【作者】 张雪梅；

【导师】 单世民；

【作者基本信息】 大连理工大学 ， 软件工程， 2025， 硕士

【摘要】 在全球贸易中,海运承担了约90%的货物运输和约80%的价值总额。其中,集装箱班轮运输作为海上运输主要形式,承担了全球70%以上的海运贸易价值总额,对全球经济贸易发展尤为重要。近年来,全球集装箱海运网络港口拥堵问题仍未缓解,多起重大港口拥堵事件频发,对全球经济、环境和社会影响深远。因此预测港口拥堵状态是全球集装箱海运网络中重要的研究问题,对于世界海运供应链稳定具有重要的现实意义。本研究提出了时空深度学习模型框架(Spatio-Temporal Deep Learning,ST-DL),以应对相关挑战。研究以船舶自动识别系统数据为基础,构建全球海运网络,并设计融合长短期记忆网络与Transformer架构的时空深度学习模型。该模型结合双解码器架构与交叉注意力机制,实现了多港口场景下未来多时间步的拥堵状态预测。模型量化港口平均待泊时长为港口拥堵状态指标,引入距离图和交互图刻画港口关系,结合多粒度时间嵌入强化时序感知。实验结果表明,ST-DL在均方根误差(RMSE)和平均绝对误差(MAE)指标上显著优于传统LSTM和XGBoost方法,并且在拥堵状态高波动场景下的预测优势进一步凸显。此外,噪声扰动以及参数敏感性分析进一步分析了模型的鲁棒性。本研究有效融合复杂网络理论与时空预测模型,为港口运营、供应链管理及绿色航运发展提供技术支撑,助力全球海运网络数字化转型与可持续发展。更多还原

【Abstract】 In global trade,maritime transport handles approximately 90%of goods transportation and around 70%of the total value.Among these,container liner shipping—being the primary form of maritime transport—accounts for over 70%of the global maritime trade value and is particularly important for the development of global economic trade.In recent years,port congestion issues within the global container shipping network have not eased,with multiple major congestion incidents occurring frequently and having far-reaching impacts on the global economy,environment,and society.Therefore,predicting port congestion status is an important research problem in the global container shipping network and holds significant practical implications for the stability of the world’s maritime supply chain.This study proposes a Spatio-Temporal Deep Learning(ST-DL)model to address these challenges.The research is based on Automatic Identification System data to construct a global maritime network and designs a spatio-temporal deep learning model that integrates Long Short-Term Memory networks with the Transformer architecture.This model,combining a dual decoder architecture with a cross-attention mechanism,achieves multi-step predictions of future congestion states in multi-port scenarios.The model measures the average berthing waiting time as an indicator of port congestion,introduces distance graphs and interaction graphs to characterize port relationships,and reinforces temporal awareness with multi-granularity time embeddings.Experimental results indicate that ST-DL significantly outperforms traditional LSTM and XGBoost methods in terms of root mean square error(RMSE)and mean absolute error(MAE),with its predictive advantages further highlighted in scenarios exhibiting high volatility in congestion status.Furthermore,noise perturbation and parameter sensitivity further analyze the robustness of the model.This study effectively integrates complex network theory with spatio-temporal forecasting models,providing technical support for port operations,supply chain management,and the development of green shipping,thereby contributing to the digital transformation and sustainable development of the global maritime network.更多还原