A convolutional neural network to predict the monthly average active sunspots in the Sun. The data is available through the Y-Space API. The Jupyter Notebook sunspots-estimator.ipynb includes the process of creating the dataset and training the model.
The model is defined using a combination of a convolutional layer, LSTM layers, and dense layers. The output is a single node layer -- the predicted number of sunspots:
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv-1 (Conv1D) (None, None, 32) 192
_________________________________________________________________
lstm-1 (LSTM) (None, None, 60) 22320
_________________________________________________________________
lstm-2 (LSTM) (None, None, 60) 29040
_________________________________________________________________
fc-1 (Dense) (None, None, 20) 1220
_________________________________________________________________
fc-2 (Dense) (None, None, 10) 210
_________________________________________________________________
output (Dense) (None, None, 1) 11
=================================================================
Total params: 52,993
Trainable params: 52,993
Non-trainable params: 0
The following plot illustrates the original values (the blue line) versus the model predictions (orange line.).
The final version of the model is saved in sunspots-estimator.h5.
Thank you to Laurence Moroney for the initial bootstrap and motivation.