Scholarlink

<p>Sea fog frequently affects Ningbo–Zhoushan Port, posing substantial risks to port operations and safety. To enhance the forecasting capabilities for sea fog in main waterways of the port, this study proposes a forecasting framework based on residual network (ResNet) combined with long short-term memory network (LSTM). The framework utilizes observational data from Liuhengdong Station (K9504) and surrounding automatic weather stations. The input fields comprise single-point time series and multi-channel grid data with varying spatiotemporal resolutions. To address class imbalance, the Focal Loss function is employed. The trained model achieves hourly multi-step sea fog forecasts, providing 24-h rolling forecasts at 1-h resolution with reliable operational performance. Validation results indicate the system achieves measurable improvements in fog prediction. For 24-h fog occurrence forecasts, the model achieves an accuracy of 0.706 and F1 score of 0.714. The system maintains strong performance in the critical 0–12-h window (F1 = 0.736), though exhibits expected gradual performance decline in the 12–24-h range (F1 = 0.581). This performance pattern follows the characteristic temporal decay, showing decreasing TS from 0.562 (1&#xa0;h) to 0.203 (24&#xa0;h) as forecast lead time increases. Comparative analysis with conventional machine learning methods, including support vector machines, random forests, and Gaussian Naive Bayes, confirms the significant performance advantages of our proposed framework across all verification metrics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A single-station hourly multi-step forecasting model for sea fog based on ResNet and LSTM networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月1日

查看全文
  • Junbo Sun,
  • Yanzhen Qian,
  • Yue Sun,
  • Huayu Shen,
  • Changyu Zhao,
  • Sihui Fan

摘要

Sea fog frequently affects Ningbo–Zhoushan Port, posing substantial risks to port operations and safety. To enhance the forecasting capabilities for sea fog in main waterways of the port, this study proposes a forecasting framework based on residual network (ResNet) combined with long short-term memory network (LSTM). The framework utilizes observational data from Liuhengdong Station (K9504) and surrounding automatic weather stations. The input fields comprise single-point time series and multi-channel grid data with varying spatiotemporal resolutions. To address class imbalance, the Focal Loss function is employed. The trained model achieves hourly multi-step sea fog forecasts, providing 24-h rolling forecasts at 1-h resolution with reliable operational performance. Validation results indicate the system achieves measurable improvements in fog prediction. For 24-h fog occurrence forecasts, the model achieves an accuracy of 0.706 and F1 score of 0.714. The system maintains strong performance in the critical 0–12-h window (F1 = 0.736), though exhibits expected gradual performance decline in the 12–24-h range (F1 = 0.581). This performance pattern follows the characteristic temporal decay, showing decreasing TS from 0.562 (1 h) to 0.203 (24 h) as forecast lead time increases. Comparative analysis with conventional machine learning methods, including support vector machines, random forests, and Gaussian Naive Bayes, confirms the significant performance advantages of our proposed framework across all verification metrics.