SEQUENTIAL CLUSTERING-BASED EVENT DETECTION FOR NONINTRUSIVE LOAD MONITORING

Jan Zizka et al. (Eds) : CCSIT, SIPP, AISC, CMCA, SEAS, CSITEC, DaKM, PDCTA, NeCoM - 2016
pp. 77–85, 2016. © CS & IT-CSCP 2016 DOI : 10.5121/csit.2016.60108
SEQUENTIAL CLUSTERING-BASED
EVENT DETECTION FOR NON-
INTRUSIVE LOAD MONITORING
Karim Said Barsim and Bin Yang
Institute of Signal Processing and System Theory,
University of Stuttgart, Germany
karim.barsim@iss.uni-stuttgart.de
bin.yang@iss.uni-stuttgart.de
ABSTRACT
The problem of change-point detection has been well studied and adopted in many signal
processing applications. In such applications, the informative segments of the signal are the
stationary ones before and after the change-point. However, for some novel signal processing
and machine learning applications such as Non-Intrusive Load Monitoring (NILM), the
information contained in the non-stationary transient intervals is of equal or even more
importance to the recognition process. In this paper, we introduce a novel clustering-based
sequential detection of abrupt changes in an aggregate electricity consumption profile with
accurate decomposition of the input signal into stationary and non-stationary segments. We also
introduce various event models in the context of clustering analysis. The proposed algorithm is
applied to building-level energy profiles with promising results for the residential BLUED
power dataset.
KEYWORDS
Event detection, change-interval detection, density-based clustering, DBSCAN, non-intrusive
load monitoring, NILM, BLUED, energy disaggregation
1. INTRODUCTION
Non-Intrusive Load Monitoring (NILM), also known as electricity disaggregation, is an energy
monitoring technique that aims at inferring the energy consumption profiles of individual
electrical loads merely from a single or a limited number of aggregate measurement points in a
building [1]. Recently, NILM has witnessed a rapidly increasing progress in both academic and
commercial research due to its promising applications in energy conservation, activity monitoring
[2], dynamic pricing [3], demand forecasting [4], and home automation [5]. Currently, the
majority of NILM systems are event-based approaches in the sense that they rely on the detection
of abrupt changes occurring in the aggregate signal which indicate state-changes of the monitored
appliances. It was observed that events attain distinctive features according to the physical
properties of their appliances such as energy storage elements, counter-electromotive force in
induction motors, or striking voltages in fluorescent lamps. Features extracted from steady and
transient intervals (such as power surges, overshoot currents, decay rate, etc) are utilized in event
clustering or classification stages of the disaggregation system. Consequently, a robust detection

