Clustering-Based Global Forecasting Models for Significant Wave Height Prediction

Abstract

Accurate wave predictions safeguard maritime operations, coastal communities, and marine ecosystems. Significant wave height, an average height of the highest one-third of the waves recorded during the sampling period, plays a crucial role in analyzing wave conditions and assessing coastal hazards among various wave fields. Forecasting significant wave height for various future timeframes, starting from 0.5 hours ahead, is vital for estimating coastal storm surges, issuing weather warnings, and preventing coastal disasters, especially during imminent large waves. Numerical methods are commonly used for wave forecasting; however, due to their computational intensity, they often require more time. In emergency situations, data-driven models offer faster wave predictions while maintaining accuracy, for shorter timeframes into the future. Data-driven forecasting models often treat data reported by buoys individually and forecast significant wave height based on the historical data of the respective buoy. Models trained on data from multiple buoys might leverage combined insights. However, training a single model on all different buoys may reduce forecasting accuracy when the data is from buoys in different environments. This study proposes a two-step approach to improve significant wave height predictions on a set of Environment and Climate Change Canada (ECCC) buoy data. First, we cluster buoys with similar data, enabling the formation of clusters with similar environmental conditions. Second, we train a global forecasting model on each cluster and predict significant wave height for individual buoys. We evaluate our proposed approach for significant wave height forecasting using data collected by 28 ECCC buoys distributed across the Atlantic, Pacific, and Great Lakes regions of Canada. Our results demonstrate that the clustering-based forecasting models, which leverage the shared patterns and relationships among multiple related buoy data, show competitive performance compared to the data-driven models trained on individual buoy data or universal model trained on all buoy data, in extreme events where wave height exceeds 6 meters.