ctrl.append(MultiPhaseDQCurrentController(current_dqp_iparams, pll_params, i_lim, droop_param_slave,
qdroop_param_slave, ts_sim=delta_t, name='slave'))
#####################################
# Definition of the optimization agent
# The agent is using the SafeOpt algorithm by F. Berkenkamp (https://arxiv.org/abs/1509.01066) in this example
# Arguments described above
# History is used to store results
agent = SafeOptAgent(mutable_params,
abort_reward,
j_min,
kernel,
dict(bounds=bounds, noise_var=noise_var, prior_mean=prior_mean,
safe_threshold=safe_threshold, explore_threshold=explore_threshold),
ctrl,
{'master': [[f'lc1.inductor{k}.i' for k in '123'],
[f'lc1.capacitor{k}.v' for k in '123'],
],
'slave': [[f'lcl1.inductor{k}.i' for k in '123'],
[f'lcl1.capacitor{k}.v' for k in '123'],
np.zeros(3)]},
history=FullHistory()
)
#####################################
# Definition of the environment using a FMU created by OpenModelica
# (https://www.openmodelica.org/)
# Using two inverters supplying a load
# - using the reward function described above as callable in the env
# - viz_cols used to choose which measurement values should be displayed
def run_experiment(len_kp, len_ki):
if isfile(f'{save_folder}/{len_kp:.4f},{len_ki:.4f}.txt'):
with open(f'{save_folder}/{len_kp:.4f},{len_ki:.4f}.txt', 'r') as f:
return strtobool(f.read().strip())
rew = Reward(i_limit=iLimit,
i_nominal=iNominal,
mu_c=mu,
max_episode_steps=max_episode_steps,
obs_dict=[[f'lc.inductor{k}.i' for k in '123'],
'master.phase', [f'master.SPI{k}' for k in 'dq0']])
#####################################
# Definitions for the GP
prior_mean = 0 # 2 # mean factor of the GP prior mean which is multiplied with the first performance of the
# initial set
noise_var = 0.001 # 0.001 ** 2 # measurement noise sigma_omega
prior_var = 2 # prior variance of the GP
bounds = None
lengthscale = None
if adjust == 'Kp':
bounds = [(0.0001, 0.1)] # bounds on the input variable Kp
lengthscale = [
.025
] # length scale for the parameter variation [Kp] for the GP
# For 1D example, if Ki should be adjusted
if adjust == 'Ki':
bounds = [(0, 20)] # bounds on the input variable Ki
lengthscale = [
10
] # length scale for the parameter variation [Ki] for the GP
# For 2D example, choose Kp and Ki as mutable parameters (below) and define bounds and lengthscale for both of them
if adjust == 'Kpi':
bounds = [(0.001, 0.07), (2, 150)]
lengthscale = [0.012, 30.]
df_len = pd.DataFrame({
'lengthscale': lengthscale,
'bounds': bounds,
'balanced_load': balanced_load,
'barrier_param_mu': mu
})
