Conference PaperPDF Available

Designing a Machine Learning based Non-intrusive Load Monitoring Classifier

Conference Paper

Designing a Machine Learning based Non-intrusive Load Monitoring Classifier

Abstract and Figures

Non-Intrusive load monitoring provides the users with detailed information about the electricity consumption of their appliances and gives energy providers a better insight about the usage of their clients. It can also be used in improving care of elderly, legal services and optimizing energy consumption. While there is plenty of work in NILM appliance classification, in this paper we investigate the design tradeoffs in the process of developing a machine learning based classification model from the perspective of feature engineering, model selection and optimisation. Our work shows that well engineered features have a greater impact on model performance than the selection of the machine learning technique. According to the results, the improvement in f1 score between non-engineered and the proposed engineered features is up to 42% while improvement between the worst non-optimised model and the best optimised one is 19%.
Content may be subject to copyright.
Designing a Machine Learning based Non-intrusive Load
Monitoring Classifier
Leo Ogrizek1, Blaz Bertalanic1,2, Gregor Cerar1, Marko Meza2, Carolina Fortuna1
1Jozef Stefan Institute, Ljubljana, Slovenia
2Faculty of Electrical Engineering, University of Ljubljana, Slovenia
E-mail: lo7909@student.uni-lj.si
Abstract. Non-Intrusive load monitoring provides
the users with detailed information about the electricity
consumption of their appliances and gives energy providers
a better insight about the usage of their clients. It can also
be used in improving care of elderly, legal services and
optimizing energy consumption. While there is plenty of
work in NILM appliance classification, in this paper we
investigate the design tradeoffs in the process of develop-
ing a machine learning based classification model from
the perspective of feature engineering, model selection
and optimisation. Our work shows that well engineered
features have a greater impact on model performance than
the selection of the machine learning technique. Accord-
ing to the results, the improvement in f1 score between
non-engineered and the proposed engineered features is
up to 42% while improvement between the worst non-
optimised model and the best optimised one is 19%.
1 Introduction
Non-Intrusive load monitoring (NILM) is a process of
estimating the electricity consumption of individual ap-
pliances in a household from their combined electricity
consumption. By providing detailed information to the
user, NILM can help reduce energy usage by 5-15%[1].
It is also beneficial for the energy providers as the addi-
tional information allows them to tune the production of
electricity and adjust the pricing plans better.
Data about appliance usage has additional applica-
tions outside of energetics. It provides an insight into
daily activities of residents which can be used in vari-
ous fields including health and medical (discovering sleep
disorders, remotely monitoring elderly), legal (monitor-
ing curfews) and commercial (customer profiling). The
benefit of measuring electricity consumption on a single
point rather than on every device is ease of deployment
and lower installation and maintenance costs.
For application such as elderly care and curfews, events
triggered by the residents, for example a toaster being
used, need to be automatically detected and distinguished
from automatic events such as for example air condition-
ing being activated by a thermostat. When unusual, pos-
sibly anomalous patterns in the usage are detected, such
as decreased home activity, then an alert needs to be sent
out to their caretakers, allowing them to intervene sooner.
This is possible through the use of NILM classification
[2]. In a similar manner, when anomalous behaviour is
detected from people under house arrest, an alert can be
dispatched to the appropriate authorities automatically.
NILM can also be used for predicting trends in en-
ergy consumption. Having a future view allows energy
producers to prepare and tune the production in accor-
dance to the projected requirements. Accurate consump-
tion predictions reduce excess energy production and in
turn, reduce the negative effects on the environment. NILM
by itself can reduce power consumption by increasing
awareness, some NILM powered systems can also turn
off unneeded devices [3, 1]. Inside small communities
NILM allows for better implementation of smart energy
distribution systems and peer-to-peer energy trading, by
providing a local producer information about whether en-
ergy is needed or can be returned into the grid [4].
Machine learning (ML) techniques have been shown
to be suitable for addressing all NILM related problems
[5]. This is realized by formulating a NILM classifica-
tion problem to be solved and selecting high quality data
that can be used by the ML techniques to discriminate
between the target classes. The data, ML techniques, the
training and evaluation process result in a model able to
distinguish the target classes based on previously learned
patterns in the data.
