def swmm_states(Network, state):
temp = []
for i in Network:
temp.append(swmm.get(i, state, swmm.SI))
temp = np.asarray(temp)
temp = np.reshape(temp, (1, len(temp)))
return temp
def swmm_track(pond, attributes=["depth", "inflow", "outflow", "flooding"], controlled=False):
att_commands = {'depth': swmm.DEPTH,
'inflow': swmm.INFLOW,
'outflow': swmm.FLOW,
'flooding': swmm.FLOODING}
temp = []
for i in attributes:
if i == 'outflow':
if controlled:
temp.append(swmm.get(pond.orifice_id, att_commands[i], swmm.SI))
else:
temp.append(0.0)
else:
temp.append(swmm.get(pond.pond_id, att_commands[i], swmm.SI))
temp = np.asarray(temp)
return temp
# Simulation Tracker
reward_sim = []
outflow_sim = []
print "episode number :", episode_tracker
print "exploration :", epsilon_value[episode_tracker]
while episode_timer < time_sim:
t_epsi += 0
episode_timer += 1
# Look at whats happening
# Heights
temp_height = np.asarray(
[swmm.get(i, swmm.DEPTH, swmm.SI) for i in nodes_list])
# Gate Positions
temp_gate = np.asarray(gates.current_gate)
# Input States
input_states = np.append(
temp_height, temp_gate).reshape(
1, len(temp_height) + len(temp_gate))
# Action
q_values = prof_x.ac_model.predict_on_batch(input_states)
# Policy
action = epsi_greedy(len(action_space), q_values,
epsilon_value[episode_tracker])
agents_dqn[i].state_vector = swmm_states(states_controlled[i],
swmm.DEPTH)
# Take action
for i in nodes_controlled.keys():
action_step = agents_dqn[i].actions_q(epsilon_value[time_sim],
action_space)
agents_dqn[i].action_vector = action_step / 100.0
swmm.modify_setting(controlled_ponds[i].orifice_id,
agents_dqn[i].action_vector)
current_gate = agents_dqn[i].action_vector
# SWMM step
swmm.run_step()
# Receive the new rewards
outflow = swmm.get('ZOF1', swmm.INFLOW, swmm.SI)
water_level = swmm.get('93-50077', swmm.DEPTH, swmm.SI)
outflow_track.append(outflow)
overflows = swmm_states(all_nodes, swmm.FLOODING)
gate_change = np.abs(current_gate - temp_gate)
temp_gate = current_gate
r_temp = reward_function(water_level, outflow, gate_change)
for i in nodes_controlled.keys():
agents_dqn[i].rewards_vector = r_temp
# Observe the new states
for i in nodes_controlled.keys():
agents_dqn[i].state_new_vector = swmm_states(
states_controlled[i], swmm.DEPTH)
# Update Replay Memory
# Simulation Tracker
reward_sim = []
outflow_sim = []
print "episode number :", episode_tracker
print "exploration :", epsilon_value[episode_tracker]
while episode_timer < time_sim:
t_epsi += 1
episode_timer += 1
# Look at whats happening
# Heights
temp_height = np.asarray(
[swmm.get(i, swmm.DEPTH, swmm.SI) for i in nodes_list])
# Gate Positions
temp_gate = np.asarray(gates.current_gate)
# Input States
input_states = np.append(temp_height, temp_gate).reshape(
1,
len(temp_height) + len(temp_gate))
# Action
q_values = prof_x.ac_model.predict_on_batch(input_states)
# Policy
action = epsi_greedy(len(action_space), q_values,
epsilon_value[episode_tracker])
swmm.initialize(inp)
# Simulation Tracker
reward_sim = []
outflow_sim = []
print "episode number :", episode_tracker
print "exploration :", epsilon_value[episode_tracker]
while episode_timer < time_sim:
t_epsi += 1
episode_timer += 1
# print(episode_timer)
# Look at whats happening
# Heights
temp_height = np.asarray(
[swmm.get(i, swmm.DEPTH, swmm.SI) for i in con_ponds])
# Gate Positions
temp_gate = np.asarray(gates.current_gate)
if episode_timer % 5000 == 0:
print(episode_timer, swmm.get())
# Input States
input_states = np.append(temp_height, temp_gate).reshape(
1,
len(temp_height) + len(temp_gate))
# Action
q_values = prof_x.ac_model.predict_on_batch(input_states)
# Policy
action = epsi_greedy(len(action_space), q_values,
# ***********************************************************************
# Step Running
# ***********************************************************************
# Main loop: finished when the simulation time is over.
while( not swmm.is_over() ):
# ----------------- Run step and retrieve simulation time -----------
time.append( swmm.get_time() )
swmm.run_step() # Step 2
# --------- Retrieve & modify information during simulation ---------
# Retrieve information about flow in C-5
f = swmm.get('C-5', swmm.FLOW, swmm.SI)
# Stores the information in the flow vector
flow.append(f)
# Retrieve information about volume in V-1
v = swmm.get('V-1', swmm.VOLUME, swmm.SI)
# Stores the information in the volume vector
vol.append(v)
# --------------------------- Control Actions ------------------------
# If the flow in C-5 is greater or equal than 2 m3/s the setting
# upstream of the link is completely closed, else it is completely
# opened.
if f >= 2:
swmm.modify_setting('R-4', 0)
swmm.DEPTH)
# Take action
for i in nodes_controlled.keys():
action_step = agents_dqn[i].actions_q(epsilon_value[time_sim],
action_space)
agents_dqn[i].action_vector = action_step / 100.0
swmm.modify_setting(controlled_ponds[i].orifice_id,
agents_dqn[i].action_vector)
current_gate = agents_dqn[i].action_vector
# SWMM step
swmm.run_step()
# Receive the new rewards
outflow = swmm.get('ZOF1', swmm.INFLOW, swmm.SI)
outflow_track.append(outflow)
overflows = swmm_states(all_nodes, swmm.FLOODING)
gate_change = np.abs(current_gate - temp_gate)
temp_gate = current_gate
depth = swmm.get("93-50077", swmm.DEPTH, swmm.SI)
floos_1 = swmm.get("93-50077", swmm.FLOODING, swmm.SI)
r_temp = reward_function(depth, outflow, gate_change, floos_1)
for i in nodes_controlled.keys():
agents_dqn[i].rewards_vector = r_temp
# Observe the new states
for i in nodes_controlled.keys():
agents_dqn[i].state_new_vector = swmm_states(states_controlled[i],
swmm.DEPTH)