Skip to contents

Run a Genetic Algorithm to optimize a wind farm layout

Source: R/genetic_algorithm.R
genetic_algorithm.Rd

Run a Genetic Algorithm to optimize the layout of wind turbines on a given area. The algorithm works with a fixed amount of turbines, a fixed rotor radius and a mean wind speed value for every incoming wind direction.

Usage

genetic_algorithm(
  area,
  wind,
  n,
  rotor,
  rotor_height,
  grid_method = "rectangular",
  fcr = 5,
  reference_height = rotor_height,
  surface_roughness = 0.3,
  proportionality = 1,
  iteration = 20,
  mutation_rate = NULL,
  terrain = FALSE,
  elitism = TRUE,
  n_elite = 3,
  selection_mode = "VAR",
  crs = NULL,
  ccl = NULL,
  ccl_roughness = NULL,
  weibull = FALSE,
  weibull_src = NULL,
  parallel = FALSE,
  n_cluster = 2,
  verbose = FALSE,
  plot = FALSE,
  on_generation = NULL
)

Arguments

area

Site polygon (sf, SpatialPolygons, or coordinate matrix). Must be projected in metres.

wind

Wind data.frame with ws, wd and optional probab. See windata_format().

n

Number of turbines (fixed; every individual has n unique cell IDs).

rotor

Rotor radius in metres.

rotor_height

Hub height in metres.

grid_method

"rectangular" or "h" / "hexagon".

fcr

Grid spacing factor. Cell size is fcr * rotor.

reference_height

Height at which wind$ws was measured.

surface_roughness

Roughness length in metres. Per-cell when terrain is on.

proportionality

Minimum fraction of a grid cell that must overlap the site (prop in grid_area()).

iteration

Generation budget.

mutation_rate

Swap probability per turbine. NULL means 2/n.

terrain

Terrain model (elevation + land cover). TRUE downloads a DEM via elevatr. Pass a DEM raster to skip the download. Per-cell values are computed once and stored in the result as terrainModel for plot_result() / random_search().

elitism

Archive the best layout and breed elite children.

n_elite

Base elite count (grows/shrinks with search phase).

selection_mode

"VAR" (parent share follows fitness) or "FIX" (50 %).

crs

CRS if area has none (EPSG code or PROJ string).

ccl

Path to a Corine Land Cover raster when terrain is on.

ccl_roughness

Path to the CLC legend CSV (Rauhigkeit_z column).

weibull

If TRUE, hub-height speed comes from Weibull rasters; wind$ws is ignored.

weibull_src

list(k, a) shape and scale rasters (e.g. Global Wind Atlas combined-Weibull-k / combined-Weibull-A).

parallel

Parallel fitness (parallel + doParallel).

n_cluster

Worker count when parallel is TRUE.

verbose

Print a line per generation.

plot

Plot the current best layout each generation.

on_generation

Optional callback after each generation: function(generation, iteration, energy, efficiency, fitness). Used by Shiny for progress. Errors in the callback are ignored.

Value

The result is a matrix with aggregated values per generation; the best individual regarding energy and efficiency per generation, some fuzzy control variables per generation, a list of all fitness values per generation, the amount of individuals after each process, a matrix of all energy, efficiency and fitness values per generation, the selection and crossover parameters, a matrix with the generational difference in maximum and mean energy output, a matrix with the given inputs, a dataframe with the wind information, the mutation rate per generation and a matrix with all tested wind farm layouts.

Details

A terrain effect model can be included in the optimization process. Therefore, a digital elevation model will be downloaded automatically via the elevatr::get_elev_raster function. A land cover raster can also downloaded automatically from the EEA-website, or the path to a raster file can be passed to ccl. The algorithm uses an adapted version of the Raster legend ("clc_legend.csv"), which is stored in the package directory ~/inst/extdata. To use other values for the land cover roughness lengths, insert a column named "Rauhigkeit_z" to the .csv file, assign a surface roughness length to all land cover types. Be sure that all rows are filled with numeric values and save the file with ";" separation. Assign the path of the file to the input variable ccl_roughness of this function.

Fitness is \(EnergyOverall \times (EfficAllDir/100)^w\) with w = getOption("windfarmGA.fitness_efficiency_weight"). Hub-height wind speeds use a logarithmic profile unless options(windfarmGA.wind_profile = "power") restores the legacy power law. Power uses options(windfarmGA.Cp) (default 0.45) and optional cut-in / rated / cut-out speeds. Layouts are encoded as n unique grid-cell IDs (set crossover and swap mutation). Selection defaults to VAR (percentage follows fitness progress). Mutation, immigrants and unused-cell injection prefer rarely visited cells. Crossover is spatial with probability options(windfarmGA.spatial_crossover) (default 0.5). Evaluated layouts are cached. A flat global max is not a stop signal. The run ends at iteration, or earlier only after options(windfarmGA.stall_generations) consecutive generations with no new layout, no newly visited cell and no new best fitness (set to 0 to disable). Operator rates cycle like seasons, still only selection / set-crossover / swap-mutation: explore (rates rise on stall) until options(windfarmGA.refine_min_gen) (default 18) and options(windfarmGA.refine_after) (default 12) generations without a new max at coverage \(\ge 0.35\); then refine (inject toward 0.15, mutation toward 2/n, selection toward about 45\ options(windfarmGA.refine_hold) generations in refine (default 25), a short disturbance pulse of options(windfarmGA.explore_pulse) generations (default 10) raises the same rates even if new maxes are still trickling in, then refine resumes. Elites get a short local search each generation: one turbine slides to a neighbouring empty cell (not a random cell anywhere on the grid).

See also

Other Genetic Algorithm Functions: crossover(), fitness(), init_population(), mutation(), selection(), set_crossover(), swap_mutation(), trimton()

Examples

if (FALSE) { # \dontrun{
## Create a random rectangular shapefile
library(sf)

area <- sf::st_as_sf(sf::st_sfc(
  sf::st_polygon(list(cbind(
    c(4498482, 4498482, 4499991, 4499991, 4498482),
    c(2668272, 2669343, 2669343, 2668272, 2668272)
  ))),
  crs = 3035
))

## Create a uniform and unidirectional wind data.frame and plot the
## resulting wind rose
data.in <- data.frame(ws = 12, wd = 0)
windrosePlot <- plot_windrose(
  data = data.in, spd = data.in$ws,
  dir = data.in$wd, dirres = 10, spdmax = 20
)

## Runs an optimization run for 20 iterations with the
## given shapefile (area), the wind data.frame (data.in),
## 12 turbines (n) with rotor radii of 30m and hub height of 100m.
result <- genetic_algorithm(
  area = area,
  n = 12,
  wind = data.in,
  rotor = 30,
  rotor_height = 100
)
plot_windfarmGA(result = result, area = area)
} # }