Optimization of In-House Energy Demand - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 270

demonstrated in Sect. 10.6 . Eventually, we conclude our study and give an outlook
on future work in Sect. 10.7 .
10.2 Background
Energy disaggregation, also referred to as nonintrusive load monitoring, is the task
of using an aggregated energy signal, such as that coming from a whole-home power
monitor, to make inferences about the different individual loads of the system [ 10 ].
This approach is seen as an intermediate between existing electricity meters (which
merely record whole-home power usage) and fully energy-aware home appliance
networks, where each individual device reports its own consumption [ 18 ].
For a thorough evaluation of various energy disaggregation mechanisms under
real-world conditions, a comprehensive collection of power consumption data is
needed [ 18 ]. Most approaches to energy disaggregation have been supervised, in
that the model is trained on individual device power signals [ 23 ]. The vast majority
of supervised disaggregation approaches have evaluated the trained models on the
same devices but in new conditions [ 1 ].
Research on energy disaggregation has been encouraged by publicly available
datasets such as REDD [ 10 ], which contains information about the power consump-
tion of several different homes on device level, and, therefore, allows cross-validation
for individual appliances. Experiments on the REDD dataset have shown that the
Factorial Hidden Markov Model (FHMM) is able to disaggregate the power data
reasonably well [ 10 ]. In that case, the disaggregation task is framed as an inference
problem and the performance of energy disaggregation is evaluated considering the
percentage of energy correctly classified.
Although FHMMs have shown to be a powerful tool [ 5 ] for learning probabilistic
models of multivariate time series, the combinatorial nature of distributed state rep-
resentation makes an exact algorithm for inferring the posterior probabilities of the
hidden state variables intractable. Approximate inference can be carried out using
Gibbs sampling or variational methods [ 5 ]. Recent work [ 8 ] on energy disaggregation
presents different FHMM variants which incorporate additional features and better
fit the probability distribution of the state occupancy durations of the appliances.
Another work [ 19 ] proposes Artificial Neural Networks (ANNs) for appliance
recognition, because they (i) do not require prior understanding of appliance behavior,
(ii) are capable of handling multiple states, and (iii) are able to learn while running.
The results show that after training the ANNwith generated appliance signatures, the
proposed system is able to recognize the previously learned appliances with relatively
high accuracy, even in demanding scenarios. To tune the ANN, the authors suggest to
use the generated signatures to create a training dataset with all possible combinations
of appliance activity. Comparing the disaggregation performance for different ANN
algorithms, additional work [ 11 ] suggests to employ back-propagation rather than
the radial-base-function.