FLORIS v3 Roadmap

[rafmudaf]

Methods and objectives

The primary objective for the next major version of FLORIS is to enable order of magnitude performance improvements for wind farm simulations over many atmospheric conditions by redesigning the underlying data structures to leverage CPU vectorization (SIMD) and support further speedups and algorithmic development in the future.

Parity with v2

• Visualization
  • @RHammond2
  • See V3/turbine grid visualization #9 for an initial implementation
  • Some functionality depends on full flow field grid implementation
• Port the FLORIS tools package
  • @rafmudaf
• Support heterogeneous wind definition
  • @rafmudaf
• Models
  • @bayc
  • Wake deficit
  • Wake deflection
  • Turbulence
• Solver instances
  • Sequential with turbine grid
  • Sequential with full flow field grid
    • @rafmudaf
• Documentation
  • Code style
  • API documentation
  • Usage from a developer's point of view

Not in FLORIS v2, but work in progress for v3

• Plane solve - AKA Tony's @bayc
• YAML input file @RHammond2 @cb
  • Wind rose input
  • Include ontology
    • Multiple turbine definitions
• Multi wind direction vectorization - wind rose calculation
• Solver
  • Iterative full-farm
• Cython implementation of one simple model

Other areas of interest

Improvements to the input file structure

How to specify wind speed and wind direction combinations

• Wind rose input; what do the following inputs mean?
• [ 270, 360 ] # wind directions
• [ 10.0, 8.0 ] # wind speeds

Data Structure

The data is structured entirely in Numpy arrays where some helper classes and functions are available for accessing the arrays in an intuitive form. The principal arrays are

• x, y, z: These hold the coordinates of the grid-type. They are dimensioned as (n wind directions, n wind speeds, n turbines, n grid points, m grid points) for the TurbineGrid
• u, v, w: These are the values of the velocity components at each grid point. Therefore, they have shape (n wind directions, n wind speeds, n turbines, n grid points, m grid points). After the first two dimensions, this is the same shape as the grid arrays (x,y,z).

The first place where data structures start to take form is in the Grid classes. Here, we create the x, y, z arrays that establish the shape (dimensions) and size of the grid. Then, FlowField.initialize_velocity_field uses the Grid's dimensions to size the velocity arrays. It also adds on dimensions for the wind speeds and wind directions.

Other considerations

Support for analytic gradients - @bayc
Automatic differentiation - @bayc

Interesting ideas

Julia backend
Other language implementation of particular models (C++, Julia, Cython)

Questions for later

The thrust curve we use for NREL 5MW has lowest wind speed of 2 m/s and the Ct is greater than 1. Is this correct? If so, we also bound Ct to 0-1 in the code, so is that correct? What does a Ct > 1 mean?