed framework achieved recall rates of 0.719 and 0.694, representing improvements of 2.36%~13.40% over GPT-4, DeepSeek-V3, GLM-4, and Kimi. The RAG-enhanced responses explicitly referenced national standards and provided hierarchically structured answers with source attribution, whereas baseline models often omitted domain-specific content such as maintenance grading criteria and dynamic constraint conditions. In referential ambiguity scenarios, the framework attained an F1-score of 0.625, marginally outperforming competing models which failed to explicitly resolve anaphoric references. The second case study examined the influence of embedding models on the RAG-based decision framework. Among Jina, Anything LLM, and Nomic embeddings, Jina demonstrated superior semantic capture of nested hierarchical structures in long power maintenance texts due to its multi-task contrastive learning and decoupled attention mechanism. Nomic embeddings exhibited broader semantic representation and supported dynamic dimensionality pruning, whereas Anything LLM showed information loss when handling compound semantic constructs. The third case study evaluated the KRAG−LLM−DM framework. The KG constructed using the proposed LLM-guided extraction method contained 18 280 nodes and 14 002 edges, with an average degree of 0.21, modularity index of 0.953, and network diameter of 9. Graph visualization was implemented using Gephi with ForceAtlas2 layout optimization. Comparative analysis against keyword matching, regular expression, and SpaCy-based dependency parsing methods revealed that the proposed approach achieved greater knowledge coverage and structural completeness while operating automatically without manual rule maintenance. Ablation experiments on three complex maintenance question types demonstrated F1-scores of 0.726 3, 0.708 1, and 0.729 0 for the KRAG−LLM−DM framework, outperforming both the LLM without RAG by 1.60%~2.63% and the standard RAG-augmented LLM by 1.80%~3.60%. Error analysis indicated that the KG structural constraints and hierarchical community semantics enhanced answer consistency and reduced logical fragmentation compared to flat RAG retrieval. In comparisons with parameter-matched baseline models including Mistral, Olmo, and LLaMA, the KRAG−LLM−DM framework achieved F1-score improvements of 3%~5% on average across device state identification, fault cause tracing, and maintenance recommendation tasks. An interactive decision support interface featuring knowledge provenance visualization, structured answer presentation, and multi-tab version switching was developed to facilitate human-machine collaborative decision-making. Conclusions: The KRAG-LLM-DM framework integrates KG construction with retrieval-augmented generation to address the deficiencies of general-purpose LLMs in DNM decision support. The LLM-guided prompt-based extraction mechanism enables automated construction of structured maintenance KGs from heterogeneous unstructured texts, achieving node and edge scales of 18,280 and 14,002 respectively without manual annotation. Hierarchical community summarization and vector-based retrieval provide semantically organized knowledge augmentation, yielding recall rates of 0.719 and 0.694 on maintenance regulation queries and grading-based planning tasks, with improvements ranging from 2.36%~13.40% over mainstream LLM baselines. The framework demonstrates superior performance compared to both non-RAG and conventional RAG configurations across complex professional queries, validating the contribution of structured semantic constraints and multi-hop reasoning paths to enhanced answer professionalism and interpretability. Despite these advances, the coupling mechanism between LLM reasoning and graph-structured inference paths remains partially opaque, limiting the transparency of causal chains from evidence nodes to final maintenance decisions. Future investigations should focus on causal reasoning within LLM−KG integration for safety-critical power system applications.关键词：large language model;retrieval-augmented generation;knowledge graph;distribution network;maintenance decision;artificial intelligence;prompt engineering23|9|0更新时间：2026-06-25
- 摘要：Objective The increasing structural complexity and uncertain operating conditions in modern power grids have made the identification of critical nodes an essential task for preventing cascading failures and ensuring system resilience. Traditional topology-based or model-driven approaches rely heavily on static assumptions, precise physical models, and extensive prior knowledge, which limits their applicability under dynamic and nonlinear power flow environments. Existing reinforcement-learning-based identification techniques also exhibit shortcomings, including insufficient representation of disturbance propagation, Q-value overestimation, and limited capability in capturing nonlinear interactions among nodes. To address these issues, this study aims to (i) enhance the state representation of power-grid nodes by incorporating disturbance diffusion characteristics, (ii) improve the stability and accuracy of reinforcement-learning-based node evaluation, and (iii) provide a robust framework for identifying nodes whose perturbation induces significant structural and operational degradation in the network. Methods