78 Computer Science & Information Technology (CS & IT)
and accurate segmentation of such change-intervals is of particular importance for event-based
NILM systems.
Basseville and Nikiforov [6] described various detection algorithms from which two approaches
have been utilized in event-based NILM systems, namely the Generalized Likelihood Ratio
(GLR) test [7, 8] and the CUmulative SUM (CUSUM) filtering [9]. Jin et al. [10] proposed a
more robust change-point detection approach based on a Goodness-of-Fit (GoF) test. In addition,
various machine learning tools such as kernel clustering [11], Hidden Markov Models (HMM)
[12], and Support Vector Machines (SVMs) [13], have been proposed as solutions to address the
change point detection problem.
Even though many previous works on NILM proposed utilizing features extracted from the
transient intervals, only few event detection approaches consider accurate segmentation of the
transient periods for the extraction of more stable transient features [9, 14]. Moreover, many
approaches need a probabilistic model for the sample distribution in the stationary segments
which is often difficult to obtain from aggregate consumption profile of several, simultaneously
operating appliances. The result is that the current event detection algorithms are not robust and
fail sometimes provide reliable event-based feature for appliance recognition in practice. In this
paper, we propose a novel clustering-based event detection algorithm for event-based NILM
systems. In contrast to other event detection algorithms, the proposed approach features accurate
segmentation of the input signal into stationary (steady) and non-stationary (transient) segments.
Such accurate segmentation is crucial for the extraction of more stable and repeatable features
from both transient and steady-state intervals. Moreover, the utilized density-based clustering
scheme does not impose any probabilistic models on the sample distribution in either of the
stationary segments and supports arbitrarily shaped, weakly stationary segments leading to an
enhanced robustness to noise. In addition, the proposed algorithm features a sequential (instead of
batch) clustering that is more efficient for real-time NILM systems.
The presented approach is modular in the sense that it can combine any clustering-based event
detection algorithm with any event model. For this purpose, we also introduce different event
models at different complexity- and robustness-levels. This paper is organized as follows. In
section 2, we introduce different event models in the context of spatial and time-series clustering.
In section 3, we describe the proposed sequential event detection algorithm in which the Density-
Based Spatial Clustering for Applications with Noise [15] is assumed and utilized sequentially in
spatial and temporal analysis of the input power signals. Section 4 shows results of application of
the proposed algorithm on the publicly available, residential BLUED [16] dataset. Finally,
section 5 concludes this paper.
2. EVENT MODELS
Event models will be introduced in the order of their increasing coverage of real events,
robustness, and complexity.
Let the matrix
‫܆‬ = ሾ࢞ଵ, ࢞ଶ, … , ࢞ேሿ, ࢞௡ ∈ ℝ௟
(1)
contain a time series of ܰ consecutive ݈-dimensional data samples (feature vectors). Typically,
࢞௡ contains the measured real ܲ and reactive ܳ powers at time instance ݊. Assume that all ܰ
samples have been clustered into ݉ non-empty, disjoint clusters (sets) ‫ܥ‬ଵ, ‫ܥ‬ଶ, … , ‫ܥ‬௠. In addition,

Computer Science & Information Technology (CS & IT) 79
we assume that a noise-aware clustering algorithm assigns un-clustered samples (i.e. outliers or
noisy samples) to the set ‫ܥ‬଴. Clearly, ∑ |‫ܥ‬௜|௠
௜ୀ଴ = ܰ where |‫ܥ‬௜| is the cardinality of the cluster
|‫ܥ‬௜|. Let
‫ݕ‬௡ = ߱(࢞௡) ∈ ሼ0,1,2, … , ݉ሽ (2)
be the corresponding cluster index of ࢞௡ (i.e. ࢞௡ ∈ ‫ܥ‬௬೙
). We then introduce the following
definitions for two metrics of a cluster and three different event models:
Definition 1: The temporal length Len(‫ܥ‬௜) of cluster ‫ܥ‬௜ is defined as the minimum window size
that contains all its elements. If
∃‫:ݑ‬ ࢞௨ ∈ ‫ܥ‬௜ and ࢞௡ ∉ ‫ܥ‬௜ ∀(݊ < ‫)ݑ‬ (3)
∃‫:ݒ‬ ࢞௩ ∈ ‫ܥ‬௜ and ࢞௡ ∉ ‫ܥ‬௜ ∀(݊ > ‫)ݒ‬ (4)
Then Len(‫ܥ‬௜) is defined as
Len(‫ܥ‬௜) = ‫ݒ‬ − ‫ݑ‬ + 1 ≥ |‫ܥ‬௜| (5)
Here ‫ݑ‬ and ‫ݒ‬ denote the time instances of the first and last samples belonging to ‫ܥ‬௜, respectively.
Definition 2: The temporal locality ratio Loc(‫ܥ‬௜) of cluster ‫ܥ‬௜ is defined as
Loc(‫ܥ‬௜) =
|‫ܥ‬௜|
Len(‫ܥ‬௜)
∈ ሿ 0, 1 ሿ (6)
The temporal locality ratio is a measure of how a cluster is spreading over time domain. A value
of one (Loc(‫ܥ‬௜) = 1) refers to the maximum temporal locality where the cluster is represented by
a single segment of consecutive observations. This measure is utilized later in the event models
as a means to control the amount of noisy samples permitted in the stationary segments.
Event model ℳଵ: In this event model, a sequence of samples ‫܆‬ is defined as an event if
(a) it does not contain any noisy samples (i.e. ‫ܥ‬଴ = ߶),
(b) it contains two clusters ‫ܥ‬ଵ and ‫ܥ‬ଶ (i.e.݉ = 2),
(c) both clusters do not interleave (overlap) in the time domain1
,
(i.e. ∃‫ݑ‬ ∶ ࢞௡ ∈ ‫ܥ‬ଵ ∀(݊ ≤ ‫)ݑ‬ and ࢞௡ ∈ ‫ܥ‬ଶ ∀(݊ > ‫.))ݑ‬
This is the simplest event model without any outliers. It consists of two stationary segments
‫܆‬௦ଵ = ሾ࢞ଵ, ࢞ଶ, … , ࢞௨ሿ and ‫܆‬௦ଶ = ሾ࢞௨ାଵ, ࢞௨ାଶ, … , ࢞ேሿ. The segment ‫܆‬௧ = ሾ࢞௨, ࢞௨ାଵሿ (including
the last sample of ‫܆‬௦ଵ and the first one of ‫܆‬௦ଶ) is called the change-interval of the event and ‫ݑ‬ is
the change point. In other words, an event ℳଵ is a change interval of length two surrounded by
two noise-free weakly stationary segments. This model is valid for switch-off events of most
loads as well as switch-on events of resistive ones in a noise-free power signals.
1
For simplicity, and without loss of generality, we assume that the first and second stationary segments of an event are assigned to the cluster sets ‫ܥ‬ଵand
‫ܥ‬ଶ, respectively.

