Energy Conversion and Management - Wind farm layout optimization based on support vector regression guided genetic algorithm with consideration of participation among landowners
[1] Xinglong Ju, Feng Liu, Li Wang, and Wei-Jen Lee. "Wind farm layout optimization based on support vector regression guided genetic algorithm with consideration of participation among landowners." Energy Conversion and Management 196 (2019): 1267-1281. 
If you have any questions, please email Feng Liu ([email protected]), or Xinglong Ju ([email protected]).
Jensen's wake model
import numpy as np
import pandas as pd
import WindFarmGenetic # wind farm layout optimization using genetic algorithms classes
from datetime import datetime
import os
from sklearn.svm import SVR
import pickle# parameters for the genetic algorithm
elite_rate = 0.2
cross_rate = 0.6
random_rate = 0.5
mutate_rate = 0.1
wt_N = 25 # number of wind turbines 15, 20, or 25
# NA_loc_array : not available location array, the index starting from 1
# L1 : wind farm, cells 121(inclusive) to 144(inclusive)
NA_loc_array = np.arange(121, 145, 1)
#
# # L2
# NA_loc_array = np.arange(61, 85, 1)
#
# # L3
# NA_loc_array = np.concatenate((np.arange(11, 144, 12), np.arange(12, 145, 12)))
#
# # L4
# NA_loc_array = np.concatenate((np.arange(6, 144, 12), np.arange(7, 145, 12)))
#
# # L5
# NA_loc_array = np.concatenate((np.arange(41, 105, 12), np.arange(42, 105, 12),
# np.arange(43, 105, 12),
# np.arange(44, 105, 12)))
#
# # L6
# NA_loc_array = np.concatenate((np.arange(1, 28, 12), np.arange(2, 28, 12),
# np.arange(12, 37, 12),
# np.arange(11, 37, 12),
# np.arange(109, 145, 12), np.arange(119, 145, 12),
# np.arange(110, 145, 12),
# np.arange(120, 145, 12),
# ))
#
# # L7
# NA_loc_array = np.arange(133, 145, 1)
#
# # L8
# NA_loc_array = np.arange(61, 73, 1)
#
# # L9
# NA_loc_array = np.arange(12, 145, 12)
#
# # L10
# NA_loc_array = np.arange(6, 145, 12)
#
# # L11
# NA_loc_array = np.concatenate((np.arange(42, 105, 12),
# np.arange(43, 105, 12)))
#
# # L12
# NA_loc_array = np.array((1, 2, 11, 12, 13, 24, 121, 132, 133, 134, 143, 144))
# convert numpy array to list, datatype convert
NA_loc = NA_loc_array.tolist()
# L0
# NA_loc = []
population_size = 120 # how many layouts in a population
iteration_times = 200 # how many iterations in a genetic algorithm run
n_inits = 100 # number of initial populations n_inits >= run_times
run_times = 100 # number of different initial populations
# wind farm size, cells
cols_cells = 12 # number of cells each row
rows_cells = 12 # number of cells each column
cell_width = 77.0 * 3 # unit : m
# all data will be save in data folder
data_folder = "data"
if not os.path.exists(data_folder):
os.makedirs(data_folder)# create an WindFarmGenetic object. Specify the number of rows and the number columns of the wind farm land. N is the number of wind turbines.
# NA_loc is the not available locations on the wind farm land. Landowners does not want to participate in the wind farm.
# pop_size: how many individuals in the population
# iteration: iteration times of the genetic algorithm
wfg = WindFarmGenetic.WindFarmGenetic(rows=rows_cells, cols=cols_cells, N=wt_N, NA_loc=NA_loc, pop_size=population_size,
iteration=iteration_times,cell_width=cell_width, elite_rate=elite_rate,
cross_rate=cross_rate, random_rate=random_rate, mutate_rate=mutate_rate)
# Specify the wind distribution
# wind distribution is discrete (number of wind speeds) by (number of wind directions)
# wfg.init_1_direction_1_N_speed_13()
# file name to store the wind power distribution SVR model
# svr_model_filename = 'svr_1s1d_N_13.svr'
# wfg.init_4_direction_1_speed_13()
# svr_model_filename = 'svr_1s4d_13.svr'
wfg.init_6_direction_1_speed_13()
# svr_model_filename = 'svr_1s6d_13.svr'# initial population saved folder
init_pops_data_folder = "{}/init_data".format(data_folder)
if not os.path.exists(init_pops_data_folder):
os.makedirs(init_pops_data_folder)
# generate initial populations to start with and store them
# in order to start from the same initial population for different methods
# so it is fair to compare the final results
for i in range(n_inits):
wfg.gen_init_pop_NA()
wfg.save_init_pop_NA("{}/init_{}.dat".format(init_pops_data_folder,i), "{}/init_{}_NA.dat".format(init_pops_data_folder,i))# results folder
# adaptive_best_layouts_N60_9_20190422213718.dat : best layout for AGA of run index 9
# result_CGA_20190422213715.dat : run time and best eta for CGA method
results_data_folder = "data/results"
if not os.path.exists(results_data_folder):
os.makedirs(results_data_folder)
# if cg,ag,sg folder does not exist, create these folders. Folders to store the running results
# cg: convertional genetic algorithm
# ag: adaptive genetic algorithm
# sg: support vector regression guided genetic algorithm
cg_result_folder = "{}/cg".format(results_data_folder)
if not os.path.exists(cg_result_folder):
os.makedirs(cg_result_folder)
ag_result_folder = "{}/ag".format(results_data_folder)
if not os.path.exists(ag_result_folder):