In this paper we aim to understand the design trade-
offs for developing a NILM classifier. The contributions
of this paper are as follows:
We systematically analyze the performance of fea-
ture engineering, model selection and optimization
for NILM on the UK-Dale dataset.
We propose a best feature set constructed to cap-
ture the shape of the time series and we show it
performs by 42 percentage points better than the
baseline raw time series.
The paper is organized as follows. Section 2 presents
related work, Section 3 formulates the problem and iden-
tifies the dataset, Section 4 discusses the feature selection
process while Section 5 elaborates on the model selec-
tion. Finally, Section 6 concludes the paper.
2 Related Work
In the last decade a large body of work on NILM classi-
fication has been published. The traditional way of de-
veloping such classifiers used various signal processing
technique. For instance, in [6] the authors classify appli-
ances based on transient effects by using FFT transforms
on higher frequency measurements.
As the performance of the ML techniques improved,
they have been increasingly considered as potential tech-
niques for realising NILM. In [5], the authors provide
an overview of the field by analysing suitable features
and ML techniques for different types of data, i.e. low
frequency, high frequency, etc. They also provide an
overview on model selection. In [7] the authors develop
a KNN classifier that is trained using U/I trajectories as
features. In [8], the authors compare four algorithms for
disaggregating appliances using multi label classification.
More recently, authors are using deep learning to solve
the NILM classification problem. In [9] low frequency
data is used to detect on/off events. Classifications are
made based on these events using a deep neural network
classifier trained on average power consumption, mini-
mum and maximum time an appliance is powered on.
In [10] the authors use transient power signals to train
a convolutional neural network to classify devices while
in [11] a recurrent neural network is trained on denoised
data.
Our work complements [5] by providing a quantita-
tive consideration of the design trade-offs when design-
ing a NILM classifier using classical explainable machine
learning models.
3 Problem statement
We define our NILM classification problem as follows.
Given an input time series T representing energy mea-
surements from households, there is a function Φ, that
maps the time series to a set of target classes C represent-
ing different household appliances as in equation 1.
C= Φ(T)(1)
where the set of target classes is C = {computer mon-
itor, laptop, television, washer dryer, microwave, boiler,
toaster, kettle, fridge }. The assumption is that a disag-
gregation in individual time series is already performed
and the classification is done on the resulting time series.
The classifier Φis realized using classical machine learn-
ing techniques and the UK-DALE dataset to develop the
model able to discriminate between the classes. We con-
sider the following diverse but explainable set of tech-
niques: SVM, KNN, Random forest (RF), MLP, Logis-
tic regression (LR) from the scikit-learn library version
0.22.2. Deep learning methods were not used due to the
limited amount of data in the UK-DALE dataset.
3.1 Dataset summary
The UK DALE (Domestic Appliance-level Electricity) [12]
dataset contains 180000 measurements of power usage
per device, taken in 6 second intervals or 0.1667 Hz for
a total of 300 hours of measurements per device. The
devices in the dataset are common household appliances
which were already presented in section 3 . Data is spread
into 1 hour long segments, each dataset sample contains
a time series with 600 data points as depicted in Figure 1.
Measurements are disaggregated, taken with a smart plug
on every device [12].
Figure 1: Samples for each appliance, showing power in rela-
tion to time over a 1 hour interval.
3.2 Methodology
UK-Dale contains univariate time series that if used raw,
may lead to suboptimal model performance. A common
practice in the ML community is to construct synthetic
features, by statistical summaries, feature interactions, etc.
using the raw time series. As some of the considered
techniques, such as SVM and Logistic regression are sen-
sitive to scaling, we used standard scaling that ensures
that all the features will be relatively proportional.
During the model development process, we use 5-
fold cross validation realized using scikit-learn K-fold
cross validation. We use the selected ML technique with
default parameter configuration as a baseline and then
provide an optimized version using the scikit-learn im-
plementation of grid search. For the SVM, grid search
was optimizing the kernel and C value; for K-nearest-
neighbours it optimized the number of neighbours; for
MLP it optimized the solver, the initial learning rate and
the learning rate schedule; for Random forest the number
of estimators while for SVM and logistic regression the
regularization parameter.
