Quick Entry

  • Latest Articles
  • Highly cited papers
  • Online First
  • Archive
更多
Volume 25 期 9，2026 2026年第25卷第9期

  • Topic\: Scenario\-driven and Industrial Implementation of Big Data and Artificial Intelligence

    LI Minglei, LIU Wenping

    DOI：10.11907/rjdk.261376
    摘要：The traditional network information evolution model heavily relies on mathematical equations， lacks semantic processing capabilities， and struggles to effectively depict the cognitive heterogeneity and dynamic decision-making of individuals when faced with distorted information. To address this， a distorted information evolution simulation system based on large language model multi-agent simulation is designed. Firstly， the system constructs a directed graph topology using real interaction datasets and generates multi-dimensional heterogeneous user personas. Secondly， it designs the SocialAgent class， which utilizes discrete time steps for asynchronous serial scheduling of the "perception-inference-decision" process. Simultaneously， relying on thought chains and prompt word engineering， it endows agents with text comprehension and logical game-playing abilities. Experimental results demonstrate that the system can accurately simulate the evolution process of distorted information， with a fitting accuracy （root mean square error） improvement of 31.5% compared to the traditional SIR model. Intervention deduction experiments confirm the importance of early intervention， showing that timely intervention at the third time step reduces the absolute value of the infection scale by 36.7%. The system can accurately simulate the evolution process of distorted information in social networks， providing an effective deduction tool for the governance of such information.  
    关键词：large language model;multi-agent simulation;disinformation;evolutionary modeling;social networks   
    76
    |
    9
    |
    0
    <HTML>
    <L-PDF><WORD><Enhanced-PDF><Meta-XML>
    <引用本文> <批量引用> 172795135 false
    更新时间：2026-09-28

    DONG Mengran, LI Shutong, MOU Jiangtao

    DOI：10.11907/rjdk.261195
    摘要：With the continuous evolution of artificial intelligence and electricity markets， accurate electricity price forecasting and transaction risk identification are critical for decision-makers and market participants in the power industry. As a cutting-edge AI technology， deep learning （DL） demonstrates substantial potential in addressing complex and time-sensitive electricity price forecasting challenges. An intelligent evaluation model for the coordinated operation of green power and green certificates based on multi-source regulatory data has been designed to enhance the intelligence level of market supervision. Firstly， the model integrates Light Gradient Boosting Machine （LightGBM） with Independently Recurrent Neural Network （IndRNN） to achieve day-ahead price prediction； Subsequently， a Federated Learning （FL） anomaly detection model based on information entropy is utilized to identify non-compliant behaviors in electricity markets； Finally， a Multi-Layer Perceptron （MLP） neural network is implemented for risk early warning， providing a reference for the assessment and early warning of electricity market operations. Simulation results demonstrate that the proposed model effectively enhances electricity price forecasting accuracy and accurately identifies anomalous trading behaviors， offering regulators dynamic and intelligent supervision tools.  
    关键词：multi-source regulatory data;green electricity and green certificates;coordinated operation;intelligent evaluation model;artificial intelligence   
    42
    |
    4
    |
    0
    <HTML>
    <L-PDF><WORD><Enhanced-PDF><Meta-XML>
    <引用本文> <批量引用> 172795291 false
    更新时间：2026-09-28

    HUANG Yunsong, GONG Jinnan, LUO Cheng, DUAN Mingjun, XU Kaihang

    DOI：10.11907/rjdk.261286
    摘要：Dynamic Facial Expression Recognition （DFER） holds significant value in scenarios such as psychological assessment and medical auxiliary diagnosis. Existing video-based expression recognition methods primarily rely on images for training， indirectly modeling the visual representation of facial muscle movements （Action Units， AUs） in pixel space， which are susceptible to interference from non-expression factors like identity appearance， lighting， and background. In contrast， explicitly characterizing facial local muscle movements through AUs can more directly and interpretable reveal the correspondence between specific actions and emotions， mitigating the interference of aforementioned pseudo-relevant information to some extent. To this end， an extended long short-term memory network （MA-X） based on AUs is proposed for DFER. First， standard AU sequences are used as the sole input， and multi-scale temporal convolution blocks model AUs across different time spans. Second， a dual-branch recalibration adapter dynamically adapts to AU combination channels and temporal frames， adjusting AU sequence features to highlight key AU combinations and their instantaneous temporal variations. Finally， an extended long short-term memory network （xLSTM） encodes global dependencies to complete expression classification. Experiments on the MEAD dataset demonstrate that MA-X significantly outperforms multiple baselines in accuracy while preserving interpretable outputs， providing a structured design approach for interpretable AU sequence-based DFER.  
    关键词：action unit;dynamic facial expression recognition;multi-scale temporal convolution;xLSTM model   
    53
    |
    6
    |
    0
    <HTML>
    <L-PDF><WORD><Enhanced-PDF><Meta-XML>
    <引用本文> <批量引用> 172795287 false
    更新时间：2026-09-30

    LAI Rongxuan, ZOU Li, WANG Zhenyi, LYU Xianwei

    DOI：10.11907/rjdk.261332
    摘要：With the rapid development of generative artificial intelligence technology， big language models have shown great potential in policy Q&A applications in the field of military human resource management. At present， the research on large-scale model evaluation of military human resources policy Q&A scenarios is still in its infancy， lacking systematic evaluation system construction methods and practical guidance. To this end， a military human resources policy Q&A evaluation system covering five dimensions including policy semantic understanding ability， policy knowledge comprehensive application ability， answer accuracy， answer consistency， and security compliance was constructed around the application evaluation problem of the military human resources big model for policy Q&A. A method for generating and evolving evaluation data based on multi strategy prompt engineering was proposed. Firstly， the leap from static generation to dynamic evolution of evaluation data has been achieved； Secondly， a military human resources policy Q&A evaluation dataset covering multiple policy areas， question types， and difficulty levels has been developed； Once again， a human centered evaluation data review and traceability mechanism has been established， effectively improving the coverage and quality of evaluation data； Finally， in order to achieve targeted iterative upgrades of the tested application， a complete reporting framework was designed that includes evaluation background， evaluation methods， evaluation results， result analysis， root cause analysis， and disposal plans. This provides a reliable attribution path for evaluation and strongly supports further iterative upgrades of the tested application. The experiment shows that the proposed method provides theoretical support and practical guidance for the scientific evaluation of large-scale model applications in the field of military human resources， and is of great significance for promoting the standardized application of generative artificial intelligence technology in military human resources management.  
    关键词：large language models;military human resources;policy question answering;evaluation system;evaluation data construction;generative artificial intelligence   
    23
    |
    3
    |
    0
    <HTML>
    <L-PDF><WORD><Enhanced-PDF><Meta-XML>
    <引用本文> <批量引用> 172795268 false
    更新时间：2026-09-28
查看更多
更多
0