CogniteDatathon

The windspeed plot for KBQX station (predicted with prophet model)

Inspiration

Tackling interesting DSCI problems. We wanted to see how we could apply our recently learned knowledge in time series modeling and data dashboarding. One of our teammates recently learned how to use Streamlit for another project and another recently learned a lot about time series modeling and we wanted to do a project that brings those together

What it does

Predicts and visualizes Wind Speed and Direction

How we built it

The model was built using R. We implemented 4 popular time series models suitable for the dataset: ARIMA (Autoregressive Integrated Moving Average), seasonal ARIMA, Garch, and prophet model. The prophet model is an automatic time series modeling and prediction framework developed at Facebook. Prophet works espeically well on data with nonlinear trend and strong seasonality effects, and it is also robust to missingness and outliers. Our data suffers from missingness and extreme values. The empirical results show that prophet outperform other models in terms of minimizing the cross validation errors, which is likely due to the seasonality effects, missingness and extreme values present in the data. See the plot in the additional material. The dashboard on the other hand, was built using streamlit and plotly to create an interactive dashboard, with interactive and animated visualizations.

Challenges we ran into

Cleaning the data was a large challenge because the delimiters were not entirely consistent and the time data was formatted in separate columns that required wrangling to understand.

Accomplishments that we're proud of

We were able to make visualizations that were both animated and interactive, depending on what the user wanted. Additionally, we were able to make use of advanced time series models, such as Facebook’s publicly available model to effectively predict complicated weather data into the future.

What we learned

Throughout this datathon, we learned a number of things. One such thing is how to use the basic features of the Streamlit Python package. The package was quite intuitive to use, allowing simple API methods for adding elements to the webpage. It would have been nice if the library allowed for finer control of the site, as the user does not have control of when elements on the site load, so things can sometimes run a bit slow.

What's next for Cognite Datathon

Most of us are seniors wrapping up our final semester at Rice. Some of us are moving straight into industry and some of us are planning to attend grad school to expand our knowledge. We’re all really excited about the future of Data Science and can’t wait to see what advances the field comes up with.