Figure 1(a) shows an example of a signal segment matching the first event model ℳଵ where the
scalar samples ‫ݔ‬௡ ∈ ℝ and their corresponding cluster indices ‫ݕ‬௡ = ߱(‫ݔ‬௡) ∈ ሼ1, 2ሽ are plotted
over time. The signal represents a step-like event that consists of two stationary segments (red,
solid) and a change interval (blue, dashed).
Event model ℳଶ: A sequence of samples ‫܆‬ is defined as an event if
a) it contains two clusters ‫ܥ‬ଵ and ‫ܥ‬ଶ (i.e. ݉ = 2) and the outliers set ‫ܥ‬଴ is not necessarily
empty allowing noisy samples,
b) both clusters ‫ܥ‬ଵ and ‫ܥ‬ଶ show a high temporal locality ratio, i.e.
‫ܥ(ܿ݋ܮ‬௜) ≥ 1 − ߳, for ݅ = 1, 2
c) both clusters do not interleave in the time domain, i.e.
∃‫,ݑ‬ ‫ݒ‬ > ‫:ݑ‬ ࢞௡ ∈ ‫ܥ‬଴ ∪ ‫ܥ‬ଵ ∀(݊ < ‫)ݑ‬ and ࢞௨ ∈ ‫ܥ‬ଵ, and
࢞௩ ∈ ‫ܥ‬଴ ∪ ‫ܥ‬ଶ ∀(݊ > ‫)ݒ‬ and ࢞௩ ∈ ‫ܥ‬ଶ
Compared with ℳଵ, this event model permits noisy samples (i.e. outliers) as well as a lengthy
transient interval. This, however, requires the utilization of a noise-aware clustering algorithm.
By definition, ࢞௡ ∈ ‫ܥ‬଴, ∀(‫ݑ‬ < ݊ < ‫.)ݒ‬ In this case, the event contains two stationary segments
‫܆‬௦ଵ and ‫܆‬௦ଶ consisting of samples belonging to ‫ܥ‬ଵ and ‫ܥ‬ଶ, respectively, and a change-interval
‫܆‬௧ = ሾ࢞௨, ࢞௨ାଵ, . . . , ࢞௩ିଵ, ࢞௩ሿ.
Figure 1: 1-dimensional signals highlighting differences between the three event models. (a) shows a step-like
event that is free of both outliers and a transient interval. In (b) random outliers as well as a transient interval
are permitted. (c) shows a repeated pattern of spikes that eventually cluster in ‫ܥ‬ଷ. Finally, (d) shows high
fluctuations in stationary segments leading to the third cluster ‫ܥ‬ଷ as well. The third event model ℳଷ fits all
segments, the second event model ℳଶ fits only (a) and (b), wherease the first model ℳଵ fits only (a).

