Forecasting Ozone Pollution using Recurrent Neural Nets and Multiple Quantile Regression

Folders and Files:

• all_scores: All scores obtained from the models in the same folder.
• data: Datasets used, containing only Summer data.
• models: Cache of the trained models.
• model.backup: When a loading is interrupted, the model try to train again, so if that new training is canceled, the cache will be overwrited, leaving slight variations in the scores.
• precalcs: Precalculate data from the dataset, this 'precalcs' are only for test and is not used in this study, but it's necessary to the properly work of the code.
• tuning.done: Tuning done with Hyperas. For each model have 3 files:
  • .py: Code with the configuration to start the tuning. the arguments <HyperLSTM | preSQP> < Number of runs >
    • i.e: file.py HyperLSTM 35
  • .temp.txt: The output of the corresponding .py
  • .txt: The tail of the corresponding .temp.txt
• work: This is the main folder. All .py files must be in this folder to work properly, also the ozone_forecasting_multi-task notebook.

For all models:

• Optimizer: Adam.
• Learning Rate: 0.001
• B₁: 0.9
• B₂: 0.999