Parallel yaw optimization under uncertainty

[MiguelMarante]

Hello everyone,

I am trying to run optimization in parallel under uncertainty. However, I have noticed that the output differs significantly when running in parallel. It appears that the optimization is performed independently for each sampled wind direction.

Below is a simple example using a wind direction standard deviation of 3. When running the optimization in series, the results show the optimal yaw set-points for the mean wind direction, as expected. However, when running in parallel, these set-points are provided for each of the sampled wind directions ([264., 267., 270., 273., 276.]).

I wonder if this is a problem in my implementation or in the coupling done between the Uncertain and Parallel classes in Floris. Is there any workaround for this problem?

Best regards,
Miguel Marante

CODE:

import numpy as np
from floris import WindRose, UncertainFlorisModel, ParallelFlorisModel
from floris.optimization.yaw_optimization.yaw_optimizer_sr import YawOptimizationSR

# Input model config
input_file = 'floris/examples/inputs/gch.yaml'

# Input climate data
ti = 0.06
ws = np.array([6., 8., 10.])
wd = np.array([270.])
wind_rose = WindRose(
    wind_speeds=ws,
    wind_directions=wd,
    ti_table=ti
    )

# Create floris object
ufmodel = UncertainFlorisModel(input_file, wd_std=3.0)
ufmodel.set(wind_data=wind_rose)

# Run optimization
yaw_opt = YawOptimizationSR(ufmodel)
df_opt_series = yaw_opt.optimize()
print(df_opt_series)


pfmodel = ParallelFlorisModel(ufmodel, max_workers=3, n_wind_condition_splits=3)
df_opt_parallel = pfmodel.optimize_yaw_angles()
print(df_opt_parallel)

[misi9170]

Hi @MiguelMarante ,

Thanks for your question! We have actually very recently introduced a new parallel option, the ParFlorisModel, in #982 . This is currently only available on FLORIS' develop branch, but it will be available in the next week or so on the main branch also (we are planning a new minor version release, 4.2, which we are just putting the finishing touches on---see #994).

To avoid breaking changes, we will not remove the existing ParallelFlorisModel---however, we have found that there is not good compatibility between UncertainFlorisModel and ParallelFlorisModel (which we have improved now with the new ParFlorisModel). To avoid confusion, in v4.2 you won't be allowed to pass the old-style ParallelFlorisModel into UncertainFlorisModel any longer, because it doesn't work properly anyway.

Note that the new ParFlorisModel does not have it's own optimize_yaw_angles() method. Rather, an instantiated ParFlorisModel object should be passed directly to YawOptimizerSR, in exactly the same way as an UncertainFlorisModel would be. In your case, the workflow would be:

1. Instantiate a ParFlorisModel (e.g., named pfmodel)
2. Instantiate an UncertainFlorisModel (e.g., named ufmodel), passing in pfmodel to the configuration argument
3. Instantiate a yaw optimizer object yaw_opt = YawOptimizationSR(ufmodel) and proceed to optimize.

Hopefully that makes sense---and sorry to switch this up on you---we hope that you'll find that the new ParFlorisModel is more directly interchangeable with FlorisModel than the previous ParallelFlorisModel!

Misha

[paulf81]

Just to add on, using the newer syntax this is working properly I think:

import numpy as np
import pandas as pd
from floris import WindRose, UncertainFlorisModel, ParFlorisModel
from floris.optimization.yaw_optimization.yaw_optimizer_sr import YawOptimizationSR

# Included for parallelization
if __name__ == "__main__":

    # Input model config
    input_file = 'inputs/gch.yaml'

    # Input climate data
    ti = 0.06
    ws = np.array([6., 8., 10.])
    wd = np.array([270.])
    wind_rose = WindRose(
        wind_speeds=ws,
        wind_directions=wd,
        ti_table=ti
        )

    # Create uncertain floris object
    ufmodel = UncertainFlorisModel(input_file, wd_std=3.0)
    ufmodel.set(wind_data=wind_rose)

    # Create parallelelized uncertain floris object
    pfmodel = ParFlorisModel(input_file, max_workers=3, n_wind_condition_splits=3)
    upfmodel = UncertainFlorisModel(pfmodel, wd_std=3.0)

    # Run optimization
    yaw_opt = YawOptimizationSR(ufmodel)
    df_opt_series = yaw_opt.optimize()

    # Run optimization with parallelization
    yaw_opt_par = YawOptimizationSR(upfmodel)
    df_opt_series_par = yaw_opt.optimize()

    # Print results
    print(df_opt_series)
    print('Parallelized results:')
    print(df_opt_series_par)

[MiguelMarante]

Hello. Thank you, I am glad to hear that you already fixed this. Perfect timing :)