Figure 1(b) shows an example of a signal segment matching the second event model M_2 (but
not the first one ℳଵ) where the event contains a slower transient interval in a noisy signal. Even
though ℳଶ is valid for most of the switch-on/off and state-change events within noisy signals, it
actually has one implicit assumption on the noise. The assumption that ݉ = 2 (maximally two
clusters representing two stationary segments) implies that the noise is random and does not
contain a repeated pattern that eventually builds up a cluster when projected to the PQ-plane. This
is not always the case as shown in the third example in Figure 1(c).
In the aggregate power signal, some appliances trigger a repeated, sometimes periodic, pattern of
high fluctuations or spikes. Such repeated patterns tackle the detection of other actual events.
This masking behaviour is resolved in the third event model.
Event Model ℳଷ: A sequence of samples ‫܆‬ is defined as an event if
(a) it contains at least two clusters ‫ܥ‬ଵ and ‫ܥ‬ଶ (i.e. ݉ ≥ 2) and the outliers set ‫ܥ‬଴ is not
necessarily empty,
(b) clusters ‫ܥ‬ଵ and ‫ܥ‬ଶ show a high temporal locality ratio, i.e.
‫ܥ(ܿ݋ܮ‬௜) ≥ 1 − ߳, for ݅ = 1, 2
(c) clusters ‫ܥ‬ଵ and ‫ܥ‬ଶ do not interleave in the time domain, i.e.
∃‫,ݑ‬ ‫ݒ‬ > ‫:ݑ‬ ࢞௡ ∉ ‫ܥ‬ଵ ∀(݊ > ‫)ݑ‬ and ࢞௨ ∈ ‫ܥ‬ଵ, and
࢞௩ ∉ ‫ܥ‬ଶ ∀(݊ < ‫)ݒ‬ and ࢞௩ ∈ ‫ܥ‬ଶ
In this model, the limitation on the clustering cardinality is released and therefore a repeated
noise pattern that eventually results in a wide (temporally wide) cluster would not mask events
occurring in the same interval. Similar to ℳଶ, the sequence in this model contains two stationary
segments ‫܆‬௦ଵ and ‫܆‬௦ଶ consisting of samples belonging to ‫ܥ‬ଵ and ‫ܥ‬ଶ respectively, and a change
interval consisting of ‫܆‬௧ = ሾ࢞௨, ࢞௨ାଵ, … , ࢞௩ିଵ, ࢞௩ሿ.
Figure 1 (b) and (c) show two event segments fit only by ℳଷ. Figure 1(a) shows the simplest
event which is fit by all defined models. In Figure 1(b), the transient period as well as the noisy
spikes can only be fit by ℳଶ and ℳଷ. Finally, the repeated noise patten in Figure 1(c) or high
fluctuations in Figure 1(d) only match the last event model ℳଷ.
2. DETECTION ALGORITHM
The main task of the event detection algorithm is to search for signal segments that match a given
event model ℳ௜. This is achieved by applying a clustering algorithm on different segments and
checking how much each segment matches the model. In all of the three models introduced in
section 2, the clustering cardinality ݉ is not known in advance. Therefore, a utilized clustering
algorithm should either be nonparametric or a model order estimation step has to take place
beforehand.
In our approach we utilized the commonly used Density-Based Spatial Clustering of Applications
with Noise (DBSCAN) algorithm [15]. The DBSCAN algorithm (or density-based clustering in
general) has several advantages that make it the best candidate for a non-parametric sequential
event detection. First, DBSCAN assumes no prior knowledge of the number of clusters. Second,
DBSCAN supports arbitrarily shaped clusters with no constraints on their samples’ distribution.
In addition, DBSCAN is a noise-aware clustering algorithm and, therefore, can be utilized with
any of the previously defined event models.

