def plot_zoomed_original_predicted_energy_consumption():
"""
Plots a zoomed time frame of the original prediction.
"""
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start="30-6-2013", end="15-7-2013")
test_building = 1
sample_period = 6
meter_keys = ['kettle']
test_elec = test.buildings[test_building].elec
results_dir = '../results/UKDALE-ACROSS-BUILDINGS-RNN-lr=1e-05-2018-02-20-14-24-46'
disag_filename = 'disag-out.h5'
for key in meter_keys:
# get predicted curve for the best epoch
result = DataSet(os.path.join(results_dir, disag_filename))
res_elec = result.buildings[test_building].elec
predicted = res_elec[key]
predicted = predicted.power_series(sample_period=sample_period)
predicted = next(predicted)
predicted.fillna(0, inplace=True)
y1 = np.array(predicted) # power
x1 = np.arange(y1.shape[0]) # timestamps
# The chosen time frame to zoom in
x1 = x1[94000:102500]
y1 = y1[94000:102500]
ground_truth = test_elec[key]
ground_truth = ground_truth.power_series(sample_period=sample_period)
ground_truth = next(ground_truth)
ground_truth.fillna(0, inplace=True)
y2 = np.array(ground_truth) # power
x2 = np.arange(y2.shape[0]) # timestamps
# The chosen time frame to zoom in
x2 = x2[94000:102500]
y2 = y2[94000:102500]
fig, (ax1, ax2, ax3) = plt.subplots(3, sharex=True, sharey=True)
ax1.plot(x1, y1, color='r', label='predicted')
ax1.plot(x2, y2, color='b', label='ground truth')
ax2.plot(x1, y1, color='r')
ax3.plot(x2, y2, color='b')
ax1.set_title('Appliance: {}'.format(key))
plt.xticks(
np.arange(94000, 102500, 2000),
('5-10-2013 12:00', '16:00', '20:00', '6-10-2013 00:00', '04:00'))
fig.legend()
fig.savefig(
os.path.join(
results_dir,
'zoomed_original_predicted_vs_ground_truth_{}.png'.format(
key)))
def plot_zoomed_new_predicted_energy_consumption():
"""
Predicts a new short window (of the given test set).
"""
train = DataSet('../data/ukdale.h5')
train.clear_cache()
train.set_window(start="13-4-2013", end="31-7-2013")
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start='16-9-2013 17:00:00', end='16-9-2013 18:00:00')
train_building = 1
test_building = 1
sample_period = 6
meter_key = 'kettle'
learning_rate = 1e-5
best_epoch = 140
train_elec = train.buildings[train_building].elec
test_elec = test.buildings[test_building].elec
train_meter = train_elec.submeters()[meter_key]
test_mains = test_elec.mains()
results_dir = '../results/UKDALE-RNN-lr=1e-05-2018-02-16-18-52-34'
train_logfile = os.path.join(results_dir, 'training.log')
val_logfile = os.path.join(results_dir, 'validation.log')
rnn = RNNDisaggregator(train_logfile,
val_logfile,
learning_rate,
init=False)
model = 'UKDALE-RNN-kettle-{}epochs.h5'.format(best_epoch)
rnn.import_model(os.path.join(results_dir, model))
disag_filename = 'disag-out-{}epochs.h5'.format(best_epoch)
output = HDFDataStore(os.path.join(results_dir, disag_filename), 'w')
results_file = os.path.join(results_dir,
'results-{}epochs.txt'.format(best_epoch))
rnn.disaggregate(test_mains,
output,
results_file,
train_meter,
sample_period=sample_period)
os.remove(results_file)
output.close()
# get predicted curve for the best epoch
result = DataSet(os.path.join(results_dir, disag_filename))
res_elec = result.buildings[test_building].elec
os.remove(os.path.join(results_dir, disag_filename))
predicted = res_elec[meter_key]
predicted = predicted.power_series(sample_period=sample_period)
predicted = next(predicted)
predicted.fillna(0, inplace=True)
y1 = np.array(predicted) # power
x1 = np.arange(y1.shape[0]) # timestamps
ground_truth = test_elec[meter_key]
ground_truth = ground_truth.power_series(sample_period=sample_period)
ground_truth = next(ground_truth)
ground_truth.fillna(0, inplace=True)
y2 = np.array(ground_truth) # power
x2 = np.arange(y2.shape[0]) # timestamps
fig, (ax1, ax2, ax3) = plt.subplots(3, sharex=True, sharey=True)
ax1.plot(x1, y1, color='r', label='predicted')
ax1.plot(x2, y2, color='b', label='ground truth')
ax2.plot(x1, y1, color='r')
ax3.plot(x2, y2, color='b')
ax1.set_title('Appliance: {}'.format(meter_key))
fig.legend()
fig.savefig(
os.path.join(results_dir, 'zoomed_new_predicted_vs_ground_truth.png'))
def plot_datasets_meter():
"""
Plots the target appliance (labels) of training, validation and test sets.
