Information Technology Reference
In-Depth Information
Table
10.2
b presents the classification accuracy of device states per appliance
averaged over all households. For example, the results show that the NB model is
able to classify the device states of the Air-Conditioning with an average accuracy of
93.15%, taking the mean of House 1-6. In general, all models achieved a relatively
high classification accuracy for appliances with distinctive energy profiles, such as
the Dishwasher or Oven, but performed less well on appliances with changes in
consumption that can easily be confused with other devices, like the Refrigerator or
Lighting.
By taking the mean over all results for (a) each household and (b) each appliance
per model, shown in the bottom row of Table
10.2
a, b respectively, we can easily see
that on average the NB model achieved the highest classification accuracy, closely
followed by FHMM and CT. Although the 1NN classifier shows relatively high clas-
sification accuracy for several individual appliances, it is unable to correctly classify
the device states of others, and, therefore, achieve the lowest average classification
performance.
10.5.4 Heating Control
In this subsection, we discuss how the classified
device states can be used
for heating control and scheduling. Since the Naive Bayes (NB) model achieved the
highest average accuracy on classifying device states of appliances in an unknown
household (see Table
10.2
), we will consider the NB approach in our following exem-
plification.
Figure
10.3
shows the (a) observed and (b) estimated
on
/
off
states for the
Washer/Dryer in House 1 over a period of 4 weeks, where every quarter of an hour
aggregates the device activities that occurred during the same weekday and time of
day. By illustrating the (a) observed activity of the Washer/Dryer, which constitutes
our ground truth, we see that this appliance is mostly used on Fridays and weekends.
The (b) estimated activity of the Washer/Dryer, inferred from the overall energy con-
sumption of House 1 by the trained NB model, shows similar behavior patterns for
weekends, but predicts false
on
/
off
states for Mondays.
By taking a closer look at the confusion matrix of observed and estimated
on
off
device states for the Washer/Dryer in House 1, shown in Table
10.3
, we are able to
gain a better understanding of the estimated appliance activity. Table
10.3
reveals
the percentage of true positives (TP) or true
on
/
on
states, true negatives (TN) or true
off
states, false positives (FP) or false
on
states, and false negatives (FN) or false
off
states. Although the NB model achieves a high classification accuracy
[
(
TP
+
TN
)/(
TP
+
TN
+
FP
+
FN
)
=
96
.
83%
]
, the percentage of falsely classified states
[
in not negligible, explaining the mistakenWasher/Dryer activity
estimated for Mondays (see Fig.
10.3
b). The FP and FN estimates imply heating
during absence and cooling during occupancy, respectively.
The cause of falsely classified states can also be explained with help of Fig.
10.2
.
By examining the distribution of
FP
+
FN
=
3
.
17%
]
on
and
off
states of the refrigerator in House 1,
Search WWH ::
Custom Search