Abstract：In response to challenges in power-grid emergency supply assurance such as dispersed information, heterogeneous states, missing knowledge, and tightly coupled decision-making, this paper proposes an integrated hydrogen knowledge graph and collaborative decision-making framework for grid emergency supply assurance. Centered on the main line of “using hydrogen as the means and grid supply assurance as the objective”, the framework constructs a three-domain coupled knowledge representation architecture covering the hydrogen side, the grid side, and the dispatch side. At the knowledge completion layer, a hybrid completion strategy is developed by integrating R-GCN with symbolic rule reasoning. At the state perception layer, time-series sensor streams, equipment drawings, and operational data are fused to build a dynamically updatable multimodal knowledge graph. At the evolution layer, an entropy-based uncertainty active learning mechanism is introduced to support continuous knowledge iteration. Experimental results show that the proposed method improves NER F1 to 92.0%, RE F1 to 88.9%, and cross-sentence RE F1 to 65.4%, while achieving 58.2% Hits@10 in knowledge completion. It also outperforms multiple baseline methods in critical-load restoration rate, restoration time, and decision latency. The proposed framework provides a technical pathway that jointly supports knowledge representation, reasoning, and collaborative decision-making for hydrogen-enabled grid resilience enhancement and emergency power supply dispatch.

Abstract：Difficult airway intubation is a critical risk in clinical anesthesia. This study proposes an automated, objective airway assessment framework named MMAAF (Multi-Modal Airway Assessment Framework) based on static mouth-opening images of patients. Using computer vision techniques, the framework automatically extracts six key anatomical features from the dataset of patient mouth-opening images. These features are combined with actual difficult intubation labels to construct a structured multi-modal prediction dataset, which is processed with a random oversampling technique. The study employs two integrated machine learning algorithms—random forest and gradient boosting decision trees—for model training and evaluation. Results show that the gradient boosting and random forest models achieve prediction accuracies of 92% and 93%, respectively, on an independent test set, demonstrating good assessment performance. Concurrently, a logistic regression model is trained within the MMAAF framework. Features are weighted according to their importance weights, and a personalized probability score for difficult intubation is calculated for each patient. This provides an intuitive quantitative reference based on multi-modal features to support clinical decision-making.

Abstract：In the context of computing-power and photovoltaic (PV) coordination, to address the issue that existing scenario generation methods struggle to capture the coordinated patterns of PV output response to computing load, a two-stage scenario generation method combining mixed-integer programming (MIP) and conditional diffusion models is proposed. First, the coupling mechanism between computing load and PV output is analyzed, and a MIP-based computing task scheduling model is constructed to generate joint PV-computing load scenarios. Second, using the MIP-generated scenarios as training data, a conditional diffusion model with U-Net as the backbone network is constructed, encoding temporal features, temperature, computing task levels, and electricity prices as conditional vectors to learn the high-dimensional joint distribution of PV-computing load, achieving controllable scenario generation. Results indicate that the proposed method can generate high-fidelity joint scenarios that provide reliable data support for distribution network planning.