Toward Forecasting Groundwater Table in Flood Prone Coastal Cities Using
Long Short-term Memory and Recurrent Neural Networks
Abstract
Coastal cities face recurrent flooding from storm events and rising
seas. A contributing factor to flooding in these low relief areas is the
groundwater table, which, already relatively shallow, can quickly rise
towards the land surface during storm events. This leads to increased
surface runoff entering stormwater drainage systems and a greater
probability of flooding. As such, groundwater table forecasts could be
an important component of real-time flood forecasting systems, but are
generally unavailable. Because traditional physics-based models require
extensive amounts of subsurface data that is difficult to obtain,
especially in urban environments, this research evaluates two types of
machine learning models, Recurrent Neural Networks (RNN) and Long
Short-term Memory neural networks (LSTM), for creating groundwater table
forecasts. The two types of networks were built with Tensorflow/Keras to
forecast the groundwater table response to forecasted storm events and
appropriate hyperparameters were tuned using the Hyperas library. Using
observed hourly groundwater levels, rainfall, and tide from the City of
Norfolk, Virginia, the networks were trained with data from 2010-2016
and tested with data from 2016-2018. Archived forecast rainfall and tide
from two large storms in the test period (Hurricane Hermine and Tropical
Storm Julia) were then used to evaluate the effect of forecast inputs on
model performance. Results indicate that LSTM is slightly more accurate
when forecasting the groundwater table than RNN, likely because of its
increased ability to preserve and learn from past information. Average
root mean squared error and Nash-Sutcliffe efficiency values for an 18hr
forecast for the LSTM were 0.06m and 0.89, respectively, and 0.07m and
0.85, respectively, for the RNN. These forecasts could provide valuable
information to aid in planning and response to storm events and will
become an increasingly important part of effectively modeling and
predicting coastal urban flooding as sea level rises.