The models were evaluated using standard classifica-
tion metrics: precision =T P
T P +F P ,recall =T P
T P +F N
and f1 = 2precisionrecall
precision+recall for every class and their mean
scores for every fold, where TP, FP and FN stand for true
positives, false positives and false negatives. The final re-
sults presented in this paper are the mean scores of all 5
folds.
4 Feature selection
By manually inspecting the appliances in the dataset, we
noticed that they differ in overall power consumption and
how constant their energy usage is (see Figure 1). For fea-
ture engineering we chose those features, that best cap-
ture the shape of the time series. The first created features
were the mathematical moments and the squares of their
values, minimum, maximum, peak-to-peak, sum and me-
dian. These features alone do not describe the energy
usage spikes very well, so a second series was created
from every sample with the derivatives of all the intervals
between measurements. The number of times this deriva-
tive changes sign, which will be referred to as nturns, de-
scribes how many power spikes a device has during the
measurement period. The sum of absolute values of all
the derivatives, which will be referred to as d abs de-
scribes the intensity of said spikes. Features were cal-
culated for each sample separately to avoid data leakage
between training and testing dataset.
Table 1: Comparison of feature sets using SVM.
Feature set Precision Recall f1
raw 0.486 0.460 0.421
min, max, median 0.746 0.664 0.661
mean, std 0.714 0.646 0.642
all 0.851 0.837 0.835
best 0.851 0.835 0.834
We performed a guided evaluation of the features de-
scribed in this section with the best results, the raw time
series as baseline and standard statistics summarized in
Table 1. Using SVM to explore the performance of vari-
ous feature vectors, our experiments showed that the most
useful features are those that describe the shape of the
time series. Inputting raw data which makes every times-
tamp a feature yields an f1 score of 0.421. The results
are better than random because of the difference in size
of the features within different classes.
Using the minimum, maximum and median value from
the time series results in an f1 score of 0.661. These
results are significantly better because they describe the
shape of the data, rather than assuming a timestamp is rel-
evant to the end results. Using synthetic features, shifting
the series would still result in the same features, unlike
when raw data is used. Because measurements can be
started/finished at any time, this is a better classifier. For
the same reason, using the mean and standard deviation
also yields better results than raw data: f1 score of 0.642.
Using all synthetic features results in an f1 score of 0.835.
This is considerably better because the features describe
more properties of the time series. However, some of
the features are redundant and can be ignored to improve
computation time.
The final feature set determined to be the best is [mean,
skew, kurtosis, peak-to-peak, variance, nturns, d abs and
standard deviation]. This means the baseline feature vec-
tor of length 600 was reduced to a feature vector of length
8. These features give almost identical results as using all
the synthetic features, with an f1 score of 0.834, but with
lower required computing performance.
Table 2: Per class performance, SVM, best feature set
class precision recall f1
computer monitor 0,882 0,720 0,780
laptop computer 0,897 0,800 0,838
television 0,958 0,927 0,941
washer dryer 0,952 0,700 0,804
microwave 0,679 0,710 0,687
boiler 0,945 0,937 0,940
toaster 0,705 0,940 0,806
kettle 0,684 0,806 0,739
fridge 0,961 0,980 0,970
A per class breakdown shown in Table 2 of the best
results reveals that the fridge, boiler and television are
classified best, while microwave and kettle are classified
poorly. This is because the former have very distinct
energy consumption patterns (fridge has low consump-
tion and it is usually turned off can be distinguished from
small devices by large and long lasting energy consump-
tion spikes, boiler has a low consumption when not in use
but consumes a lot of power when turned on, television
uses a medium amount of energy with much more empha-
sised variations in consumption), while the latter share
properties between each other and other low consump-
tion devices such as the toaster and monitor (all are pow-
ered on for short periods of time, they also share highly
variable energy usage).
Certain samples contain devices that are powered off
throughout the whole sample, such samples can not be
distinguished from one another because they all have the
same characteristic of no consumption.
