A Machine Learning Framework for Extending Wave Height Time Series Using Historical Wind Records

TL;DR

This paper introduces a machine learning framework using ConvLSTM-1D to extend wave height time series from historical wind data, demonstrated on Lake Michigan for a 70-year hindcast, showing improved accuracy with more wind stations.

Contribution

The study develops a novel ML framework that effectively models wave heights from wind observations, including an ensemble approach and optimal input station number, for long-term wave prediction.

Findings

Single wind station provides reasonable wave height estimates.

Adding more stations improves accuracy, plateauing after four stations.

Optimal lookback period identified as 10 hours.

Abstract

This study presents a novel machine learning-based (ML) framework that utilizes the ConvLSTM-1D model to hindcast or forecast wave heights at coastal locations using a nonuniform array of wind observations. This approach was applied to Lake Michigan to perform a 70-year ice-free hindcast of waves near Chicago, IL (USA). The Wave Information System model (WIS) served as the training, validation, and testing dataset for the ML model. Ensemble learning-optimized ML models forced by different numbers of observation stations were tested, showing that a single wind station alone as an input feature produced a reasonably accurate wave height model. However, the wave height model accuracy increased as more wind input data was included from around the lake, largely plateauing beyond the inclusion of four stations that spanned Lake Michigan's southern basin. The optimized model lookback period was found to be 10 hours for all models, suggestive of a fetch-limited temporal coupling between the wind observations and nearshore waves. The ML framework offers a promising avenue for utilizing historical wind records worldwide to extend wave height time series for nearshore locations, particularly in enclosed and semi-enclosed basins where waves are strongly linked to local winds.