Experiments in climatological time series analysis using deep learning.
Loss (MSE) 0.00037
on SSN with 64-layer LSTM and 400 epochs. See the notebook.
Interestingly, the network is trained on 66% of the SSN data but correctly predicts the weakness of solar cycle 24.
Next step: predict solar cycles 25, 26, 27!
- create a generic LSTM framework/notebook
- analyze SSN (Solar Sunspot Numbers) monthly series
- analyze Global/Local Datasets (temperature, precipitation, etc)
- analyze climatic indices (ENSO, etc)
- modify the network in order to accept Continuous Wavelet Transform output
- generate signal from predicted CWT spectra
John Abbot et al.: The application of machine learning for evaluating anthropogenic versus natural climate change, GeoResJ (2017). DOI: 10.1016/j.grj.2017.08.001
Qin Zhang et al.: Prediction of Sea Surface Temperature using Long Short-Term Memory. arXiv:1705.06861 [cs.CV]
Bao W, Yue J, Rao Y: A deep learning framework for financial time series using stacked autoencoders and long-short term memory. Podobnik B, ed. PLoS ONE. 2017;12(7):e0180944. doi:10.1371/journal.pone.0180944
Franco Zavatti: Clima, Reti Neurali, Dati di Prossimità e Analisi Spettrali. http://www.climatemonitor.it/?p=46061
https://github.com/simaaron/kaggle-Rain
https://thesai.org/Downloads/Volume8No2/Paper_43-Prediction_by_a_Hybrid_of_Wavelet_Transform.pdf
SSN Sunspot Number - Source: WDC-SILSO, Royal Observatory of Belgium, Brussels