# The performance should not drop below the safe threshold, which is defined by the factor safe_threshold times
# the initial performance: safe_threshold = 0.8 means. Performance measurement for optimization are seen as
# unsafe, if the new measured performance drops below 20 % of the initial performance of the initial safe (!)
# parameter set
safe_threshold = 0
j_min = cal_j_min(phase_shift, amp_dev) # Used for normalization
# The algorithm will not try to expand any points that are below this threshold. This makes the algorithm stop
# expanding points eventually.
# The following variable is multiplied with the first performance of the initial set by the factor below:
explore_threshold = 0
# Factor to multiply with the initial reward to give back an abort_reward-times higher negative reward in case of
# limit exceeded
# has to be negative due to normalized performance (regarding J_init = 1)
abort_reward = 100 * j_min
# Definition of the kernel
kernel = GPy.kern.Matern32(input_dim=len(bounds),
variance=prior_var,
lengthscale=lengthscale,
ARD=True)
#####################################
# Definition of the controllers
mutable_params = None
current_dqp_iparams = None
if adjust == 'Kp':
# mutable_params = parameter (Kp gain of the current controller of the inverter) to be optimized using
# the SafeOpt algorithm
mutable_params = dict(currentP=MutableFloat(0.04))
# Define the PI parameters for the current controller of the inverter
current_dqp_iparams = PI_params(kP=mutable_params['currentP'],
kI=12,
limits=(-1, 1))
# For 1D example, if Ki should be adjusted
elif adjust == 'Ki':
mutable_params = dict(currentI=MutableFloat(5))
current_dqp_iparams = PI_params(kP=0.005,
kI=mutable_params['currentI'],
limits=(-1, 1))
# For 2D example, choose Kp and Ki as mutable parameters
elif adjust == 'Kpi':
mutable_params = dict(currentP=MutableFloat(0.04),
currentI=MutableFloat(11.8))
current_dqp_iparams = PI_params(kP=mutable_params['currentP'],
kI=mutable_params['currentI'],
limits=(-1, 1))
# Define a current sourcing inverter as master inverter using the pi and droop parameters from above
ctrl = MultiPhaseDQCurrentSourcingController(current_dqp_iparams,
delta_t,
undersampling=undersample,
name='master',
f_nom=net.freq_nom)
i_ref = MutableParams([MutableFloat(f) for f in i_ref1])
#####################################
# Definition of the optimization agent
# The agent is using the SafeOpt algorithm by F. Berkenkamp (https://arxiv.org/abs/1509.01066) in this example
# Arguments described above
# History is used to store results
agent = SafeOptAgent(
mutable_params,
abort_reward,
j_min,
kernel,
dict(bounds=bounds,
noise_var=noise_var,
prior_mean=prior_mean,
safe_threshold=safe_threshold,
explore_threshold=explore_threshold),
[ctrl],
dict(master=[[f'lc.inductor{k}.i' for k in '123'], i_ref]),
history=FullHistory(),
)
#####################################
# Definition of the environment using a FMU created by OpenModelica
# (https://www.openmodelica.org/)
# Using an inverter supplying a load
# - using the reward function described above as callable in the env
# - viz_cols used to choose which measurement values should be displayed (here, only the 3 currents across the
# inductors of the inverters are plotted. Labels and grid is adjusted using the PlotTmpl (For more information,
# see UserGuide)
# - inputs to the models are the connection points to the inverters (see user guide for more details)
# - model outputs are the the 3 currents through the inductors and the 3 voltages across the capacitors
if include_simulate:
# Defining unbalanced loads sampling from Gaussian distribution with sdt = 0.2*mean
# r_load = Load(R, 0.1 * R, balanced=balanced_load, tolerance=0.1)
# l_load = Load(L, 0.1 * L, balanced=balanced_load, tolerance=0.1)
# i_noise = Noise([0, 0, 0], [0.0023, 0.0015, 0.0018], 0.0005, 0.32)
# if no noise should be included:
r_load = Load(R, 0 * R, balanced=balanced_load)
