Optimization of In-House Energy Demand - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 276

The accuracy of the appliance state classification and the implications for heating
control will be scrutinized in the following section.
10.5 Empirical Evaluation
The goal of our evaluation is twofold: (i) we investigate which of the considered
machine learning models is most accurate for the the appliance state classification
task; and (ii) we assess the use of the identified appliance or rather occupancy states
for heating control.
10.5.1 Energy Data
We consider the REDD dataset [ 10 ], which comprises electricity consumption mea-
surements from six household at the granularity level of individual devices, and rep-
resents to date one of the largest and richest publicly available collections of power
consumption data [ 2 ]. There are approximately 20 consecutive days of measurements
available for each house, providing data from the two main phases and each individ-
ual circuit at 1Hz frequency rate. Measured appliances include main consumers such
as Air Conditioning, Dishwasher, Disposal, Electrical Heating, Microwave, Oven,
Refrigerator, Stove, Washer/Dryer as well as other miscellaneous electronics and
outlets (see Table 10.2 ).
10.5.2 Experimental Design
In our empirical evaluation, we compare the classification accuracy of the introduced
machine learning models (see Table 10.1 ) on the REDD data set. Strictly speaking,
we assess the appliance state classification accuracy for all considered models on
a granularity level of individual devices. The training of the respective models is
done on appliance-specific consumption measurements of one particular device for
all households but one. The aggregated electricity consumption signal of the left-
out household is then used for testing the performance of the trained models for
each individual device. This evaluation principle is also commonly known as cross-
validation with leave-one-out.
10.5.3 Classification Accuracy
Table 10.2 illustrates the classification accuracy per (a) household and (b) appli-
ance for all examined models, including Naive Bayes (NB), Factorial Hidden
Markov Model (FHMM), Classification Trees (CT), and One-Nearest-Neighbor