A multi-site tide level prediction model based on graph convolutional recurrent networks. (1st February 2023)

• Record Type: Journal Article
• Authors: Zhang, Xinlong; Wang, Tengfei; Wang, Weiping; Shen, Ping; Cai, Zhongya; Cai, Huayang
• Is Part Of: Ocean engineering. Volume 269(2023)
• Publication Date: 2023-02-01
• Subjects: Multi-site tide -- Tide level prediction -- Machine learning -- Graph convolutional recurrent networks; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.113579
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Predicting regional tide levels is vital for engineering and catastrophe avoidance along the shore. Data-driven method is capable of fast prediction of tide levels. However, current data-driven algorithms only make predictions for individual tide stations, rather than aiming to a regional network system of tide stations. This paper proposes a model based on graph convolutional recurrent networks to predict tidal levels at regional multiple tide stations. The model captures spatial and temporal features from historical tide level and meteorological data. Future tidal levels for multiple tide stations are the model's output. In this work, 48-year historical data from five tidal stations in Pearl River Delta were utilized for model training and evaluation. The results show that: (1) The model outperforms five commonly used baseline models in terms of evaluation metrics RMSE and MAE, and is able to predict future tide levels at multiple tide stations; (2) Short-term forecasts (1 and 3 h) are more accurate than long-term forecasts (12 h); (3) The model retains a high degree of accuracy for short-term predictions and satisfactory accuracy for long-term prediction during typhoons. The method provides a new instrument for regional prediction of tide levels and forecasting of storm surges. Highlights: A regional tide level prediction model based on graph convolutional recurrent networks is proposed. The model captures feature information from spatial and temporal tide levels and meteorological factors at nearby tide stations. Historical data from five tidal stations in Pearl River Delta were utilized for model training and evaluation. The model outperforms five commonly used baseline models in terms of evaluation metrics. The model retains satisfactory accuracy for both short-term (1–3 h forecast lead time) and long-term prediction (12 h) during typhoons.