Machine Learning-Based Flood Forecasting System for Window Cliffs State Natural Area, Tennessee

Abstract

The prevalence of unforeseen floods has heightened the need for more accurate flood simulation and forecasting models. Even though forecast stations are expanding across the United States, the coverage is usually limited to major rivers and urban areas. Most rural and sub-urban areas, including recreational areas such as the Window Cliffs State Natural Area, do not have such forecast stations and as such, are prone to the dire effects of unforeseen flooding. In this study, four machine learning model architectures were developed based on the long short-term memory, random forest, and support vector regression techniques to forecast water depths at the Window Cliffs State Natural Area, located within the Cane Creek watershed in Putnam County, Tennessee. Historic upstream and downstream water levels and absolute pressure were used to forecast the future water levels downstream of the Cane Creek watershed. The models were tested with lead times of 3, 4, 5, and 6 h, revealing that the model performances reduced with an increase in lead time. Even though the models yielded low errors of 0.063–0.368 ft MAE, there was an apparent delay in predicting the peak water depths. However, including rainfall data in the forecast showed a promising improvement in the models’ performance. Tests conducted on the Cumberland River in Tennessee showed a promising improvement in model performance when trained with larger data.

Keywords

• flood forecasting
• machine learning
• flood early warning system (FEWS)
• long short-term memory (LSTM)
• random forest (RF)
• support vector regression (SVR)