Predicting the Evolution of Extreme Water Levels with Long Short-Term
Memory Station-based Approximated Models and Transfer Learning
Techniques
Abstract
Extreme water levels (EWLs) resulting from cyclones pose significant
flood hazards and risks to coastal communities and interconnected
ecosystems. To date, physically-based models have enabled accurate
prediction of EWLs despite their inherent high computational cost.
However, the applicability of these models is limited to data-rich sites
with diverse characteristics. The dependence on high quality
spatiotemporal data, which is often computationally expensive, hinders
the applicability of these models to regions of either limited or
data-scarce conditions. To address this challenge, we present a Long
Short-Term Memory (LSTM) network framework to predict the evolution of
EWLs beyond site-specific training stations. The framework, named
LSTM-Station Approximated Models (LSTM-SAM), consists of a collection of
bidirectional LSTM models enhanced with a custom attention mechanism
layer embedded in the architecture. LSTM-SAM incorporates a transfer
learning approach applicable to target (tide-gage) stations along the
U.S. Atlantic Coast. Importantly, LSTM-SAM helps analyze: (i) the
underlying limitations associated with transfer learning, (ii) evaluate
EWL predictions beyond training domains, and (iii) capture the evolution
of EWL caused by tropical and extratropical cyclones. The framework
demonstrates satisfactory performance with “transferable” models
achieving Kling-Gupta Efficiency (KGE), Nash-Sutcliffe Efficiency (NSE),
and Root-Mean Square Error (RMSE) ranging from 0.78 to 0.92, 0.90 to
0.97, and 0.09 m to 0.18 m at the target stations, respectively. We show
that LSTM-SAM can accurately predict not only EWLs but also their
evolution over time, i.e., onset, peak, and dissipation, which could
assist in operational flood forecasting in regions with limited
resources to set up physically-based models.