l_load = Load(L, 0 * L, balanced=balanced_load)
def reset_loads():
r_load.reset()
l_load.reset()
plotter = PlotManager(agent,
save_results=save_results,
save_folder=save_folder,
show_plots=show_plots)
def ugly_foo(t):
if t >= .05:
i_ref[:] = i_ref2
else:
i_ref[:] = i_ref1
return partial(l_load.give_value, n=2)(t)
env = gym.make(
'openmodelica_microgrid_gym:ModelicaEnv_test-v1',
# reward_fun=Reward().rew_fun,
reward_fun=rew.rew_fun_c,
# time_step=delta_t,
viz_cols=[
PlotTmpl([[f'lc.inductor{i}.i'
for i in '123'], [f'master.SPI{i}' for i in 'abc']],
callback=plotter.xylables_i_abc,
color=[['b', 'r', 'g'], ['b', 'r', 'g']],
style=[[None], ['--']]),
PlotTmpl([[f'master.m{i}' for i in 'abc']],
callback=lambda fig: plotter.update_axes(
fig,
title='Simulation',
ylabel='$m_{\mathrm{abc}}\,/\,\mathrm{}$')),
PlotTmpl([[f'master.CVI{i}'
for i in 'dq0'], [f'master.SPI{i}' for i in 'dq0']],
callback=plotter.xylables_i_dq0,
color=[['b', 'r', 'g'], ['b', 'r', 'g']],
style=[[None], ['--']])
],
log_level=logging.INFO,
viz_mode='episode',
max_episode_steps=max_episode_steps,
model_params={
'lc.resistor1.R': partial(r_load.give_value, n=0),
'lc.resistor2.R': partial(r_load.give_value, n=1),
'lc.resistor3.R': partial(r_load.give_value, n=2),
'lc.inductor1.L': partial(l_load.give_value, n=0),
'lc.inductor2.L': partial(l_load.give_value, n=1),
'lc.inductor3.L': ugly_foo
},
model_path='../../omg_grid/grid.paper.fmu',
# model_path='../omg_grid/omg_grid.Grids.Paper_SC.fmu',
net=net,
history=FullHistory(),
action_time_delay=1 * undersample)
runner = MonteCarloRunner(agent, env)
runner.run(num_episodes,
n_mc=n_MC,
visualise=True,
prepare_mc_experiment=reset_loads)
with open(f'{save_folder}/{len_kp:.4f},{len_ki:.4f}.txt', 'w') as f:
print(f'{agent.unsafe}', file=f)
return agent.unsafe
qdroop_param,
undersampling=undersample,
name='master')
#####################################
# Definition of the optimization agent
# The agent is using the SafeOpt algorithm by F. Berkenkamp (https://arxiv.org/abs/1509.01066) in this example
# Arguments described above
# History is used to store results
agent = SafeOptAgent(
mutable_params,
abort_reward,
j_min,
kernel,
dict(bounds=bounds,
noise_var=noise_var,
prior_mean=prior_mean,
safe_threshold=safe_threshold,
explore_threshold=explore_threshold), [ctrl],
dict(master=[[f'lc.inductor{k}.i'
for k in '123'], [f'lc.capacitor{k}.v' for k in '123']]),
history=FullHistory())
if include_simulate:
#####################################
# Definition of the environment using a FMU created by OpenModelica
# (https://www.openmodelica.org/)
# Using an inverter supplying a load
# - using the reward function described above as callable in the env
# - viz_cols used to choose which measurement values should be displayed.
# Labels and grid is adjusted using the PlotTmpl (For more information, see UserGuide)
qdroop_param,
undersampling=2)
#####################################
# Definition of the optimization agent
# The agent is using the SafeOpt algorithm by F. Berkenkamp (https://arxiv.org/abs/1509.01066) in this example
# Arguments described above
# History is used to store results
agent = SafeOptAgent(
mutable_params,
abort_reward,
kernel,
dict(bounds=bounds,
noise_var=noise_var,
prior_mean=prior_mean,
safe_threshold=safe_threshold,
explore_threshold=explore_threshold),
ctrl,
dict(master=[
np.array([f'lc1.inductor{i + 1}.i' for i in range(3)]),
np.array([f'lc1.capacitor{i + 1}.v' for i in range(3)]), i_ref
], ),
history=FullHistory())
#####################################
# Definition of the environment using a FMU created by OpenModelica
# (https://www.openmodelica.org/)
# Using an inverter supplying a load
# - using the reward function described above as callable in the env
# - viz_cols used to choose which measurement values should be displayed (here, only the 3 currents across the
# inductors of the inverters are plotted. Alternatively, the grid frequency and the currents transformed by