5 Model selection and optimisation
As discussed in Section 3.2 and summarized Table 3 the
model performance depends both on the selected ML tech-
nique as well as on the parameters. It can be seen from
the table that the best performing models are based on
MLP and SVM.
For the SVM, the default radial kernel proved to be
optimal, however the default C value of 1 was far too
Table 3: Impact of algorithm optimisation
Algorithm Precision Recall f1
SVM default 0.704 0.672 0.647
SVM optimised 0.851 0.835 0.834
KNN default 0.784 0.776 0.772
KNN optimised 0.787 0.776 0.776
MLP default 0.786 0.769 0.766
MLP optimised 0.849 0.830 0.828
LR default 0.742 0.691 0.674
LR optimised 0.788 0.770 0.768
RF default 0.822 0.815 0.813
RF optimised 0.822 0.815 0.813
low, resulting in a separating hyperplane with a larger
margin, that misclassifies more points. After optimisa-
tion the best C value was determined to be around 38000.
This resulted in a significant increase in accuracy with the
f1 score going from 0.647 to 0.834 as can be seen from
the first two rows of the table.
In K-nearest-neighbours grid search found that check-
ing 4 closest neighbours is better than the default 5, how-
ever this resulted in only a very small improvement of f1
from 0.772 to 0.776 as can be seen from rows 3 and 4 of
the table.
MLP initially had convergence issues. The first can-
didates for optimisation were thus the initial learning rate
and the learning rate schedule. Different solvers were
also tested. After optimisation, the parameters were changed
from default adam solver with initial learning rate of 0.001
and constant learning rate to sgd solver with 0.45 initial
learning rate and adaptive learning rate. We used one hid-
den layer with 100 neurons. Optimisation resulted in a
large increase in f1, from 0.766 to 0.828 as can be seen
from rows 5 and 6 of the table.
Like SVM, logistic regression also suffered from a
low C value. After optimisation, it was changed from the
default of 1 to 20000. The solver was changed from lbfgs
to liblinear. This resulted in a large increase in f1 going
from 0.674 to 0.768 as can be seen from rows 7 and 8 of
the table.
In random forest the number of estimators was changed
from 100 to 500 but that only resulted in a change of f1
from 0.813 to 0.815 as can be seen from rows 9 and 10 of
the table.
6 Conclusions
In this paper we showed the tradeoffs in developing an
accurate appliance classification system and proposed a
new feature set for improved classifier performance.
First, we showed that by using statistical and custom
quantities able to capture the shape of the raw time series
can improve the F1 score performance by up to 42 per-
centage points compared to the raw time series baseline.
Our experiments showed that the best performing fea-
tures were mean, skew, kurtosis, peak-to-peak, variance,
the number of times the derivative changed sign on the
observed interval, sum of absolute values of the deriva-
tive on every interval within the time series and standard
deviation.
Second, we also showed that the choice of the ma-
chine learning technique and the optimal parameters are
also important. The best preforming model using opti-
mised SVM has an f1 score 0.187 better than the worst
performing non optimized SVM. However, all optimized
SVM, MLP and random forest models work well with f1
scores between 0.81 and 0.84 for the considered problem.
References
[1] S. Darby, “THE EFFECTIVENESS OF FEEDBACK ON
ENERGY CONSUMPTION,” p. 24, 2006.
[2] J. Alcal´
a, J. Ure˜
na, and A. Hern´
andez, “Activity supervi-
sion tool using Non-Intrusive Load Monitoring Systems,
Sep. 2015, pp. 1–4, iSSN: 1946-0759.
[3] I. Abubakar, S. N. Khalid, M. W. Mustafa, H. Shareef, and
M. Mustapha, “Application of load monitoring in appli-
ances’ energy management – A review,Renewable and
Sustainable Energy Reviews, vol. 67, pp. 235–245, Jan.
2017.
[4] “Smart home energy management systems survey,” Nov.
2014, pp. 167–173.
[5] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajasegarar,
“Non-intrusive load monitoring approaches for disaggre-
gated energy sensing: A survey,Sensors, vol. 12, no. 12,
pp. 16 838–16866, 2012.
