Deep Uncertainty Quantification (DUQ)

DUQ: A Machine Learning Approach for Weather Forecasting

  1. Sequential deep uncertainty quantification (DUQ) produces more accurate weather forecasting based on the observation and NWP prediction. Our online rank-2 (CCIT007) in Global AI Challenger-Weather Forecasting ( indicates deep learning is very considerable for large-scale meteorological data modeling and forecasting!

  2. Pragmatical loss function for sequence-to-sequence uncertainty quantification is proposed.

  3. Important experimental phenomenon was reported and analysized experimentally, which may be noteworthy in the furture deep learning researches for spatio-temporal data and time series forecasting.





3 mins promotional videos:

If you feel it helpful, please cite our paper:

 author = {Wang, Bin and Lu, Jie and Yan, Zheng and Luo, Huaishao and Li, Tianrui and Zheng, Yu and Zhang, Guangquan},
 title = {Deep Uncertainty Quantification: A Machine Learning Approach for Weather Forecasting},
 booktitle = {Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \&\#38; Data Mining},
 series = {KDD '19},
 year = {2019},
 isbn = {978-1-4503-6201-6},
 location = {Anchorage, AK, USA},
 pages = {2087--2095},
 numpages = {9},
 url = {},
 doi = {10.1145/3292500.3330704},
 acmid = {3330704},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {deep learning, uncertainty quantification, urban computing, weather forecasting},


I test it on MacOS and Ubuntu. It is based on Python 3.6. Required packages like keras, tensorflow etc. are iincluded in requirements.txt. Run bellow command to install them.

pip install -r requirements.txt

Pipeline for quick start.

I noticed that the competition link has been closed. However, competitors have uploaded the raw dataset.

Training set:
Validation set:
Test set (Taking one-day test data on 28/10/2018 as an example):

After downloaded, set three original dataset into the folder ./data/raw/ (For quick start, we just take ai_challenger_wf2018_testb1_20180829-20181028 as a test example, you can easily re-name the related arguments and apply it on other test set, which can be introduced in later section How to change test dataset for other days? )

Ensure that you are located in the project root directory and implement commands strictly according to bellow order:

  1. make make_train_data (Prepare training data)

  2. make make_val_data (Prepare validation data)

  3. make make_TestOneDayOnlineData_for_submit (Prepare one-day test data)

  4. make train_from_scratch (Train)

  5. make load_single_model_and_predict (Test)

    • Then a file called 'Seq2Seq_MVE_layers_50_50_loss_mae_dropout0-2018102803_demo.csv' is created in the path './src/weather_forecasting2018_eval/pred_result_csv/'. To evaluate, locate yourself to the path'./src/weather_forecasting2018_eval' and run:
  6. make evaluate_1028_demo

    • You will get the resulted score (This day i.e., 28/10 is relatively hard to predict, we only got score of 0.2408 using this demo model). BTW, to reproduce our online scores, you can evaluate all submitted files in 'src/weather_forecasting2018_eval/pred_result_csv/submit_csv' by changing evaluated file name.
    • Since ensemble can improve the prediction accuracy and stability, I have trained many models for you to try ensemble! Go back to the the project root directory and run next command 7:
  7. make load_multi_models_pred

    • Three different models are loaded and predict in turn. The resulted .csv file of each model is saved in './src/weather_forecasting2018_eval/ensemble_2018102803/'
    • Then locate yourself to ./src/weather_forecasting2018_eval/ensemble_2018102803/ and run next command 8:
  8. python

    • Here ensemble is to simply calculate mathematical mean. You will get the file called ensemble_avg_2018102803.csv. Locate yourself to ./src/weather_forecasting2018_eval and run:
  9. make evaluate_1028_demo_ensemble

    • You will find the score of ensemble. Maybe sometimes you can find that ensemble result is lower than single model. Don't be confused about this. The performance of a single model can fluctuate very much according to different dataset. However stability is a great trait of ensemble learning. We got score of 0.3440 using this demo ensemble model, which is a little higher than our online score 0.3358.

How to change test dataset for other days?

