Optimization of In-House Energy Demand - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 271

In another study [ 21 ], the authors propose a disaggregation algorithm that consists
of several consecutive steps including normalization, edge detection via thresholding
and smoothing techniques, extraction of power-level and delta-level consumption,
matching of known appliances from a signature database with extracted delta vectors,
and labeling of recognized devices. The proposed system does not require setup or
training, because the user is able to label appliance signatures via her smart phone. In
that case, the appliance signatures are based on apparent, reactive, real, and distortion
power measured by the smart meter.
The classification of household items based on their electricity usage profile over
a fixed time interval is discussed in yet another study [ 13 ]. The authors consider the
time series classification problem of identifying device types through daily or weekly
demand profiles. The proposed approach concentrates on bespoken features such as
mean, variance, kurtosis, skewness, slope, and run measures. The experiments show
that classification using the bespoken features performs better than classification
using the raw data. However, the nature of similarity captured strongly depends on
the features extracted.
In a similar work [ 18 ], the authors present an appliance identification approach
based on characteristic features of traces collected during the 24h of a day. The
extracted features include temporal appliance behavior, power consumption levels,
shape of the consumption, active phase statistics, and noise level characteristics.
Each resulting feature vector is annotated by the actual device class and used to train
the underlying model of the selected classifier. Among various tested classifiers, the
Random Committee algorithm performs best in categorizing new and yet unseen
feature vectors into one of the previously trained device types. Additional work [ 11 ]
demonstrates that the solution from any single-feature, single-algorithm disaggrega-
tion approach could be combined under a committee decision mechanism to render
the best solution.
Yet another work [ 20 ] presents a nonintrusive appliance load monitoring tech-
nique based on integer programming. Since the overall load current is expressed as
a superposition of each current of the operating appliance, the monitoring problem
can be formulated as an integer quadratic programming problem by expressing the
operating conditions as integer variables. Besides that the proposed method does not
require relearning when a new appliance is installed in the house, it is furthermore
able to distinguish between different device modes and some-type appliances that
operate simultaneously.
To monitor the states of multiple appliances via electricity consumption measure-
ments, another work [ 12 ] introduces the Bayes filter approach, which computes the
posterior distribution over the current state given all observations to date. Since the
state transition of an appliance is a continuous process, the authors employ a sliding
window to take the temporal factor into consideration and extract the past records of
data to be features. The estimated states are represented as binary strings, where each
bit denotes the on/off state of one individual appliance. According to the results, the
Bayes filter outperforms the KNN, Naive Bayes, and SVM classifier.
Leveraging recent advances in device and appliance power supplies, another series
of studies [ 4 , 6 ] extends the energy disaggregation approach by using high-frequency