This is the official repo for the paper E2E-AT: A Unified Framework for Tackling Uncertainty in Task-aware End-to-end Learning, to be appeared in AAAI-24. You can find the preprint here.

The authors include:

• Wangkun Xu: Control and Power Research Group, Department of EEE, Imperial College London.
• Dr. Jianhong Wang: University of Manchester.
• Dr. Fei Teng: Control and Power Research Group, Department of EEE, Imperial College London.

This repo is under development but should contain the main functionality to support the paper results. We are currently working on merging the E2E-AT into a more general code base for reasearches in E2E learning and power system community, to support diverse power system operations, such as stochastic programming, integer programming, etc. Please stau updated.

The core packages include cvxpy for solving optimization problems, cvxpylayers for embedding optimization problems as differentiable layers in NN and pytorch. pypower is also used to retrieve the power system configuration. We have tested with the versions:

torch==2.0.1+cpu
pypower==5.1.16
cvxpy==1.4.1                    
cvxpylayers==0.1.4

We use open source dataset from A Synthetic Texas Backbone Power System with Climate-Dependent Spatio-Temporal Correlated Profiles. You can download/read the descriptions of the dataset from the official webpage.

Download the the .zip file into the data/ folder and change the name into raw_data.zip, unzip the file by

unzip data/raw_data.zip -d data/

This should give you data/Data_public/ folder.

Then to generate the dataset used in the paper, run

python clean_data.py --no_bus 14

This will generate the feature and load pairs for each bus in bus-14 system.

We have modified the standard IEEE case files from pypower. The modifications can be found in 'config.json'.

All the configurations can be found in configs.json.

For clean training using mse loss:

python trainer_clean_mse.py

This will store the best model and also the model that is used to warm start the training with objective-based loss.

For robust training using mse loss (attack on the input only):

python trainer_robust_mse.py --max_eps {eps_parameter}

For clean training using objective-based loss:

python trainer_clean_spo.py -p

where -p represents to retrain the model using the clean_mse. The pre-trained model is saved as trained_model//full/mse_warm.pth.

For robust training using objective-based loss:

python trainer_robust_spo.py -p spo_clean -t {train_method} -a {parameter} -i {random} --eps_parameter {eps_parameter} --eps_input {eps_input} --alpha {alpha}

where -t: choose from 'normal' for regular adversarial training; 'free' for adversarial training for free. -a: choose the attack method from 'input', 'parameter', or 'both'. -i: choose the attack initializationg method: 'random' for random initialization in the range, 'previous' for use the previous generated attack for initialization, or 'default' for start at the clean sample. --eps_parameter: the attack strength on the parameter (in ratio, e.g. 0.1) --eps_input: the attack strength on the input. --alpha: the parameter to balance the clean and adversarial training loss.

To certify the input space attack:

python exp_certify.py -t spo_clean --eps_input 0.005

solver_args: According to the document of cvxpylayers, you can change the solver_args in the forward pass of solving the optimization problems. You can change the configs under "nn:solver_args". For example, increase the "eps" can significantly improve the training speed.

To cite our paper:

@article{xu2023e2e,
  title={E2E-AT: A Unified Framework for Tackling Uncertainty in Task-aware End-to-end Learning},
  author={Xu, Wangkun and Wang, Jianhong and Teng, Fei},
  journal={AAAI24},
  year={2024}
}