Everyday, we have three opportunities to submit. We will use 3 models

1. Ensemble model.

During the first competition days: we use models for ensemble including:

MODEL_LIST=seq2seq_subnet_50_swish_dropout \
        seq2seq_subnet_30_30_best \
        seq2seq_subnet_200_200 \
        seq2seq_model_100 \

During the last competition days: we try switching to models:

            seq2seq_subnet_200_200 \

2. Single model 1: Seq2Seq_MVE_layers_222_222_loss_mve_dropout0. (At ./models/model_for_official)

3. Single model 2: Seq2Seq_MVE_layers_50_50_loss_mae_dropout0 (At ./models/model_for_official)

Pipeline steps:

  1. Download raw test data oneline to the path: ./data/raw/

  2. Edit file '/src/data/'. Set file_name to the same of the downloaded file in bellow snippet:

    elif process_phase == 'OnlineEveryDay': file_name=''

  3. Modify Makefile file for the rule 'make_TestOneDayOnlineData_for_submit':

    • 3.1 Set --process_phase=OnlineEveryDay
    • 3.2 Set --datetime=2018xxxx to month and day of prediction.
    • 3.3 Run 'make make_TestOneDayOnlineData_for_submit'
  4. Modify Makefile for rule 'load_multi_models_pred': (Take datetime=2018xxxx as instance)

    • 4.1 Set --test_file_name = 'OnlineEveryDay_2018xxxx_norm.dict'
    • 4.2 Set --saved_csv_path = './src/weather_forecasting2018_eval/ensemble_2018xxxx03/'
    • 4.3 Set --saved_csv_name=$$model-2018xxxx03.csv
    • 4.4 Run 'make load_multi_models_pred'
    • 4.5 Locate to './src/weather_forecasting2018_eval/ensemble_2018xxxx03/', run 'python' and create 'ensemble_avg.csv' file. Copy and rename it to 'forecast-2018xxxx03.csv' then SUBMIT (1) it.
  5. Modify Makefile for rule 'load_single_model_and_predict'

    • 5.1 set MODEL_NAME_prediction = Seq2Seq_MVE_layers_50_50_loss_mae_dropout0 (Or another model you have trained)
    • 5.2 set --test_file_name='OnlineEveryDay_2018xxxx_norm.dict' \
    • 5.3 set --saved_csv_name=$(MODEL_NAME_prediction)-2018xxxx03.csv
    • 5.4 run 'make load_single_model_and_predict'
    • 5.5 set MODEL_NAME_prediction = Seq2Seq_MVE_layers_222_222_loss_mve_dropout0
    • 5.6 run 'make load_single_model_and_predict'
  6. Locate to './src/weather_forecasting2018_eval/pred_result_csv/':

    • 6.1 Rename Seq2Seq_MVE_layers_50_50_loss_mae_dropout0-2018xxxx03.csv to 'forecast-2018xxxx03.csv' and SUBMIT (2)
    • 6.2 Rename Seq2Seq_MVE_layers_222_222_loss_mve_dropout0-2018xxxx03.csv to 'forecast-2018xxxx03.csv' and SUBMIT (3)
  7. Sanity check by visualization (not try).

    • We'd better plot for Sanity check before submitting! (in notebook or Excel). But we do not try this.

Customize model parameters

You can dive into the './src/models/'. Due to the parameters of the deep model are too many. Here we do not play exhaustedly. We mainly play different parameters of 'self.layers' and use ensemble to combine shallow and deep seq2seq model.

Project Organization

├── Makefile           <- Makefile with commands like `make data` or `make train`
├──          <- The top-level README for developers using this project.
├── data
│   ├── external       <- Data from third party sources.
│   ├── interim        <- Intermediate data that has been transformed.
│   ├── processed      <- The final, canonical data sets for modeling.
│   └── raw            <- The original, immutable data dump.
├── docs               <- A default Sphinx project; see for details
├── models             <- Trained and serialized models, model predictions, or model summaries
├── requirements.txt   <- The requirements file for reproducing the analysis environment, e.g.
│                         generated with `pip freeze > requirements.txt`
├── src                <- Source code for use in this project.

Project based on the cookiecutter data science project template. #cookiecutterdatascience