Ideally, the detection algorithm searches the input signal sequentially for segments that match a
given event model. However, we control the matching process with a proximity measure that
shows how much a segment matches the given model.
Definition 3: The model loss between an event model ℳ௜ and a signal segment ‫܆‬ is defined as
ℒ(ℳ୧, ‫,܆‬ ‫,ݑ‬ ‫)ݒ‬ = |ሼ࢞௡:݊ ≤ ‫ݑ‬ and ࢞௡ ∈ ‫ܥ‬ଶሽ| +
|ሼ࢞௡: ݊ ≥ ‫ݒ‬ and ࢞௡ ∈ ‫ܥ‬ଵሽ| + (7)
|ሼ࢞௡: ‫ݑ‬ < ݊ < ‫ݒ‬ and ࢞௡ ∈ ‫ܥ‬ଵ ∪ ‫ܥ‬ଶሽ|
where ‫ݑ‬ and ‫ݒ‬ are the indices of the first and last sample of the change-interval, respectively. In
the case of ℳଵ where ‫ݒ‬ = ‫ݑ‬ + 1, the last term in Equation 7 becomes zero regardless of ‫.ݑ‬
The model loss function counts the number of samples that need to be corrected (i.e. reassigned
to a different set ‫ܥ‬௝ of the clustering structure) in order for the segment ‫܆‬ to match the event
model ℳ௜. The lower the loss, the more the signal segment matches the event model.
The proposed detection algorithm can then be presented as to two sub-tasks, the forward
detection step which is the main process for finding an event, and the backward reduction step
that is responsible for a more accurate segmentation.
In the forward detection step, new samples are received one at a time and inserted into the
clustering space. Upon insertion of a new sample, the clustering indices are updated and the
model loss is re-estimated. Once a match is encountered (i.e. the model loss is zero or less than a
predefined threshold ߣ), a detection is declared with the current change point ‫ݑ‬ of the matched
segment and the change-interval ‫܆‬௧ = ሾ࢞௨, ࢞௨ାଵ, … , ࢞௩ିଵ, ࢞௩ሿ where ࢞௩ is the first sample of the
second stationary segment.
Once an event is declared, the backward reduction step begins. In this step, samples are removed
from the clustering space in a First-In-First-Out (FIFO) fashion while updating the clustering
structure upon each deletion and re-estimating the model loss. The reduction ends by the last
sample that satisfies the matching condition (i.e. if that sample is deleted, the segment will no
longer matches the event model within the predefined threshold loss ߣ). The complete detection
algorithm can be described as follows. Given an event model ℳ௜
1. Receive new sample ࢞ேାଵ and append it to ‫܆‬
2. Update the clustering vector ࢟ and the clustering structure ൛‫ܥ‬௝ൟ௝ୀଵ
௠
3. Check ℒ(ℳ୧, ‫,܆‬ ‫,ݑ‬ ‫)ݒ‬ ≤ ߣ for all ‫,ݑ‬ ‫,ݒ‬ if not satisfied, go to step (1)
4. Declare event detection with change-interval ‫܆‬௧ = ሾ࢞௨, ࢞௨ାଵ, … , ࢞௩ିଵ, ࢞௩ሿ and change-point is
‫ݑ‬ where ‫ݑ‬ and ‫ݒ‬ result in the minimum model loss between ℳ௜ and the current segment ‫܆‬
(i.e. argmin௨,௩ ℒ(ℳ୧, ‫,܆‬ ‫,ݑ‬ ‫.))ݒ‬
5. Delete oldest sample ࢞ଵ from the segment
6. Update the clustering vector ࢟ and the clustering structure ൛‫ܥ‬௝ൟ௝ୀଵ
௠
7. Check ℒ(ℳ୧, ‫,܆‬ ‫,ݑ‬ ‫)ݒ‬ ≤ ߣ for all ‫,ݑ‬ ‫,ݒ‬ if satisfied, go to step (5)
8. Re-insert last sample and declare current segment ‫܆‬ as a balanced event.