[6] S. R. Shaw, S. B. Leeb, L. K. Norford, and R. W.
Cox, “Nonintrusive load monitoring and diagnostics in
power systems,IEEE Transactions on Instrumentation
and Measurement, vol. 57, no. 7, pp. 1445–1454, 2008.
[7] A. Kelati, H. Gaber, J. Plosila, and H. Tenhunen, “Im-
plementation of non-intrusive appliances load monitoring
(NIALM) on k-nearest neighbors (k-NN) classifier,AIMS
Electronics and Electrical Engineering, vol. 4, no. 3, pp.
326–344, 2020.
[8] D. Li and S. Dick, “Whole-house Non-Intrusive Appli-
ance Load Monitoring via multi-label classification,” in
2016 International Joint Conference on Neural Networks
(IJCNN), Jul. 2016, pp. 2749–2755, iSSN: 2161-4407.
[9] M. A. Devlin and B. P. Hayes, “Non-intrusive load moni-
toring and classification of activities of daily living using
residential smart meter data,” IEEE Transactions on Con-
sumer Electronics, vol. 65, no. 3, pp. 339–348, 2019.
[10] D. d. Paiva Penha and A. R. Garcez Castro,
“Home Appliance Identification for Nilm Systems
Based on Deep Neural Networks,” IJAIA, vol. 9,
no. 2, pp. 69–80, Mar. 2018. [Online]. Available:
http://aircconline.com/ijaia/V9N2/9218ijaia06.pdf
[11] J. Kim, T.-T.-H. Le, and H. Kim, “Nonintrusive
Load Monitoring Based on Advanced Deep Learn-
ing and Novel Signature, Computational Intelli-
gence and Neuroscience, vol. 2017, p. e4216281,
Oct. 2017, publisher: Hindawi. [Online]. Available:
https://www.hindawi.com/journals/cin/2017/4216281/
[12] J. Kelly and W. Knottenbelt, “The uk-dale dataset, domes-
tic appliance-level electricity demand and whole-house
demand from five uk homes,Scientific data, vol. 2, no. 1,
pp. 1–14, 2015.
... For instance, in [4] they used ML to find patterns in the data and extract useful information such as type of load, electricity consumption detail and the running conditions of appliances [4]. More recently, [5] focused on the study of design trade-offs during the feature and model selection steps of the development of the ML-based classifier for ILM. In their study they considered various statistical summaries for feature engineering and classical machine learning techniques for model selection. ...
... In their study they considered various statistical summaries for feature engineering and classical machine learning techniques for model selection. We complement the work in [5] by extending the feature set with additional shape capturing values and considering deep learning (DNN) and gradient boosted trees (XGBoost) as promising modelling techniques. The contributions of this paper are as follows: ...
... • We show that our best DNN based model slightly outperforms our best XGBoost by 2 percentage points at the expense of increased training time. We also show that our models outperform the results from [5] by 5 percentage points. ...
Conference Paper
Full-text available
Appliance load monitoring (ALM) is a technique that enables increasing the efficiency of domestic energy usage by obtaining appliance specific power consumption profiles. While machine learning have been shown to be suitable for ALM, the work on analyzing design trade-offs during the feature and model selection steps of the ML model development is limited. In this paper we show that 1) statistical features capturing the shape of the time series, yield superior performance by up to 20 percentage points and 2) our best deep neural network-based model slightly outperforms our best gradient descent boosted decision trees by 2 percentage points at the expense of increased training time.
Article
Full-text available
This paper develops an approach for household appliance identification and classification of household Activities of Daily Living (ADLs) using residential smart meter data. The process of household appliance identification, i.e. decomposing a mains electricity measurement into each of its constituent individual appliances, is a very challenging classification problem. Recent advances have made deep learning a dominant approach for classification in fields such as image processing and speech recognition. This paper presents a deep learning approach based on multi-layer, feedforward neural networks that can identify common household electrical appliances from a typical household smart meter measurement. The performance of this approach is tested and validated using publicly-available smart meter data sets. The identified appliances are then mapped to household activities, or ADLs. The resulting ADL classifier can provide insights into the behaviour of the household occupants, which has a number of applications in the energy domain and in other fields.