A reinforcement-learning framework based on an improved Diffusion Double DQN is developed. The critical-node identification problem is formulated as a Markov Decision Process, where each state represents a multi-attribute description of a node, including electrical, topological, and operational metrics. Actions correspond to selecting a candidate node for evaluation, and the reward function integrates indicators reflecting voltage deviation, load sensitivity, power flow redistribution, and disturbance propagation influence. A diffusion mechanism is introduced to model the spread of perturbations across the network topology. Starting from a node, disturbance information is iteratively diffused through the adjacency structure to produce a multi-step neighborhood representation. The final diffusion state encodes both direct and indirect interactions, which enhances the agent’s perception of local vulnerability, load redistribution potential, and cascading risk. A multi-head attention (MHA) module is embedded into the feature-extraction stage to capture nonlinear dependencies among node features. This mechanism allows the model to assign adaptive importance to different node attributes and neighbor interactions, strengthening the representation of structurally subtle but operationally influential nodes. To improve training stability, an advantage-constrained update strategy is incorporated into the Double-DQN framework. This mechanism reduces bias in Q-value estimation, mitigates instability during policy updates, and increases learning efficiency in high-dimensional state spaces. The overall system integrates diffusion-enhanced state expansion, attention-based feature modeling, and stabilized Q-learning into a unified model for evaluating and ranking node importance. Results and Discussions Simulation studies were performed on the IEEE 30-bus and IEEE 118-bus systems using network efficiency, largest connected component, voltage stability index (VSI), and load-shedding amount (LSA) as evaluation metrics. In the IEEE 30 system, removing the top 10 nodes ranked by each method shows that all approaches lead to a pronounced decline in network efficiency and largest connected component. Nodes selected by the proposed Diffusion-DDQN cause consistently larger reductions—particularly in the first several removals—indicating that the method tends to identify nodes whose perturbation has stronger structural and operational impact. Because the 30-bus system is relatively small and different algorithms produce partially overlapping top-10 node sets, the curves converge after removing all 10 nodes. For the IEEE 118 system, nodes were removed based on percentage thresholds due to the larger network size. When the top 10% of nodes identified by each method were removed, the Diffusion-DDQN produced the most significant degradation, reducing network efficiency to 32.8% and the largest connected component to 0.44 of their original values. Relative to conventional DDQN, the network-efficiency decline was 3.5% greater, while the largest connected component decline exceeded that of the closeness-based method by 0.06. When 20% of the nodes were removed, network efficiency and largest connected component further dropped to 14.6% and 0.15, respectively; both reductions exceeded those under the improved PageRank baseline by 4.7% and 0.24, indicating stronger consistency with disturbance-impact simulations in larger systems. Further evaluations were conducted by applying controlled perturbations (5% voltage disturbance and 10% load disturbance) to the top-ranked nodes determined by each method. For VSI, nodes selected by the Diffusion-DDQN produced the largest voltage-stability deviations, with changes approximately 10% higher than those obtained from conventional DDQN. In terms of LSA, when perturbing the top 5% of nodes, both Diffusion-DDQN and DDQN produced load-shedding effects about 60% higher than the improved PageRank method. When extending the perturbation to 20% of nodes, the Diffusion-DDQN yielded the highest LSA among all evaluated methods. Across all metrics, the diffusion-enhanced state representation enabled the agent to detect nodes that trigger greater cascading stress, producing ranking results more consistent with physical disturbance simulations. Conclusions This study presents an improved Diffusion-DDQN framework for critical node identification in complex power grids. By incorporating a diffusion mechanism into the state space, the method captures essential spatial propagation characteristics that are overlooked by conventional RL and graph-theoretic approaches. The integration of a multi-head attention module improves feature-interaction modeling, while the advantage-constrained estimation stabilizes Q-value updates and enhances learning efficiency. Simulation results on IEEE-30 and IEEE-118 demonstrate that the proposed method provides more accurate, robust, and interpretable assessments of node importance. The framework offers a practical and extensible tool for analyzing vulnerability, supporting protection planning, and enhancing resilience assessment in modern power systems.关键词：deep double Q-network;diffusion model;multi-head attention mechanism;critical node identification13|4|0更新时间：2026-06-25
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0