随着车联网的发展和智能交通系统的完善,移动边缘计算和边缘缓存技术不断发展,在车联网服务迁移问题上发挥了巨大作用。为了更加合理利用边缘缓存资源、减小时延,本文提出了一种联合优化时延和缓存策略的“车–边–云”协同服务迁移方案。在该架构下,网络中的车载终端、路侧单元和云服务器可以协同进行任务计算,进行动态服务迁移。通过使用基于Hawkes过程的方法,根据历史请求信息更新不同内容类型的流行度,选择合适的边缘计算节点和缓存内容。本文以最大化系统的长期效益为目标,降低时延和提高命中率,形式化服务迁移问题。这是一个具有挑战性的非凸问题。通过将该问题建模为马尔科夫决策过程,并引入深度强化学习求解最优问题,提出了一种结合长短期记忆网络和近端策略优化算法的服务迁移算法。仿真实验结果表明,提出的策略优化方法比其他策略有更好的性能。With the development of the Internet of Vehicles (IoV) and the improvement of intelligent transportation system, mobile edge computing (MEC) and edge cache technology are constantly developing, which play a huge role in the service migration of the Internet of vehicles. In order to make more reasonable use of edge cache resources and reduce the delay, this paper proposes a “Vehicle-Edge-Cloud” collaborative service migration scheme which jointly optimizes the delay and cache strategy. In this architecture, vehicles, Roadside Units (RSUs) and cloud server in the network can collaborate on task computation and dynamic service migration. By using a Hawkes process-based approach, the popularity of different content types is updated based on historical request information, and appropriate edge compute nodes and cached content are selected. To maximize the long-term benefits of the system, reduce the delay and improve the hit rate, formalize the problem of service migration. This is a challengi
监控和预测PM2.5浓度变化对人类健康和环境污染治理至关重要。本文旨在研究PM2.5浓度长期预测任务中精度较低的问题。通过融合空间特征提取、空间注意力机制增强以及长时间序列特征提取,提出了一种预测模型,能够精准捕捉长序列中PM2.5浓度变化趋势。该模型首先通过CNN提取空间特征,并利用空间注意力机制强化关键空间信息。然后,由XLSTM捕捉时间序列中的动态变化和长期依赖关系。本章模型在两个大城市的数据集上进行了实验,并与FXX、LSTM、XLSTM以及CNN-XLSTM进行了对比分析。结果表明,本文模型在所有评估指标上均优于对比模型,充分验证了其有效性和泛化能力。Monitoring and predicting changes in PM2.5 concentration is crucial for human health and environmental pollution control. This paper aims to investigate the issue of low accuracy in long-term PM2.5 concentration prediction tasks. By integrating spatial feature extraction, spatial attention mechanism enhancement, and long-term sequence feature extraction, a predictive model is proposed that is capable of accurately capturing the trends of PM2.5 concentration variations over extended sequences. Specifically, the model first extracts spatial features using CNN and enhances key spatial information through a spatial attention mechanism. Subsequently, XLSTM captures dynamic changes and long-term dependencies within the time series. The model is evaluated on datasets from two major cities. The results show that the proposed model outperforms comparison models, including FNN, LSTM, XLSTM, and CNN-XLSTM, across all evaluation metrics, fully validating its effectiveness and generalization capability.
