Session: 08-02-01 Wave–Structure Interaction, Noise Modeling, and Marine Hydrodynamics

Submission Number: 179599

Hybrid Timegan Model for Enhanced Coastal Sea Level Forecasting 

Climate change is causing sea levels to rise, posing increasing threats to coastal communities and underscoring the urgent need for accurate regional sea level predictions. However, existing deep learning models for sea level forecasting—such as Convolutional Long Short-Term Memory (ConvLSTM) and Convolutional Gated Recurrent Unit (ConvGRU) networks—are constrained by limited training data and struggle to capture long-term temporal dependencies. To address these challenges, we propose a hybrid forecasting approach that first uses a Time-series Generative Adversarial Network (TimeGAN) to generate realistic synthetic sea level data, augmenting the limited historical dataset. We then introduce Mamba, a new linear-time sequence model designed to efficiently capture long-term dependencies, and train it on the enriched data. In addition, we apply Seasonal-Trend decomposition using Loess (STL) to remove periodic seasonal effects from the data, allowing the model to focus on underlying long-term trends and variability.
Experimental results on two representative coastal regions, New York and Lisbon, demonstrate that our method significantly outperforms ConvLSTM and ConvGRU baselines, yielding substantially lower prediction errors. In New York, the hybrid TimeGAN-Mamba model reduces the Average Mean Squared Error (AMSE) by 58.9% compared to ConvLSTM and by 46.3% compared to ConvGRU. Similarly in Lisbon, we observe AMSE reductions of 29.9% and 14.7% relative to ConvLSTM and ConvGRU, respectively. This consistent pattern of improved performance confirms the efficacy of our solution. Moreover, these improvements highlight the potential of combining synthetic data generation with advanced sequence models as a promising direction for enhancing sea level forecast accuracy, especially in coastal regions with limited historical data.

Presenting Author: Guanchao Tong Wenzhou-Kean University

Presenting Author Biography: Dr. Guanchao Tong is an Assistant Professor in the Mathematical Science Department at Kean University at Wenzhou-Kean University. He received his Ph.D. in Applied Mathematics and Statistics from the State University of New York at Stony Brook in 2023. He has extensive research experience in data analysis, data mining, and time series analysis and forecasting. He also conducts interdisciplinary research on climate change, and medical data analysis, aiming to apply advanced statistical and machine learning methods to address global sustainability and public health challenges. The work reflects an interdisciplinary approach aimed at tackling complex problems in the environmental and health sciences by combining applied mathematics, data-driven methods, and deep learning techniques.

Authors:

Guanchao Tong Wenzhou-Kean University