"""
windows = {
'train': [["13-4-2013", "30-9-2013"]],
'validation': ["13-4-2013", "31-7-2013"],
'test': ["6-1-2014", "11-1-2014"]
}
train = []
for window in windows['train']:
t = DataSet('../data/ukdale.h5')
t.clear_cache()
t.set_window(start=window[0], end=window[1])
train += [t]
validation = DataSet('../data/ukdale.h5')
validation.clear_cache()
validation.set_window(start=windows['validation'][0],
end=windows['validation'][1])
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start=windows['test'][0], end=windows['test'][1])
train_buildings = [1]
val_buildings = [2]
test_building = 5
meter_key = 'microwave'
results_dir = '../results/UKDALE-RNN-lr=1e-05-2018-02-16-18-52-34'
train_meterlist = []
val_meterlist = []
for i, b in enumerate(train_buildings):
train_elec = train[i].buildings[b].elec
train_meterlist += [train_elec.submeters()[meter_key]]
for i in val_buildings:
val_elec = validation.buildings[i].elec
val_meterlist += [val_elec.submeters()[meter_key]]
test_elec = test.buildings[test_building].elec
test_meter = test_elec.submeters()[meter_key]
fig, (ax1, ax2, ax3) = plt.subplots(3, sharex=False, sharey=True)
fig.set_size_inches(15.5, 10.5)
train_meterlist[0].plot(
ax=ax1,
plot_kwargs={
'color': 'g',
'label': 'Train set - building {}'.format(train_buildings)
},
plot_legend=False)
val_meterlist[0].plot(
ax=ax2,
plot_kwargs={
'color': 'y',
'label': 'Validation set - building {}'.format(val_buildings)
},
plot_legend=False)
test_meter.plot(ax=ax3,
plot_kwargs={
'color': 'r',
'label': 'Test set - building {}'.format(test_building)
},
plot_legend=False)
ax1.set_title('Appliance: {}'.format(meter_key))
fig.legend()
plt.savefig(os.path.join(results_dir, 'datasets.png'))
def generate_vertices():
"""
Predicts the power demand of the target appliance using the intermediate models which are exported during training.
Generates a polygon from those predictions.
"""
train = DataSet('../data/ukdale.h5')
train.clear_cache()
train.set_window(start="13-4-2013", end="31-7-2013")
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start='7-2-2014 08:00:00', end='7-3-2014')
train_building = 1
test_building = 5
sample_period = 6
meter_key = 'kettle'
learning_rate = 1e-5
train_elec = train.buildings[train_building].elec
test_elec = test.buildings[test_building].elec
train_meter = train_elec.submeters()[meter_key]
test_mains = test_elec.mains()
results_dir = '../results/UKDALE-ACROSS-BUILDINGS-RNN-lr=1e-05-2018-02-03-11-48-12'
train_logfile = os.path.join(results_dir, 'training.log')
val_logfile = os.path.join(results_dir, 'validation.log')
rnn = RNNDisaggregator(train_logfile,
val_logfile,
learning_rate,
init=False)
verts = []
zs = [] # epochs
for z in np.arange(10, 341, 10):
# disaggregate model
model = 'UKDALE-RNN-kettle-{}epochs.h5'.format(z)
rnn.import_model(os.path.join(results_dir, model))
disag_filename = 'disag-out-{}epochs.h5'.format(z)
output = HDFDataStore(os.path.join(results_dir, disag_filename), 'w')
results_file = os.path.join(results_dir,
'results-{}epochs.txt'.format(z))
rnn.disaggregate(test_mains,
output,
results_file,
train_meter,
sample_period=sample_period)
os.remove(results_file)
output.close()
# get predicted curve for epoch=z
result = DataSet(os.path.join(results_dir, disag_filename))
res_elec = result.buildings[test_building].elec
os.remove(os.path.join(results_dir, disag_filename))
predicted = res_elec[meter_key]
predicted = predicted.power_series(sample_period=sample_period)
predicted = next(predicted)
predicted.fillna(0, inplace=True)
ys = np.array(predicted) # power
xs = np.arange(ys.shape[0]) # timestamps
verts.append(list(zip(xs, ys))) # add list of x-y-coordinates
zs.append(z)
ground_truth = test_elec[meter_key]
ground_truth = ground_truth.power_series(sample_period=sample_period)
ground_truth = next(ground_truth)
ground_truth.fillna(0, inplace=True)
ys = np.array(ground_truth) # power
xs = np.arange(ys.shape[0]) # timestamps
verts.append(list(zip(xs, ys))) # add list of x-y-coordinates
zs.append(350)
zs = np.asarray(zs)
for i in range(len(verts)):
verts[i].insert(0, [0, np.array([0])])
verts[i].append([len(verts[i]), np.array([0])])
pickle.dump(verts, open(os.path.join(results_dir, 'vertices.pkl'), 'wb'))
pickle.dump(zs, open(os.path.join(results_dir, 'zs.pkl'), 'wb'))
pickle.dump(ys, open(os.path.join(results_dir, 'ys.pkl'), 'wb'))
def plot_prediction_over_epochs_ploty():
"""
Predicts the power demand of the target appliance using the intermediate models which are exported during training.
Plots the prediction curves using plotly.
"""
train = DataSet('../data/ukdale.h5')
train.clear_cache()
train.set_window(start="13-4-2013", end="31-7-2013")
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start="23-7-2014 10:00:00", end="23-7-2014 11:00:00")
train_building = 1
test_building = 5
sample_period = 6
meter_key = 'kettle'
learning_rate = 1e-5
train_elec = train.buildings[train_building].elec
test_elec = test.buildings[test_building].elec
train_meter = train_elec.submeters()[meter_key]
test_mains = test_elec.mains()
results_dir = '../results/UKDALE-ACROSS-BUILDINGS-RNN-lr=1e-05-2018-02-03-11-48-12'
train_logfile = os.path.join(results_dir, 'training.log')
val_logfile = os.path.join(results_dir, 'validation.log')
rnn = RNNDisaggregator(train_logfile,
val_logfile,
learning_rate,
init=False)
data = []
for i in range(10, 401, 10):
# disaggregate model
model = 'UKDALE-RNN-kettle-{}epochs.h5'.format(i)
rnn.import_model(os.path.join(results_dir, model))
disag_filename = 'disag-out-{}epochs.h5'.format(i)
