max_SMS_SoC=12 / 3,
max_LION_SoC=54,
# Period length in minutes:
PERIODEN_DAUER=period_min,
# Q-Table can only take discrete inputs and make dicrete outputs
ACTION_TYPE='discrete',
OBS_TYPE='discrete',
# Set number of discrete values:
discrete_space=22,
# Size of validation data:
val_split=0.1)
# Setup agent:
agent = Q_Learner(
env=env,
memory_len=update_num,
# Training parameter:
gamma=0.85,
epsilon=0.8,
epsilon_min=0.1,
epsilon_decay=0.999996,
lr=0.5,
tau=0.125)
# Train:
training(agent, epochs, update_num, num_warmup_steps)
# Test with dataset that includes val-data:
env.use_all_data()
testing(agent)
def run_agent(name='',gamma=.9, lr=0.1, tau=0.15, update_num=500,
epsilon_decay='linear', input_list=['norm_total_power','normal','seq_max'],
hidden_size=256, pre_trained_model=None, target_update_num=None):
'''
Trains and tests a DQN based on the passed parameter.
'''
# Naming the agent:
now = datetime.now()
NAME = 'agent_DQN_'+name+'_t-stamp'+now.strftime("_%d-%m-%Y_%H-%M-%S")
# Import dataset and logger based from the common settings
df, power_dem_df, logger, period_min = dataset_and_logger(NAME)
# Number of warm-up steps:
num_warmup_steps = 100
# Number of epochs and steps:
epochs = 100
# Setup reward_maker
r_maker = reward_maker(
LOGGER = logger,
# Settings:
COST_TYPE = 'exact_costs',
R_TYPE = 'savings_focus',
R_HORIZON = 'single_step',
# Parameter to calculate costs:
cost_per_kwh = 0.2255,
LION_Anschaffungs_Preis = 34100,
LION_max_Ladezyklen = 6000,
SMS_Anschaffungs_Preis = 55000,#115000/3,
SMS_max_Nutzungsjahre = 25,
Leistungspreis = 102,)
# Setup common_env
env = common_env(
reward_maker = r_maker,
df = df,
power_dem_df = power_dem_df,
# Datset Inputs for the states:
input_list = input_list,
# Batters stats:
max_SMS_SoC = 25,
max_LION_SoC = 54,
LION_max_entladung = 50,
SMS_max_entladung = 100,
SMS_entladerate = 0.72,
LION_entladerate = 0.00008,
# Period length in minutes:
PERIODEN_DAUER = period_min,
# DQN inputs can be conti and outputs must be discrete:
ACTION_TYPE = 'discrete',
OBS_TYPE = 'contin',
# Set number of discrete values:
discrete_space = 22,
# Size of validation data:
val_split = 0.1)
# Setup Agent:
agent = DQN(
env = env,
memory_len = update_num,
# Training parameter:
gamma = gamma,
epsilon = 0.99,
epsilon_min = 0.1,
epsilon_decay = epsilon_decay,
lr = lr,
tau = tau,
activation = 'relu',
loss = 'mean_squared_error',
hidden_size = hidden_size,
pre_trained_model=pre_trained_model,
target_update_num=target_update_num)
# Train:
training(agent, epochs, update_num, num_warmup_steps)
# Test with dataset that includes val-data:
env.use_all_data()
testing(agent)
def run_agent(name='',gamma=.9, lr=0.1, update_num=100, load_table=None,
epsilon_decay='linear', input_list=['norm_total_power','normal','seq_max']):
# Naming the agent:
now = datetime.now()
NAME = 'agent_Q-Table_'+name+'_t-stamp'+now.strftime("_%d-%m-%Y_%H-%M-%S")
# Import dataset and logger based on the common settings
df, power_dem_df, logger, period_min = dataset_and_logger(NAME)
# Number of warm-up steps:
num_warmup_steps = 100
# Number of epochs and steps:
epochs = 100
# Setup reward_maker
r_maker = reward_maker(
LOGGER = logger,
# Settings:
COST_TYPE = 'exact_costs',
R_TYPE = 'savings_focus',
R_HORIZON = 'single_step',
# Parameter to calculate costs:
cost_per_kwh = 0.2255,
LION_Anschaffungs_Preis = 34100,
LION_max_Ladezyklen = 6000,
SMS_Anschaffungs_Preis = 55000,#115000/3,
SMS_max_Nutzungsjahre = 25,
Leistungspreis = 102,)
# Setup common_env
env = common_env(
reward_maker = r_maker,
df = df,
power_dem_df = power_dem_df,
# Datset Inputs for the states:
input_list = input_list,
# Batters stats:
max_SMS_SoC = 25,
max_LION_SoC = 54,
LION_max_entladung = 50,
SMS_max_entladung = 100,
SMS_entladerate = 0.72,
LION_entladerate = 0.00008,
# Period length in minutes:
PERIODEN_DAUER = period_min,
# Q-Table can only take discrete inputs and make discrete outputs
ACTION_TYPE = 'discrete',
OBS_TYPE = 'discrete',
# Set number of discrete values:
discrete_space = 22,
# Size of validation data:
val_split = 0.1)
# Setup agent:
agent = Q_Learner(
env = env,
memory_len = update_num,
# Training parameter:
gamma = gamma,
epsilon = 0.99,
epsilon_min = 0.1,
epsilon_decay = epsilon_decay,
lr = lr,
load_table = load_table)
# Train:
training(agent, epochs, update_num, num_warmup_steps)
# Test with dataset that includes val-data:
env.use_all_data()
testing(agent)