os.makedirs(ag_result_folder)
sg_result_folder = "{}/sg".format(results_data_folder)
if not os.path.exists(sg_result_folder):
os.makedirs(sg_result_folder)
# resul_arr: run_times by 2 , the first column is the run time in seconds for each run and the second column is the conversion efficiency for the run
result_arr = np.zeros((run_times, 2), dtype=np.float32)# CGA: Conventional genetic algorithm
for i in range(0, run_times): # run times
print("run times {} ...".format(i))
# load initial population
wfg.load_init_pop_NA("{}/init_{}.dat".format(init_pops_data_folder,i), "{}/init_{}_NA.dat".format(init_pops_data_folder,i))
# run the conventional genetic algorithm and return run time and conversion efficiency
run_time, eta = wfg.conventional_genetic_alg(i,result_folder=cg_result_folder)
result_arr[i, 0] = run_time
result_arr[i, 1] = eta
time_stamp = datetime.now().strftime("%Y%m%d%H%M%S")
# save the run time and etas to a file
filename = "{}/result_conventional_{}.dat".format(cg_result_folder,time_stamp)
np.savetxt(filename, result_arr, fmt='%f', delimiter=" ")# AGA: adaptive genetic algorithm
for i in range(0,run_times): # run times
print("run times {} ...".format(i))
wfg.load_init_pop_NA("{}/init_{}.dat".format(init_pops_data_folder,i),"{}/init_{}_NA.dat".format(init_pops_data_folder,i))
run_time, eta = wfg.adaptive_genetic_alg(i,result_folder=ag_result_folder)
result_arr[i, 0] = run_time
result_arr[i, 1] = eta
time_stamp = datetime.now().strftime("%Y%m%d%H%M%S")
filename = "{}/result_adaptive_{}.dat".format(ag_result_folder,time_stamp)
np.savetxt(filename, result_arr, fmt='%f', delimiter=" ")n_mc_samples = 10000 # svr train data, number of layouts to average
wds_data_folder = "{}/wds".format(data_folder)
if not os.path.exists(wds_data_folder):
os.makedirs(wds_data_folder)
# mc : monte-carlo
# number of layouts to generate as the training data for regression
# to build the power distribution surface
# mc_layout.dat file stores layouts only with 0s and 1s. 0 means no turbine here. 1 means one turbine here.
# mc_layout_NA.dat file stores layouts with 0s, 1s and 2s. 2 means no turbine and not available for turbine.
# These two files are used to generate wind power distribution.
# Each file has 10000 lines. Each line is layout.
# gen_mc_grid_with_NA_loc function generates these two files.
train_mc_layouts, train_mc_layouts_NA = WindFarmGenetic.LayoutGridMCGenerator.gen_mc_grid_with_NA_loc(rows_cells,
cols_cells,
n_mc_samples,
wt_N, NA_loc,
"{}/mc_layout.dat".format(wds_data_folder),
"{}/mc_layout_NA.dat".format(wds_data_folder))
# wfg.init_1_direction_1_N_speed_13()
# file name to store the wind power distribution SVR model
# svr_model_filename = 'svr_1s1d_N_13.svr'
# wfg.init_4_direction_1_speed_13()
# svr_model_filename = 'svr_1s4d_13.svr'
# wfg.init_6_direction_1_speed_13()
svr_model_filename = 'svr_1s6d_13.svr'
# load Monte-Carlo layouts from a text file. 10000 random layouts
layouts = np.genfromtxt("{}/mc_layout.dat".format(wds_data_folder), delimiter=" ", dtype=np.int32)
# generate the location index coordinate and average power output at each location index coordinate
# location index coordinate : in the cells, the cell with index 1 has location index (0,0) and the cell 2 has (1,0)
# store the location index coordinate in x.dat and average power in y.dat
wfg.mc_gen_xy_NA(rows=rows_cells, cols=cols_cells, layouts=layouts, n=n_mc_samples, N=wt_N, xfname="{}/x.dat".format(wds_data_folder),
yfname="{}/y.dat".format(wds_data_folder))
# read index location coordinates
x_original = pd.read_csv("{}/x.dat".format(wds_data_folder), header=None, nrows=rows_cells * cols_cells, delim_whitespace=True, dtype=np.float32)
x_original = x_original.values
# read the power output of each index location coordinate
y_original = pd.read_csv("{}/y.dat".format(wds_data_folder), header=None, nrows=rows_cells * cols_cells, delim_whitespace=True, dtype=np.float32)
y_original = y_original.values.flatten()
# create a SVR object and specify the kernal and other parameters
svr_model = SVR(kernel='rbf', C=2000.0, gamma=0.3, epsilon=.1)
# build the SVR power distribution model
svr_model.fit(x_original, y_original)
# save the SVR model to a file
pickle.dump(svr_model, open("{}/{}".format(wds_data_folder,svr_model_filename), 'wb'))
# This is how to load SVR model from a file
# svr_model = pickle.load(open("{}/{}".format(wds_data_folder,svr_model_filename), 'rb'))#SUGGA: support vector regression guided genetic algorithm
for i in range(0, run_times): # run times
print("run times {} ...".format(i))
wfg.load_init_pop_NA("{}/init_{}.dat".format(init_pops_data_folder,i),"{}/init_{}_NA.dat".format(init_pops_data_folder,i))
run_time, eta = wfg.sugga_genetic_alg(i,svr_model=svr_model,result_folder=sg_result_folder)
result_arr[i, 0] = run_time
result_arr[i, 1] = eta
time_stamp = datetime.now().strftime("%Y%m%d%H%M%S")
filename = "{}/result_sugga_{}.dat".format(sg_result_folder,time_stamp)
np.savetxt(filename, result_arr, fmt='%f', delimiter=" ")