An Hourly Short-Term Sea Fog Forecasting Model for Ningbo-Zhoushan Port Based on Dynamic Graph Neural Networks

Sea fog is a critical meteorological factor affecting production safety and operational efficiency at Ningbo-Zhoushan Port. To enhance the fine-grained forecasting capabilities for sea fog in the port area, this study exploits the strengths of graph neural networks (GNNs) in handling irregularly distributed meteorological station data. Based on station locations and wind direction/speed, a dynamic graph was constructed by incorporating the advection term in atmospheric motion equations. A novel forecasting framework was developed, integrating Graph Attention Networks (GAT), Long Short-Term Memory (LSTM), and multi-step classification output, to train a sea fog forecasting model. The proposed model delivers hourly rolling 24-hour sea fog forecasts for 72 stations within the port area. Operational validation in 2023 demonstrates robust performance in forecasting fog occurrence. For 24-hour fog occurrence forecasting, the model achieves a TS score of 0.369 (F1=0.539), with an average TS score of 0.190 for fixed-time forecasts. The forecasting performance shows characteristic temporal decay, achieving optimal TS score (0.441) at 1-hour lead time. Furthermore, the model exhibits superior capability in predicting localized patchy fog and fog evolution, outperforming conventional machine learning approaches including Random Forest and Gaussian Naive Bayes methods. 海雾是影响宁波舟山港生产安全和作业效率的重要气象因素。为提升港区海雾精细化预报能力，本研究利用图神经网络在处理不规则分布气象站点数据方面的优势，基于站点位置和风向风速信息，结合大气运动方程中的平流项构建动态图结构。搭建了融合图注意力神经网络、长短期记忆网络结构与多步长分类输出的网络框架，训练得到海雾预报模型。该模型可为港区72个站点提供逐小时滚动更新的24小时海雾预报。2023年业务检验表明，该模型在24小时内雾的出现预报中TS评分达到0.369（F1=0.539），定时预报平均TS评分为0.190。预报性能随时效延长呈规律性下降，其中1小时预报TS评分最优（0.441）。此外，模型在团雾、雾区范围和雾演变的预报上表现更优，同时也优于随机森林和高斯朴素贝叶斯等传统机器学习方法。