GE-STDGN

Spatio-temporal weather prediction model based on graph evolution

Requirement

• pytorch (see https://pytorch.org/get-started/locally/ for how to install it)
• GPU
• Numpy (see http://www.numpy.org/)

Dataset

• The source dataset is in Google Drive and Baiduyun with code ni44.

• The source dataset and processed dataset are in Google Drive

Directory description

├── Readme.md                   // help
├── requirements.txt            // 
├── GE                          // graph evolution
│   ├── Geo_threshold.py        // the adjacency matrix generation method based on geographical threshold
│   ├── KNN.py                  // the adjacency matrix generation method based on k nearest-neighbor 
│   ├── TIN.py                  // the adjacency matrix generation method based on triangulated irregular network
│   ├── parallel.py             // multi-GPUs parapall control
│   ├── population.py           // define about population
│   ├── utils.py                // define about genetic operations
│   ├── individual.py           // define about individuals
├── baseline.py                 // baselines based on traditional machine learning methods
├── feature_selection.py        // feature selection
├── graph_evolution.py          // setup of graph evolution
├── main.py
├── measure.py                  // evaluation standard
├── model.py                    // model of STDGN
├── model_fs.py                 // model used in feature selection
├── tester.py
└── trainer.py

Expiremental setup

graph generation

• k->[6,23], β is adopted to weigh distance and altitude is 0.8 in KNN
• threshold in distance is 300km, and altitude is 1200m

graph evolution

• mutation 0.1
• crossing 0.8

|