Power transformers are critical assets in power systems, and their reliable operation is essential for grid stability. Conventional fault diagnosis methods suffer from delayed response and limited adaptability, while existing artificial intelligence-based approaches face challenges related to data heterogeneity, limited interpretability, and weak integration of physical mechanisms. To address these issues, this paper proposes a physics-informed enhanced transformer-based framework for power transformer fault diagnosis. A unified temporal representation scheme is developed to integrate heterogeneous monitoring data using Dynamic Time Warping and physics-guided feature projection. Physical priors derived from thermodynamic laws and gas diffusion principles are embedded into the attention mechanism through multi-physics coupling constraints, improving physical consistency and interpretability. In addition, a multi-task diagnostic strategy is adopted to jointly perform fault classification, severity assessment, and fault localization. Experiments on 3000 samples from 76 power transformers demonstrate that the proposed method achieves high diagnostic accuracy and superior robustness under noise and interference, indicating its effectiveness for practical predictive maintenance applications.

With breakthroughs in the power electronics industry, the stability and rapid power regulation of wind power generation have been improved. Its power generation technology is becoming more and more mature. However, there are still weaknesses in the operation and control of power systems under the influence of extreme weather events, especially in real-time power dispatch. To optimally distribute the power of the regulation resources in a more stable manner, a wind energy forecasting-based power dispatch model with time-control intervals optimization is proposed. In this model, the outage of the wind energy under extreme weather is analyzed by an autoregressive integrated moving average model (ARIMA). Additionally, the other regulation resources are used to balance the corresponding wind power drop and power mismatch. Meanwhile, an algorithm names weighted mean of vectors (INFO) is employed to solve the real-time power dispatch and minimize the power deviation between the power command and real output. Lastly, the performance of the proposed optimal real-time power dispatch is executed in a simulation model with ten regulation resources. The simulation tests show that the combination of ARIMA and INFO can effectively improve the power control performance of the PD-WEF system.

Retrieval-augmented generation (RAG) has shown strong potential for knowledge-intensive tasks, yet its performance degrades sharply when applied to structured long-context documents in power systems, where dense entity–relation dependencies, cross-document references, and strict traceability requirements exist. To address this Structured Long-Context RAG (SLCRAG) challenge, this paper proposes a hierarchical graph-enhanced RAG (HG-RAG) framework tailored for power system question answering. HG-RAG constructs a globally consistent knowledge graph via sliding-window entity–relation extraction to mitigate semantic fragmentation, and employs multi-granularity structured indexing for precise entity/relation retrieval. A hierarchical structured retrieval mechanism with multi-hop expansion and semantic distillation maximizes recall while minimizing redundancy. Furthermore, a regex-enhanced retrieval module records authoritative file_path provenance and constrains downstream retrieval to the same source documents, effectively eliminating cross-document interference—especially in cases where different documents contain similar entities and relations. Combined with version control and deduplication-merging, HG-RAG supports incremental knowledge updates with minimal forgetting and negligible token overhead. Experimental results on a domain-authentic power system QA dataset demonstrate that HG-RAG outperforms LightRAG and GraphRAG, achieving up to 85.47% accuracy in short-answer tasks with significantly lower token consumption. Ablation studies confirm that semantic distillation primarily improves precision and efficiency, while regex-enhanced retrieval safeguards recall in edge cases.

With the widespread integration of renewable energy sources and distributed energy storage systems, intraday dispatch of distribution networks (DNs) has gradually evolved into a multistage sequential decision-making problem with intertemporal state coupling and progressively revealed uncertainty. Stochastic optimization under the stagewise independence assumption and limited horizon prediction cannot adequately capture temporal transition characteristics, which may lead to biased future cost estimation and myopic decisions. To address this obstacle, this paper proposes a conditional-cut-enhanced Markov chain stochastic dual dynamic programming strategy (MC-SDDP-CC) for multistage intraday dispatch of DNs. Feature encoding and scenario-lattice-driven trajectory sampling with Markovian path dependence are introduced to accurately characterize the temporal dependence in DNs. To improve computational efficiency, a node-wise conditional cut management scheme is developed to accelerate the recursion and tighten the conditional value function approximation. Numerical studies on modified IEEE 33-bus and 123-bus networks, as well as a practical system in a southwestern province of China, verify the effectiveness of the proposed method in terms of optimality, scalability and ablation performance.

Reliable state assessment of railway traction power transformers is challenged by heterogeneous operating environments, measurement disturbances, coupled gas-generation mechanisms, and uneven fault-sample distributions. Conventional dissolved gas analysis (DGA) ratio rules and single-model classifiers often show insufficient generalization when rare faults and boundary-ambiguous operating states are encountered. To address this issue, this paper proposes a feature-guided stacking framework for state assessment of oil-immersed railway power transformers. First, a DGA-oriented fusion-feature representation is established by combining raw gas concentrations, gas-ratio descriptors, and an aggregated dissolved-gas analysis factor. Second, DBSCAN-assisted sample structuring is introduced to identify density patterns, sparse rare fault regions, and boundary samples, thereby improving the organization of imbalanced monitoring records. Third, a monitoring-feature-embedded stacking model is developed in which heterogeneous base learners are adaptively weighted according to feature-reliability information and integrated through a cross-validated meta-learner. This synthetic-data-based validation provides a controlled and reproducible proof-of-concept. Therefore, the reported results should be interpreted as evidence of methodological feasibility. Under the default synthetic setting, the proposed feature-guided stacking (FE-stacking) method achieves an accuracy of 99.70% and a macro-F1 of 99.55%. Under the severe minority-retention setting in which only 25% of low-energy discharge (LD) and low-temperature overheating (LT) training samples are preserved, it obtains an accuracy of 99.62%, a macro-F1 of 99.40%, and an LT recall of 96.61%, slightly surpassing random forest (RF) and outperforming Original Stacking in rare fault robustness. These results indicate that feature-guided ensemble learning can improve the generalization stability of DGA-based transformer state assessment under imbalanced and boundary-ambiguous conditions. From a practical perspective, the proposed framework can serve as a decision-support module for transformer condition screening, maintenance prioritization, and alarm verification in railway traction power-supply systems.