In the scientific community it is more or less established that solar activity has a seasonality of 11 years. In general, scientists use sunspots per unit time to quantify the strength of a solar cycle.
In an era increasingly dominated by machine learning, our daily lives are profoundly intertwined with satellite technology. This evolving landscape underscores the critical need for a reliable solar cycle prediction model, which my project aims to address through the innovative application of neural networks.
Sunspots are generated by a stochastic process, making it harder to predict analytically. The models we have so far does not work quite well. For example, in July 2023, the prediction form the top model from NASA's heliophysics division is half of the actual numbers observed by National Oceanic and Atmospheric Administration (NOAA).
I used the data from Kaggle. The data as a csv file is in the folder "Data"
-
!pip install kaggle
-
!mkdir ~/.kaggle
-
copy the kaggle.json file to .kaggle folder
!cp json/file/path .kaggle/file/path -
!kaggle datasets download -d abhinand05/daily-sun-spot-data-1818-to-2019.
5.Unzip the dataset.
!unzip daily-sun-spot-data-1818-to-2019.zip -d data/
There are a lot of algorthms that can be used for time series forecasting. I implemented these algorithms using tensorflow and compared their performance.
|
tensorflow |
seaborn |
pandas |