Sea fog is a critical meteorological phenomenon that affects production safety and operational efficiency at Ningbo–Zhoushan Port. To improve fine-grained sea fog forecasting in the port area, this study constructs a dynamic graph by leveraging the strengths of graph neural networks in processing irregularly distributed meteorological station data. The graph is designed based on station locations and wind direction/speed, incorporating the advection term in the atmospheric motion equation. On this basis, a novel forecasting framework is developed that integrates graph attention networks, long short-term memory, and multistep classification output, to train a sea fog forecasting model. The proposed model provides hourly rolling 24-h sea fog forecasts for 72 stations across the port area. Operational validation in 2023 demonstrates its robust performance in forecasting fog occurrence. For 24-h fog occurrence forecasts, the model achieves a threat score (TS) of 0.369 (F1 = 0.539), with an average TS of 0.190 for fixed-time forecasts. The forecasting performance exhibits a typical temporal decay pattern, reaching its highest TS (0.441) at the 1-h lead time. Furthermore, the model demonstrates superior skill in predicting localized patchy fog and fog evolution, outperforming conventional machine learning approaches such as random forest and Gaussian naive Bayes.

This study established an S-band dual-polarization radar quantitative precipitation estimation system (NDRQPES) using raindrop size distribution (DSD) data observed during the Meiyu seasons of 2021–2023 in Ningbo, a coastal city in eastern China. The NDRQPES was evaluated by comparing its quantitative precipitation estimation (QPE) accuracy with the national radar mosaic operational product (CHN_QPE) during the Meiyu periods of 2023 and 2024. The results show that NDRQPES, incorporating linear rainfall relationships from localized DSD data, has significantly improved hourly precipitation estimate accuracy, with a correlation coefficient (CC) > 0.90 and relative bias (RB) within ±10%. Furthermore, it outperforms CHN_QPE in critical success index (CSI) across various rainfall rates. However, it exhibits biases in accumulated rainfall estimates, overestimating by 8.89% in 2023 and underestimating by 13.87% in 2024. NDRQPES is also less effective than CHN_QPE in detecting extreme heavy rainfall (> 64 mm h⁻¹). These deficiencies may be attributed to the microphysical complexity of Meiyu rainfall, including variabi-lity in hydrometeor morphology and noise interference in weak rainfall, as well as the limited representativeness of single-point DSD data. While CHN_QPE effectively captures intense precipitation events, it exhibits significant systematic overestimation, with an RB of 161.77% in 2024. This study provides valuable insights for refining and localizing dual-polarization radar precipitation algorithms for regional applications.