meng-zha / PM2.5Forecast

PM2.5Forecast

This repository represents the final project of pattern recognition in 2020 spring for PM2.5 prediction with a Seq2Seq network and attention. If you want to know more, please read report.pdf

Dataset

The data can be downloaded from Tsinghua Cloud.

Requirements

Python 3.7 or later with all requirements.txt dependencies installed, including torch >= 1.3. To install run:

$ pip install -U -r requirements.txt

Modules

• create.py: read the csv files and convert them into a h5 file.
• clean.py: clean the data and split the data for dataloader.
• train.py: train the model.
• test.py: evaluate and test the model.
• dataloader.py: define the AirDataset.
• model.py: define the model of encoder and decoder.

Prepare data

The data before 20140429 and on 20141231 is abandoned. Please delete them at first. To prepare the database:

$ python create.py 'trainval'
$ python create.py 'test'

The prepared database can be downloaded from PM2.5 database.

To get the final train/val/test split:

$ python clean.py create_dataset --seed=0

Train and test

To train the model, please run:

$ python train.py --lr=1e-3
                  --epochs=100
                  --hidden=8

To evaluate the model, please run:

$ python test.py --mode=val
                 --checkpoints = YOUR_MODEL_PATH
                 --model= YOUR_MODEL_NUMBER

To test the model, please run:

$ python test.py --mode=test
                 --checkpoints = YOUR_MODEL_PATH
                 --model= YOUR_MODEL_NUMBER

Results and Model

Our baseline get 83.29% acc on val dataset and 64.61% acc on test dataset.

mAP

	1hour mAP	2hour mAP	3hour mAP	4hour mAP	5hour mAP	6hour mAP	total mAP
val	83.69	83.10	83.50	83.70	83.56	82.21	83.29
test	76.09	70.33	65.53	61.65	58.53	55.51	64.61

MAE

	1hour MAE	2hour MAE	3hour MAE	4hour MAE	5hour MAE	6hour MAE	total MAE
val	8.718	8.946	8.731	8.601	8.736	9.395	8.855
test	13.86	17.53	20.96	24.01	27.78	29.35	22.08

models

Our baseline models can be found in baseline_model. Or you can download from PM2.5_baseline_models.

Contact

Please don't contact me if you find some bugs. If not, my mailbox will be overflowing.