Predicting Daily Stream Flow with Deep Learning Model

Abstract:

Predicting Daily Streamflow with Deep Learning Model

REU SCHOLAR: NELL FLORES (1)

MENTORS: AREZOO BYBORDI (2), DR. RONAK ETEMADPOUR (3)

• Engineering Science, Queensborough Community College
• The Graduate Center, CUNY
• Computer Science Department, City College of New York, CUNY

Streamflow tells us how much water is flowing through a river, or stream. As a part of the water cycle stream flow can be combined with other data to tell a story. One tragic story is that of California’s droughts. Droughts are often associated with a lack of precipitation, lack of aquifer, etc. For example, lack of precipitation influences stream flow. Less water dropping from the sky means less water to be transported back to the ocean or other bodies of water. Our goal is to be able to predict streamflow so that California drought crisis area management could potentially use the data to combat drought. Our project will involve the use of weather and streamflow data from about 300 sites. In order to reduce the dimensionality of the data we used watersheds which group multiple sites together. Our watershed of focus is Lower Cosumnes watershed. To use the data, we will implement various deep learning models of varying complexity. These models will take all our climate data inputs and ascertain discrete information on their relationships and in the end give one value for the predicted streamflow.

We have examined two deep learning models. One was a straightforward linear regression which only used one variable while the other one required sequential matrices and used a LSTM layer. The complex model required the building of a normalized data frame of 6 sites each with their climate variables. We were also able to calculate seasonality’s for each climate variable and use those values to preprocess my data or in other words make it cleaner. We will use visualization techniques on the model to get better insight of the model’s functionality in the hidden layers.

Keywords: Streamflow, Drought, Deep Learning, Prediction, Watershed, California