After each detection, the process restarts from the first sample of the second stationary segment
࢞௩. The main objective of the backward reduction step is to extract balanced stationary segments
(i.e. |‫ܥ‬ଵ| ≈ |‫ܥ‬ଶ|) around the transient interval. Balanced segments lead to more stable steady-state
features as well as an enhanced robustness to missed detections (i.e. false negatives).
2. EXPERIMENTS AND RESULTS
The proposed event detection approach has been evaluated on different power datasets among
them is the Building-Level fUlly labelled Electricity Disaggregation (BLUED) dataset [16]. In
the following, we show the results of applying the event detection algorithm with event model
ℳଷ and the DBSCAN clustering scheme on the BLUED dataset. We only show evaluation of
detection results. Evaluation of the accuracy of transient interval segmentation and the stability of
extracted features is beyond the scope of this paper.
Table 1 shows the event detection results on the real and reactive power signals from the BLUED
dataset. BLUED include aggregate measurements from a two-phase residential building (phase A
and B) and each is evaluated separately. True Positives (TP) is the number of successful
detections, False Positives (FP) is the number of detections that do not correspond to actual
events, while False Negatives (FN) is the number of missed events. Finally, False Positive
Percentage (FPP), precision, recall, and the F1-score measures are defined as
FPP =
‫ܲܨ‬
‫ܧ‬
(8)
precision =
ܶܲ
ܶܲ + ‫ܲܨ‬
(8)
recall =
ܶܲ
ܶܲ + ‫ܰܨ‬
(9)
‫ܨ‬ଵ − score =
2 × precision × recall
precision + recall
(10)
where ‫ܧ‬ is the number of events. Results show highly precise detection rates where the number
of false positives is relatively low in both phases. It is also observed that, noise in the second
phase (phase B) still masks a relatively large number of events.
Table 1. Event detection results on BLUED [16] dataset.
Phase A Phase B Total
Number of events ‫ܧ‬ 892 1609 2501
Number of detections 874 1176 2050
True Positives (TP) 867 1097 1964
False Positives (FP) 7 79 86
False Negatives (FN) 25 512 537
FPP 0.78% 4.91% 3.44%
precision ૢૢ. ૛૙% ૢ૜. ૛ૡ% ૢ૞. ૡ૚%
recall (TPR) ૢૠ. ૛૙% ૟ૡ. ૚ૡ% ૠૡ. ૞૜%
‫ܨ‬ଵ-score ૢૡ. ૚ૢ% ૠૡ. ૠૡ% ૡ૟. ૜૚%

