Power Grid Gans

Generative Adversarial Networks (GANs) are nowadays not only used to create novel digital artworks, but can also be found in various other domains such as text generation, music, medicine, finance, smart grids, general training data generation and many more.

This work also addresses the area of smart grids by building on the knowledge gained from other areas and existing GAN extensions to create an extensible GAN framework that makes it possible to generate novel power grid simulation data, initially using weather data, for MIDAS.

This is an experiment-based work, where the developed methods and concepts were first applied to sinusoidal synthetic data in the form of a feasibility study, and then extended to weather data of the power grid simulation of MIDAS, tested against a set of metrics, and additionally validated visually with the help of plots.

A GAN framework was created that allows the generation of similarly distributed weather simulation data using a CGAN. Great attention was paid to the extensibility and versatility of the GAN framework through an elaborated software architecture. During the development many problems of the GAN training arose and became visible. The generated data using the developed framework currently contains an increased temporal oscillation in contrast to the training data, which could potentially be reduced by a variety of techniques based on this work.

Table of Contents

• Requirements
• Getting Started
• Development
  • Running Tests
  • Profiling
  • Problems
• Misc
  • Create .gif from images
• Docs
  • GAN Framework
    • Dataflow
    • Architecture
    • CLI

Requirements

To run the code python3.9 is required. All package requirements are listed in the requirements.txt

Getting Started

It is recommended to run this code in a virtual python environment. To do this, install virtualenv using pip with pip3 install virtualenv on Linux. It is also recommended to use pip-tools.

To create the virtual environment use virtualenv:

virtualenv -p python3 power-grid-gans
# or
python3 -m venv power-grid-gans

Activate the virtual environment by running

source power-grid-gans/bin/activate

Next install the requirements from the requirements.txt If you want to use pip-tools, install it first and run:

pip-sync

If you are using pip execute the following:

pip install -r requirements.txt

Deactivate virtual environment by running

deactivate

Development

Running Tests

python3 setup.py pytest

Profiling

1. Tensorboard

When training models we can use tensorboard to keep track of out models. Run tensorboard with the following command:

tensorboard --logdir . 

Problems

1. Missing CUDA Version of torch

Currently pip-tools cannot specify where to download torch and what specific flavor so for cuda support do this manually. Pip command can be generated here: https://pytorch.org/get-started/locally/

Sample:

pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio===0.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Maybe you also need to uninstall pytorch first and purge the cache manually:

pip uninstall torch
pip cache purge

2. Atari FileNotFoundError: Could not find module ale_c.dll

1. Uninstall gym and atari-py (If already installed):

pip uninstall atari-py
pip uninstall gym[atari]

2. Download VS build tools here: https://visualstudio.microsoft.com/thank-you-downloading-visual-studio/?sku=BuildTools&rel=16

3. Run the VS build setup and select "C++ build tools" and install it.

4. Restart PC.

5. Install cmake, atari-py and gym

pip install cmake
pip install atari-py
pip install gym[atari]

Misc

Create .gif from images

ffmpeg -f image2 -framerate 10 -i img_%03d.png -loop -1 animated.gif

Docs

GAN Framework

Dataflow

Architecture

CLI

Help overview

python3 -m src.pggan --help

Train

python3 -m src.pggan

Eval

python3 -m src.pggan --mode eval