Intro

Introducing SYNTHETIC HARIBROS Team solution

Features

1. Time-Variant GAN

• Minibatch predictions
  
• Detect unusual patterns
  
• Learn complex patterns
  
• Discriminator teaches the Generator until we find optimums for synthetic data
  

2. Generator and Discriminator Loss Functions

Getting Started

To recreate our application, the easiest way to get started is to clone the repository:

# Get the latest snapshot
git clone https://github.com/balintdecsi/synthetic-haribros-watts-up.git myproject

# Change directory
cd myproject

# Install dependencies
python3 -m pip install -r requirements.txt

Note that you should have access to GPUs to run this code.

About our Solution

Challange of Energy Networks
• Energy Transition is on the corner
• Managing renewables demand smart grids
• Smart meters for smart grids
• Roll-out and GDPR concerns
• Synthetic data paves the way

Model architecture decisions...

GANs

Pros

• Minibatch predictions
• Detect unusual patterns
• Learn complex patterns

Cons

• Data quality dependent
• Resource hungry

LLMs

Pros

• Reports and summaries
• Contextual insights

Cons

• Not great for numbers
• One token at a time
• Sticks to first part of data

Generative Adversarial Networks

• Discriminator teaches the Generator until we find optimums for synthetic data
• We chose Time-Variant GANs

Grid search and findings

Ideal number of EPOCHs

• cca. 1050

Target:

• Generator loss decreases as Discriminator loss increases
• Find optimum

Alternatives for optimization

• Different training patterns
• Alternated Optimizers
• Feature Engineering

Discriminator and Generator Loss

• After cca. step 875 losses diverge in multiple scenarios(5;6)
• Early stopping is very difficult due to volatility
• Semi-Manual optimums should be used

End result

• After cca. step 875 losses diverge in multiple scenarios(5;6)
• Early stopping is very difficult due to volatility
• Manual optimums should be used

Acknowledgement

Synthetic Energy Data Generation Using Time Variant Generative Adversarial Network
Generative Adversarial Nets for Synthetic Time Series Data

Team Members:

Balint Decsi
Gabor Schwimmer
Tamas Sueli
Zoltan Takacs