Machine Learning Based Water Level Predictions for Southern Florida

Along the coast of Florida there is an increased risk of inundation and compound flooding in large part due to relative sea level rise. As a result, it has become imperative that local stakeholders and residents have access to better water level predictions in their area. This work focuses on the development, performance, and the operationalization of regression Multilayer Perceptron (MLP) AI models capable of predicting short-term water levels (12 - 96 hours) for Virginia Key, Biscayne Bay, Florida (FL).

The study location, Virginia Key, Biscayne Bay, FL, was selected to take advantage of the sensitive location near Miami and the historical and operational measurements of the NOAA Tides & Currents station (ID of 8723214). Data variables were retrieved from the tide station: 6-minute verified water level measurements and hourly observed winds. The 6–minute water levels were converted to an hourly interval (meaning only water levels at the hour were kept) to reduce computational costs. These datasets were taken from 1998-2023. These datasets were used to select the set of inputs that were to be fed to the machine learning models. Based on the observed winds (wind direction and wind speed) the u-wind (N-S) and v-wind (E-W) wind components were calculated to be used as model inputs. Before model calibration the data went through various preprocessing steps to make the dataset AI ready. There were no missing values in the water level dataset, but there were data gaps in the observed winds time series. To handle these missing values, any gaps that were greater than four hours were dropped from the dataset while the smaller remaining gaps (less than or equal to four hours) were interpolated linearly.

Water level predictions are computed hourly using 5 ML models in total, each with a different lead time: 12, 24, 48, 72, and 96 hours. Each model uses past (from the forecast time) water level measurements, past u- and v-wind measurements, and future u- and v-wind "predictions" as inputs. Predictions are in quotations because for model calibration and assessment the u-wind and v-wind measurements were treated as predictions using the perfect prognosis approach. For the operational model, real-time predictions are extracted from the National Weather Service National Digital Forecast Database. Each model implemented the Multilayer Perceptron architecture. Each model used 50 hidden neurons (with the exception of the 48-hour model which found the best performance using 20 hidden neurons), one hidden layer, Adam Optimizer, ReLU activation function, and Mean Squared Error (MSE) as the loss function. These hyperparameters were lightly tuned using RandomSearch from the KerasTuner Python library. Early stopping (50 epochs), model checkpoint (saves the model with lowest validation loss), and dropout (2% dropout rate) were also used.

Almost every single model (except the 96-hour model, which came very close at 89.2%) met the NOAA operational central frequency CF (15cm) requirement of greater than or equal to 90% for the predictions to be within 15 cm of the eventual measurements. The models performed well based on other metrics as well, such as MSE, Root Mean Square Error (RMSE), R Squared (R2), Mean Absolute Error (MAE), and Median Absolute Error (MedAE). Each water level model has been transitioned to real-time operations using the student-developed software platform called Semaphore which semi-automates the operationalization of AI models. The predictions generated by these models can be viewed in real-time through an interactive chart located on this webpage: https://sherlock-prod.tamucc.edu/cbocp/virginia-key.html.