A110-06
Streamflow Prediction using Deep Learning from the NOAA-CREST Observation Site in Millbrook, New York
Abstract:
REU SCHOLAR: ANA ABREU AUTHOR: Dr. RONAK ETEMADPOUR
Mentor: Arezoo Bybordi
CREST Research Experience for Undergraduate Program
Mathematics, Bronx Community College, CUNY
Mathematics and Computer Science Department, The City College of New York, CUNY
ABSTRACT –
Streamflow is the flow of water in streams, rivers and other channels. Changes in streamflow can influence the amount of water available for crops, the generation of electricity, fishery, many plants and animals. The ultimate goal of this project is to predict daily streamflow values of locations in California using 300 sites with deep learning model. The data is provided by the US Geological Survey, which has daily streamflow data from the year 1980 to 2019. These measurements include daily data, streamflow and precipitation from different locations. Our challenges are having many dry locations in West of the US, and the use of deep learning methods in all the West of the US.
We analyzed the factors that affect streamflow, which are the seasonality, precipitation, and dry and wet months. We predicted streamflow using windows of 30-days of streamflow and precipitation data. The method used was a deep learning model including a LSTM layer. Also, we determined the months that have been driest for the locations inside a specific watershed (Lower Cosumnes watershed) since the locations inside a watershed have similar climatic patterns. We have used bar charts and heatmaps for this analysis. Then, using Support Vector Regression as a machine learning model, we built a baseline model using windows of 30-days of streamflow data as the input. Besides, we are going to predict streamflow using a more complicated deep learning model, which is called a variational autoencoder.
Keywords: US Geological Survey, precipitation, seasonality, LSTM layer, Lower Cosumnes watershed.