Application of Offshore Visibility Forecast Based on Temporal Convolutional Network and Transfer Learning

Area of Science:

• Meteorology
• Artificial Intelligence
• Data Science

Background:

• Offshore visibility forecasting is challenging due to sparse observational data and complex weather patterns.
• Existing methods struggle with data scarcity, impacting prediction accuracy.

Purpose of the Study:

• To propose an intelligent offshore visibility prediction method using temporal convolutional network (TCN) and transfer learning (TL).
• To address the limitations of low observational data in offshore forecasting.

Main Methods:

• Data preprocessing for source and target visibility datasets.
• Development of a TCN-based model integrated with transfer learning (TCN_TL).
• Knowledge transfer from a large source domain dataset to a small target domain dataset.

Main Results:

• The TCN_TL model demonstrated significantly lower forecast errors and improved forecast scores (e.g., +0.11 within 0-1 km, 24h forecast period) compared to pre-transfer learning models.
• TCN_TL outperformed CUACE forecast results, showing smaller errors and a TS score improvement of 0.16.
• The study confirmed that transfer learning enhances model prediction performance in small-data scenarios.

Conclusions:

• The proposed TCN_TL method effectively improves offshore visibility forecasting accuracy, particularly when dealing with limited data.
• TCN_TL offers superior performance compared to other deep learning methods and conventional forecasting models like CUACE.
• This approach provides a robust solution for enhancing visibility predictions in challenging offshore environments.