su22-reu-382

Name: Junyang (Eric) He
Email: fay5du@virginia.edu

To Do List

• Review previous research applying deep learning to Hydrology
• Find hydrology datasets for other nations based on the standards set by CAMELS-US
• Compare datasets and find the list of common static and dynamic variables
• Preprocess CAMELS-US, CAMELS-UK, and CAMELS-CL datasets to contain the same set of variables with Pandas
• Preprocess CAMELS-US, CAMELS-UK, and CAMELS-CL datasets into numpy arrays of the same format suitable for training
• Train LSTM model with the three input datasets
• Train Science Transformer model with the three input datasets
• Train TFT model with the three input datasets
• Compare efficiency of the three models
• Apply transfer learning

What I Knew

Programming Languages:

• Python
• JAVA
• R

Data Structures:

• Array
• Linked list
• Tree
• Stack
• Queue
• Hash table

Machine Learning:

• Unsupervised learning (K-means clustering)

Data Analysis:

• Pandas library
• Matplotlib library
• Seaborn library
• Linear and logistic regression
• Statistical graphs

What I Learned

Programming:

I had the chance to strengthen my general Python programming skills and apply the knowledge I learned in class to solve real problems. I applied my knowledge in Pandas library to preprocess CAMELS data from US, UK, and Chile into a form suitable for the deep learning algorithm. I also learned how to create my own repository on Github and some basic command line operations.

Machine Learning - Deep Learning:

Before the REU experience, I had limited knowledge about reinforcement learing and deep learning. All I knew was the name of some of the deep learning neural networks such as the CNN widely used in image recognition, the LSTM widely used in time series prediction, and the Transformer widley used in Seq2seq prediction. During the REU, I got to know the detailed structure of the LSTM and Transformer arthitecture and the strengths and weaknesses of each type of neural network. I gained hands-on experience writing deep learning code with the Tensorflow library in Python.

Research Process

Through the two months, I learned the proper research process. I learned how to choose the proper method at each step of research to ensure that our research keeps up with others in the field of Hydrology. In data preprocessing and normalization, I learned the importance of taking every normalization step with caution to ensure that hidden patterns embedded in some numerical and categorical variables weren't lost. Generally, I learned how to solve obstacles by reading other papers on the topic of DL in Hydrology.

Universal Approach to Science Time Series: Deep Learning on Hydrology Poster

• PDF and PPTX versions of the poster can be found here.
• Images used in the poster can be found here.