output = HDFDataStore(os.path.join(results_dir, disag_filename), 'w')
results_file = os.path.join(results_dir,
'results-{}epochs.txt'.format(i))
rnn.disaggregate(test_mains,
output,
results_file,
train_meter,
sample_period=sample_period)
os.remove(results_file)
output.close()
# plot predicted curve for epoch=i
result = DataSet(os.path.join(results_dir, disag_filename))
res_elec = result.buildings[test_building].elec
os.remove(os.path.join(results_dir, disag_filename))
predicted = res_elec[meter_key]
predicted = predicted.power_series(sample_period=sample_period)
predicted = next(predicted)
predicted.fillna(0, inplace=True)
power = predicted.tolist()
length = len(power)
timestamps = list(range(length))
x = []
y = []
z = []
ci = int(255 / 420 * i) # ci = "color index"
for j in range(length):
x.append([timestamps[j], timestamps[j]]) # timestamps
y.append([i, i + 5]) # epochs
z.append([power[j], power[j]]) # power
data.append(
dict(
z=z,
x=x,
y=y,
colorscale=[[i, 'rgb(%d,%d,255)' % (ci, ci)]
for i in np.arange(0, 1.1, 0.1)],
showscale=False,
type='surface',
))
# plot ground truth curve as the last curve
ground_truth = test_elec[meter_key]
ground_truth = ground_truth.power_series(sample_period=sample_period)
ground_truth = next(ground_truth)
ground_truth.fillna(0, inplace=True)
power = ground_truth.tolist()
length = len(power)
timestamps = list(range(length))
i = 410
x = []
y = []
z = []
ci = int(255 / 410 * i) # ci = "color index"
for j in range(length):
x.append([timestamps[j], timestamps[j]]) # timestamps
y.append([i, i + 5]) # epochs
z.append([power[j], power[j]]) # power
data.append(
dict(
z=z,
x=x,
y=y,
colorscale=[[i, 'rgb(%d,%d,255)' % (ci, ci)]
for i in np.arange(0, 1.1, 0.1)],
showscale=False,
type='surface',
))
layout = dict(title='prediction over epochs',
showlegend=False,
scene=dict(xaxis=dict(title='timestamps'),
yaxis=dict(title='epochs'),
zaxis=dict(title='power'),
camera=dict(eye=dict(x=-1.7, y=-1.7, z=0.5))))
fig = dict(data=data, layout=layout)
plotly.offline.plot(fig, filename='filled-3d-lines')
IMPORT = False # TODO: True if continue training
windows = {
'train': [['13-4-2013', '13-5-2013'], ['13-6-2013', '13-7-2013'],
['13-5-2013', '13-6-2013'], ['13-7-2014', '31-7-2014']],
'validation': [['13-5-2013', '20-5-2013'], ['13-7-2013', '20-7-2013'],
['13-4-2013', '20-4-2013'], ['7-6-2014', '13-7-2014']],
'test': ['30-6-2013', '15-7-2013']
}
train = []
print("========== OPEN DATASETS ============")
for window in windows['train']:
t = DataSet('../data/ukdale.h5')
t.clear_cache()
t.set_window(start=window[0], end=window[1])
train += [t]
validation = []
for window in windows['validation']:
v = DataSet('../data/ukdale.h5')
v.clear_cache()
v.set_window(start=window[0], end=window[1])
validation += [v]
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start=windows['test'][0], end=windows['test'][1])
train_mainslist = []
from rnndisaggregator import RNNDisaggregator
from plots import plot_loss
IMPORT = False # TODO: True if continue training
windows = {
'train': ["13-4-2013", "31-7-2013"],
'validation': ["31-7-2013", "31-8-2013"],
'test': ["13-9-2013", "30-9-2013"]
}
print("========== OPEN DATASETS ============")
train = DataSet('../data/ukdale.h5')
train.clear_cache()
train.set_window(start=windows['train'][0], end=windows['train'][1])
validation = DataSet('../data/ukdale.h5')
validation.clear_cache()
validation.set_window(start=windows['validation'][0], end=windows['validation'][1])
test = DataSet('../data/ukdale.h5')
test.clear_cache()
test.set_window(start=windows['test'][0], end=windows['test'][1])
train_building = 1
validation_building = 1
test_building = 1
sample_period = 6
meter_key = 'kettle'
learning_rate = 1e-5