High-Level Project Summary
The objective was to develop a neural network pipeline to track and follow the changes in the peak solar wind speed. In this project, I developed and deployed a neural network model to predict solar flares and also to predict changes in magnetic fields. I also analysed the dataset and trained the model on relevant features. Since the dataset has a record of solar flare activity for almost 10 years, we can forecast all the features that the classification model takes as input (by building time series regression models) and predict future solar flares. I used the TabNet model, for predicting solar flares and also for time series regression. The accuracy of the classification model is 95%.
Link to Final Project
Link to Project "Demo"
Detailed Project Description
The objective was to develop a neural network pipeline to track and follow the changes in the peak solar wind speed. The dataset used was a result of the research paper titled: Data set for solar flare prediction using helioseismic and magnetic imager vector magnetic field data
In this project, I developed and deployed a neural network model to predict solar flares and also to predict changes in magnetic fields. I also analysed the dataset and trained the model on relevant features.
Since the dataset has a record of solar flare activity for almost 10 years, we can forecast all the features that the classification model takes as input (by building time series regression models) and predict future solar flares. I was able to do this by training models on data ranging from 2010 to 2017 and testing them on data ranging from 2018 to 2019. A sample plot is shown below. (Notice the low error values at the bottom left)
I used the TabNet model, for predicting solar flares and also for time series regression. The accuracy of the classification model is 95%. The final project also allows you to interact with the classification model.
The final project is available at Project Link
Tools used: Pandas, streamlit, sckit-learn, pytorch_tabnet
Coding language: Python
Space Agency Data
The dataset used was a result of the research paper titled: Data set for solar flare prediction using helioseismic and magnetic imager vector magnetic field data
Hackathon Journey
I had a very fruitful and enjoyable experience in this hackathon. I learnt a lot about time series data, big data handling and model deployment to the web. The willingness to work with large data and make deployable models from it inspired me to choose this challenge. Initially, I tried various classification models until I found the best. Then I tried to make time series models using the autots library but I could not load saved models using that, so I used the TabNet architecture for time series forecasting. I also faced some issues with model deployment.
References
Tags
#AI, #Software, #SolarFlares