Climate Prediction - 2021 CIS&CMU Project

Problem

Predicting future temperature with machine learning methods.

Dataset

National Centers for Environmental Information Global Surface Summary of Day Data (GSOD). See GSOD README for information on how to obtain it and/or its terms and restrictions.

In this project, we are mainly using station 722860 since it has a long recording history. To run the notebooks, place the dataset at runtime/GSOD, so that runtime/GSOD/2021 contains .op or .op.gz data files. For example, runtime/GSOD/2021/583620-99999-2021.op.gz should exist.

Items

Use those Notebooks by this order to start.

• gsod.py: Dataset helper and common functions.
• 00-Search_Stations.ipynb: Some helpers to find good stations (optional).
• 01-FFT.ipynb: Real Fourier transform analysis of the recurrent period (optional).
• 10-EMD.ipynb: Complete empirical mode decomposition of periods and trends.
• 11-LSTM.ipynb: Building a long short-term memory model.
• 12-MLP.ipynb: Building a multilayer perceptron model. (Model by @stevenli-phoenix)
• 13-LSTM_Only.ipynb: A comparative LSTM-only approach (optional).
• 14-Baseline_and_LR.ipynb: Building a long short-term memory model.
• 15-ARIMA_and_SVM.ipynb: Comparative ARIMA and SVM models (optional).
• 20-Ensemble.ipynb: Combining different models for different mode functions.

Historical Models

Early experiments. Kept here for reference.

• ClimateMarkov.ipynb: First-degree Markov-like model.
• ClimateMarkov2.ipynb Any-degree Markov-like model.
• ClimateDiscreteMarkov.ipynb Any-degree discrete Markov-like model.

Citation

BibTeX entry:

@inproceedings{10.1117/12.2637187,
    author = {Maiyun Zhang and Shuyu Li and Yiqi Wang},
    title = {Short-term temperature prediction based on hybrid {CEEMDAN}, neural networks, and linear regression methods},
    volume = {12285},
    booktitle = {International Conference on Advanced Algorithms and Neural Networks (AANN 2022)},
    editor = {Rajeev Tiwari},
    organization = {International Society for Optics and Photonics},
    publisher = {SPIE},
    pages = {301 -- 308},
    year = {2022},
    doi = {10.1117/12.2637187},
    URL = {https://doi.org/10.1117/12.2637187}
}

License

Private until the project is finished. TBD.