This is our solution to the KDD Cup 2022 spatial dynamic wind power forecast challenge, see the competition webpage for more information of the challenge itself.
Team name: didadida_hualahuala
Placement: 6th (of 2490 teams)
Final score (3rd phase): -45.18139
The solution uses a combination of two models: MDLinear and XGTN, see the technical report for the details. A quick summary can be found in our presentation slides and our video presentation. The trained models used for the final score are included in this repository.
The training data can be downloaded on the competition website: https://aistudio.baidu.com/aistudio/competition/detail/152/0/datasets.
Put this file into the data folder before starting to train the models.
All parameter settings are adjusted in the methods/prepare.py file. The default settings were used for the competition results.
To train the models, run
python train_mdlinear.pyand
python train_xtgn.pyin any order. The trained models and any relevant files are saved to the methods/checkpoints folder (this folder is shared for both methods).
To evaluate our method, we use the provided test dataset (in data/test_x and data/test_y). The input data contains 14 days and since we do not require that much we use a sliding window to create more test data (see the techincal report). The code for this is included in data/split_test_file.py. To use the single test file instead, adjust the values of path_to_test_x and path_to_test_y in methods/prepare.py.
To run the forecast and evaluate the score, use:
python evaluate.py