In [1]:
from __future__ import print_function, division
import matplotlib
matplotlib.use('nbagg') # interactive plots in iPython. New in matplotlib v1.4
# %matplotlib inline
In [2]:
import matplotlib.pyplot as plt
from nilmtk import DataSet, MeterGroup
import pandas as pd
import numpy as np
from time import time
In [3]:
from pybrain.supervised import RPropMinusTrainer
from pybrain.datasets import SequentialDataSet
from pybrain.structure import RecurrentNetwork, FullConnection
from pybrain.structure.modules import LSTMLayer, BiasUnit, LinearLayer, TanhLayer, SigmoidLayer
In [28]:
CONFIG = dict(
EPOCHS_PER_CYCLE = 5,
CYCLES = 10,
HIDDEN_LAYERS = [15, 15],
PEEPHOLES = True,
TRAINERCLASS = RPropMinusTrainer,
# instead, you may also try
# TRAINERCLASS = BackpropTrainer(net, dataset=trndata, verbose=True,
# momentum=0.9, learningrate=0.00001)
INPUTS = ['fridge'], #, 'hour of day (int)', 'outside temperature', 'is business day (-1, 1)'
EXPERIMENT_NUMBER = 1
)
In [5]:
# Load dataset
dataset = DataSet('/data/mine/vadeec/merged/ukdale.h5')
dataset.set_window("2014-01-01", "2014-03-01")
elec = dataset.buildings[1].elec
In [6]:
# Select top-5 meters identified in UK-DALE paper
# APPLIANCES = ['kettle', 'dish washer', 'HTPC', 'washer dryer', 'fridge freezer']
APPLIANCES = ['fridge freezer']
selected_meters = [elec[appliance] for appliance in APPLIANCES]
selected_meters.append(elec.mains())
selected = MeterGroup(selected_meters)
In [7]:
df = selected.dataframe_of_meters()
In [8]:
# Use human-readable column names
df.columns = selected.get_labels(df.columns)
In [9]:
mains = df['Site meter'].diff().dropna()
appliances = df.iloc[:,:-1].fillna(0).diff().dropna()
del df
In [10]:
# standardise input
mains = (mains - mains.mean()) / mains.std()
# Constrain outputs to [-1,1] because we're using TanH
appliances /= appliances.max()
In [11]:
mains.plot()
plt.show()
In [12]:
appliances.plot()
plt.show()
In [13]:
# Build PyBrain dataset
N_OUTPUTS = appliances.shape[1]
N_INPUTS = 1
N = len(mains)
ds = SequentialDataSet(N_INPUTS, N_OUTPUTS)
ds.newSequence()
ds.setField('input', pd.DataFrame(mains).values)
ds.setField('target', appliances.values)
In [14]:
ds.getSequence(0)
Out[14]:
In [15]:
# Build network
net = RecurrentNetwork()
def lstm_layer_name(i):
return 'LSTM{:d}'.format(i)
# Add modules
net.addInputModule(LinearLayer(ds.indim, name='in'))
net.addOutputModule(TanhLayer(dim=ds.outdim, name='out'))
for i, n_cells in enumerate(CONFIG['HIDDEN_LAYERS']):
net.addModule(LSTMLayer(n_cells, name=lstm_layer_name(i+1), peepholes=CONFIG['PEEPHOLES']))
# Bias
bias = BiasUnit()
net.addModule(bias)
net.addConnection(FullConnection(bias, net['out'], name='c_output_bias'))
# Add connections
n_hidden_layers = len(CONFIG['HIDDEN_LAYERS'])
prev_layer_name = 'in'
for i in range(n_hidden_layers):
hidden_layer_i = i + 1
layer_name = lstm_layer_name(hidden_layer_i)
net.addRecurrentConnection(FullConnection(net[layer_name], net[layer_name], name='c_' + layer_name + '_to_' + layer_name))
net.addConnection(FullConnection(bias, net[layer_name], name='c_' + layer_name + '_bias'))
net.addConnection(FullConnection(net[prev_layer_name], net[layer_name], name='c_' + prev_layer_name + '_to_' + layer_name))
prev_layer_name = layer_name
layer_name = lstm_layer_name(n_hidden_layers)
net.addConnection(FullConnection(net[layer_name], net['out'], name='c_' + layer_name + '_to_out'))
net.sortModules()
print(net)
In [16]:
# define a training method
net.randomize()
net._setParameters(np.random.uniform(-0.1, 0.1, size=net.paramdim))
trainer = CONFIG['TRAINERCLASS'](net, dataset=ds, verbose=True)
In [17]:
# carry out the training
net.reset()
# train_errors = []
t0 = time()
EPOCHS = CONFIG['EPOCHS_PER_CYCLE'] * CONFIG['CYCLES']
# trainer.trainUntilConvergence(maxEpochs=EPOCHS, verbose=True)
# start_time = time()
print("Starting training with", EPOCHS, "epochs...")
for i in xrange(CONFIG['CYCLES']):
trainer.trainEpochs(CONFIG['EPOCHS_PER_CYCLE'])
# train_errors.append(trainer.testOnData())
# epoch = (i+1) * CONFIG['EPOCHS_PER_CYCLE']
# seconds_elapsed = time() - start_time
# seconds_per_epoch = seconds_elapsed / epoch
# seconds_remaining = (EPOCHS - epoch) * seconds_per_epoch
# td_elapsed = timedelta(seconds=seconds_elapsed)
# td_elapsed_str = str(td_elapsed).split('.')[0]
# eta = (datetime.now() + timedelta(seconds=seconds_remaining)).time()
# eta = eta.strftime("%H:%M:%S")
# print("\r epoch = {}/{} error = {} elapsed = {} ETA = {}"
# .format(epoch, EPOCHS, train_errors[-1], td_elapsed_str, eta),
# end="")
# stdout.flush()
print("Finished training. total seconds =", time() - t0)
In [18]:
# Disaggregate!
START = "2014-01-01"
END = "2014-01-03"
print("Starting disaggregation...")
net.reset()
estimates = pd.DataFrame(columns=appliances.columns, index=appliances[START:END].index)
for date, mains_value in mains[START:END].iteritems():
estimates.loc[date] = net.activate(mains_value)
In [19]:
estimates.plot()
plt.show()
In [20]:
appliances[START:END].plot()
plt.show()
In [24]:
appliances[START:END].cumsum().plot()
estimates[START:END].cumsum().plot()
plt.show()
In [25]:
mains[START:END].cumsum().plot()
plt.show()
In [31]:
estimates.cumsum().to_hdf('neuronilm_estimates_{:03d}.hdf'.format(CONFIG['EXPERIMENT_NUMBER']), 'df')
In [34]:
ls -alh
In [ ]: