Transformer-Based SAC for Wind Farm Yaw Control

A transformer-based Soft Actor-Critic (SAC) agent that learns generalizable yaw control policies for wind farms. Trained on diverse farm layouts, it can deploy zero-shot to unseen configurations.

Key Features

• Zero-shot generalization — Train on small farms (3-6 turbines), deploy on larger ones (20-25 turbines) without retraining
• Layout transfer — Learn from grid layouts, transfer to irregular and circular arrangements
• Interpretable attention — Attention weights reveal which turbines influence each control decision
• Wind-relative encoding — Positions are transformed to a canonical wind frame, making the policy invariant to absolute wind direction
• Modular architecture — Pluggable positional encodings (MLP, sinusoidal, polar, ALiBi, relative bias) and profile encoders (Fourier, CNN, dilated, attention-based)

Quick Start

Training:

python transformer_sac_windfarm.py \
    --train_layouts T1 T2 T3 T4 T5 T6 \
    --eval_layouts E1 E2 E3 \
    --total_timesteps 500000 \
    --pos_encoding_type absolute_mlp \
    --use_profiles \
    --seed 1

Evaluation:

python evaluate.py \
    --checkpoint runs/<run_name>/checkpoints/step_500000.pt \
    --eval_layouts E1 E2 E3 E4 E5

Project Structure

Path	Description
transformer_sac_windfarm.py	Main training script (SAC + transformer architecture)
agent.py	WindFarmAgent — wraps the actor for inference
evaluate.py	Evaluation pipeline
eval_utils.py	Evaluation helper functions
helper_funcs.py	Checkpoint I/O, coordinate transforms, env utilities
MultiLayoutEnv.py	Multi-layout environment for training across farm configurations
positional_encodings/	Positional encoding modules (absolute, bias, GAT, neighborhood)
profile_encodings/	Wake profile encoders (Fourier, CNN, dilated, attention)
receptivity_profiles.py	Compute turbine receptivity/influence profiles via PyWake
geometric_profiles.py	Geometry-based profile approximations
pretrain.py	Behavioral cloning pretraining
extract_attention.py	Extract attention weights for analysis
Notebooks/	Plotting and analysis notebooks (WES paper figures)
archive/	Historical development code (old iterations, experiments)

Approach

Each turbine is treated as a token in a transformer sequence:

1. Per-turbine tokenization — Local observations (wind speed, direction, yaw) become token features
2. Wind-relative positional encoding — Turbine positions are rotated so wind always comes from a canonical direction, then encoded via MLP (or other schemes)
3. Wake profile conditioning — Optional Fourier-encoded receptivity/influence profiles provide layout-aware context
4. Permutation-equivariant output — Shared actor/critic heads produce actions for all turbines simultaneously

The transformer naturally handles variable-length sequences, enabling a single policy to control farms of different sizes.

License

This project is licensed under the MIT License. See LICENSE for details.