3. CONCLUSIONS
We introduced a novel clustering-based approach for sequential event detection. The proposed
algorithm features accurate segmentation of the stationary and non-stationary intervals for more
stable feature extraction, support of arbitrarily shaped stationary segments with no prior
assumptions on their sample distribution, and more robustness to noise as well as parameter
variations.
REFERENCES
[1] G. W. Hart, “Nonintrusive appliance load monitoring”, in proceedings of the IEEE: vol.80, no.12, pp.
1870-1891, Dec. 1992. doi:10.1109/5.192069
[2] A. I. Cole and A. Albicki, "Data extraction for effective non-intrusive identification of residential
power loads", in proceedings of the Instrumentation and Measurement Technology Conference
(IMTC) 1998 IEEE: vol.2, pp.812-815, May 1998. doi:10.1109/IMTC.1998.676838
[3] S. Drenker and A. Kader, "Nonintrusive monitoring of electric loads", in Computer Applications in
Power, IEEE: vol.12, no.4, pp.47-51, Oct 1999. doi:10.1109/67.795138
[4] M. El Hachemi Benbouzid, “A review of induction motors signature analysis as a medium for faults
detection”, IEEE Transactions on Industrial Electronics, vo.47, no.5, pp.984-993, Oct 2000.
[5] S. N. Patel, T. Robertson, J. A. Kientz, M. S. Reynolds, and G. D. Abowd, “At the Flick of a Switch:
Detecting and Classifying Unique Electrical Events on the Residential Power Line”, in UbiComp
2007: Ubiquitous Computing, vol.4717, pp.271-288, 2007.
[6] M. Basseville and I. V. Nikiforov. Detection of Abrupt Changes: Theory and Application. Prentice
Hall, 1993.
[7] M. Berges, E. Goldman, L. Soibelman, H. S. Matthews, and K. Anderson, “User-centred non-
intrusive electricity load monitoring for residential buildings”, Journal of Computing in Civil
Engineering, vol.25, no.1, 2011.
[8] K. D. Anderson, M. E. Berges, A. Ocneanu, D. Benitez, and J. M. F. Moura, “Event Detection for
Non-Intrusive Load Monitoring”, in IECON 2012, 38th Annual Conference on IEEE Industrial
Electronics Society, October 2012.
[9] K. N. Trung, E. Dekneuvel, B. Nicolle, and O. Zammit, “Event Detection and Disaggregation
Algorithms for NIALM System”, in the 2nd International Non-Intrusive Load Monitoring (NILM)
Workshop, Jun 2014.
[10] Y. Jin, E. Tebekaemi, M. Berges, and L. Soibelman, “A time-frequency approach for event detection
in nonintrusive load monitoring” in proceedings of the Signal Processing, Sensor Fusion, and Target
Recognition, Orlando, Florida, USA, 2011.
[11] M. Volpi, D. Tuia, G. Camps-Valls, and M. Kanevski, "Unsupervised Change Detection With
Kernels", in Geoscience and Remote Sensing Letters, IEEE: vol.9, no.6, pp.1026-1030, Nov. 2012.
doi: 10.1109/LGRS.2012.2189092
[12] M. Luong, V. Perduca, and G. Nuel, “Hidden Markov Model Applications in Change-Point
Analysis”, arXiv Journal, preprint arXiv:1212.1778v1, 2012.

[13] G. L. Grinblat, L. C. Uzal and P. M. Granitto, “Abrupt change detection with one-class time-adaptive
support vector machines”, Expert Systems with Applications Journal, vol. 40, pp. 7242–7249, 2013.
[14] S. B. Leeb, S. R. Shaw, and Jr. Kirtley, J. L. “Transient event detection in spectral envelope estimates
for nonintrusive load monitoring”, IEEE Transactions on Power Delivery, 10(3):1200–1210, July
1995.
[15] M. Ester, H. P. Kriegel, J. Sander, and X. Xu, “A density based algorithm for discovering clusters in
large spatial databases with noise” in proceedings of Knowledge Discovery and Data mining (KDD),
1996.
[16] K. Anderson, A. Ocneanu, D. Benitez, D. Carlson, A. Rowe, and M. Berges, “BLUED: a fully
labeled public dataset for Event-Based Non-Intrusive load monitoring research”, in proceedings of
the 2nd Knowledge Discovery and Data mining (KDD) Workshop on Data Mining Applications in
Sustainability (SustKDD), Beijing, China, August 2012. URL: http://nilm.cmubi.org/

SEQUENTIAL CLUSTERING-BASED EVENT DETECTION FOR NONINTRUSIVE LOAD MONITORING

More Related Content

What's hot (19)

Viewers also liked (14)

Similar to SEQUENTIAL CLUSTERING-BASED EVENT DETECTION FOR NONINTRUSIVE LOAD MONITORING (20)

Recently uploaded (20)

SEQUENTIAL CLUSTERING-BASED EVENT DETECTION FOR NONINTRUSIVE LOAD MONITORING