windopt: Multi-Fidelity Wind Farm Layout Optimization

A framework for efficiently and accurately optimizing wind turbine layouts using Bayesian optimization, combining cheap analytical wake models with gold-standard large-eddy simulations.

Overview

Wind farms produce more power when turbines are optimally positioned, but finding the best layout is challenging. Gold-standard approaches using high-fidelity large-eddy simulations (LES) are computationally expensive, while simplified analytical wake models may miss important flow physics.

This project adaptively combines information from both approaches to efficiently find optimal layouts.

Preliminary results indicate that the multi-fidelity approach can find near-optimal layouts with significantly (60%) fewer expensive LES evaluations compared to single-fidelity methods.

Multi-fidelity optimization (right) reaches near-optimal designs in <20 simulations (4 batches), while single-fidelity LES optimization (left) requires 60+ simulations (12+ batches) to achieve similar performance. Both significantly outperform a standard grid layout (red line).

The framework:

• Combines the approximate Gauss-curl hybrid (GCH) wake model with large-eddy simulations (LES)
• Employs multi-task and multi-fidelity Bayesian optimization to efficiently balance between fidelities
• Supports batch parallel evaluation of candidate layouts

Requirements

• Python 3.10+
• Ax Platform 0.4.0+
• FLORIS 4.2.1+
• WInc3D (LES simulator)
• f90nml 1.4.4+
• seaborn 0.13.2+
• kaleido 0.2.1

WInc3D must be installed separately following instructions at WInc3D repository.

Usage

Optimization

Optimization campaigns are run through a command-line interface:

# Start a new single-fidelity campaign using only LES
python -m windopt.optim.cli new my-campaign \
    --strategy les-only \
    --n-turbines 4 \
    --les-batch-size 5 \
    --max-les-batches 20

# Start a new multi-fidelity campaign alternating between GCH and LES
python -m windopt.optim.cli new my-mf-campaign \
    --strategy multi-alternating \
    --n-turbines 4 \
    --les-batch-size 4 \
    --gch-batch-size 50 \
    --gch-batches-per-les 1 \
    --max-les-batches 20

# Resume an existing campaign
python -m windopt.optim.cli restart my-campaign

The main optimization workflow consists of:

1. Running precursor simulations to generate inflow conditions
2. Evaluating candidate layouts using both GCH and LES models
3. Fitting multi-task Gaussian processes to observations
4. Selecting new layouts via batch expected improvement

Visualization

The package includes visualization utilities for analyzing wind farm simulations and optimization results:

# Plot instantaneous and mean velocity fields, where the instantanous field is taken 1 hour after spinup
python -m windopt.viz.velocity_flow run_dir 1.0 --save_path flow.png

# Create animated visualization of velocity field evolution
python -m windopt.viz.vector_field run_dir \
    --field ux \
    --start 0 \
    --n_steps 100 \
    --save anim.html

Displayed: Instantaneous and mean streamwise velocity (in m/s) throughout a grid-layout wind farm.

The visualization tools allow you to:

• Examine velocity field development through wind farms
• Understand wake interactions between turbines
• Monitor statistical convergence of simulations
• Track optimization campaign performance

Limitations

• Currently supports optimization of up to 4 turbines as proof of concept
• Handles single wind direction/speed (extension to varying conditions planned)
• Uses fixed simulation durations (adaptive durations planned)
• Requires manual alternation between fidelities (automatic selection planned)

Acknowledgements

• Prof. Nikos Bempedelis (Queen Mary University of London) for guidance on LES configuration
• Prof. John Cunningham (Columbia University) for Bayesian optimization guidance
• Columbia University's Terremoto computing cluster for computational resources

Project Structure

.
├── campaigns/    # Optimization experiment results
├── config/       # Configuration files
├── data/         # Input data and initial samples
├── doc/          # Documentation and reports
├── src/
│   └── windopt/  # Main package
│       ├── config/     # Configuration handling
│       ├── initial/    # Initial sampling routines
│       ├── optim/      # Optimization implementations
│       ├── viz/        # Visualization utilities
│       └── winc3d/     # WInc3D LES interface