This is our solution to the KDD Cup 2022 spatial dynamic wind power forecast challenge.
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 [to be added when available on the competition homepage] 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