Article
Full-text available
Monitoring electricity consumption in the home is an important way to help reduce energy usage. Nonintrusive Load Monitoring (NILM) is existing technique which helps us monitor electricity consumption effectively and costly. NILM is a promising approach to obtain estimates of the electrical power consumption of individual appliances from aggregate measurements of voltage and/or current in the distribution system. Among the previous studies, Hidden Markov Model (HMM) based models have been studied very much. However, increasing appliances, multistate of appliances, and similar power consumption of appliances are three big issues in NILM recently. In this paper, we address these problems through providing our contributions as follows. First, we proposed state-of-the-art energy disaggregation based on Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) model and additional advanced deep learning. Second, we proposed a novel signature to improve classification performance of the proposed model in multistate appliance case. We applied the proposed model on two datasets such as UK-DALE and REDD. Via our experimental results, we have confirmed that our model outperforms the advanced model. Thus, we show that our combination between advanced deep learning and novel signature can be a robust solution to overcome NILM’s issues and improve the performance of load identification.
Article
Full-text available
Many countries are rolling out smart electricity meters. These measure a home's total power demand. However, research into consumer behaviour suggests that consumers are best able to improve their energy efficiency when provided with itemised, appliance-by-appliance consumption information. Energy disaggregation is a computational technique for estimating appliance-by-appliance energy consumption from a whole-house meter signal. To conduct research on disaggregation algorithms, researchers require data describing not just the aggregate demand per building but also the `ground truth' demand of individual appliances. In this context, we present UK-DALE: an open-access dataset from the UK recording Domestic Appliance-Level Electricity at a sample rate of 16 kHz for the whole-house and at 1/6 Hz for individual appliances. This is the first open access UK dataset at this temporal resolution. We recorded from five houses, one of which was recorded for 655 days, the longest duration we are aware of for any energy dataset at this sample rate. We also describe the low-cost, open-source, wireless system we built for collecting our dataset.
Article
Full-text available
Appliance Load Monitoring (ALM) is essential for energy management solutions, allowing them to obtain appliance-specific energy consumption statistics that can further be used to devise load scheduling strategies for optimal energy utilization. Fine-grained energy monitoring can be achieved by deploying smart power outlets on every device of interest; however it incurs extra hardware cost and installation complexity. Non-Intrusive Load Monitoring (NILM) is an attractive method for energy disaggregation, as it can discern devices from the aggregated data acquired from a single point of measurement. This paper provides a comprehensive overview of NILM system and its associated methods and techniques used for disaggregated energy sensing. We review the state-of-the art load signatures and disaggregation algorithms used for appliance recognition and highlight challenges and future research directions.
Article
Nonintrusive Appliance Load Monitoring (NIALM) is used to analyze individual’s house energy consumption by distinguishing variations in voltage and current of appliances in a household. The method identifies load consumption of each appliance from the aggregated home energy consumption. NIALM will also provide information of load consumptions of each appliance by indirectly detecting the abnormal changes of appliance usage. The proposed NIALM approach is based on features extraction from load consumptions measurements of electrical power signals in order to classify appliance’s state of operation. In this work, we have improved the identification accuracy and the detection of appliances based on their operational state by employing Machine Learning (ML) technique; namely k-nearest neighbor (k-NN) classification algorithm. The dataset used to perform this process is from the publicly available (PLAID) of power, voltage and current signals of appliances from several houses. This is used as benchmark data set. The PLAID dataset is collected and processed for each appliance and our classification results based on k-NN algorithm achieved high accuracy and is able to gain cost-effective solution. In addition, the result shows that k-NN classifier is a proven as an efficient method for NIALM techniques when compared with other proposed different ML options. Based on the used dataset, the average F-score measure obtained using the k-NN classifier is 90%. Possible reasons behind these findings are discussed and areas for further exploration are proposed.
Article
Energy monitoring is one of the important aspects of the energy management, as such there is a need to monitor the power consumption of a premises before planning some of the technical measures to minimize the energy consumption. This paper presents the current state of art of appliances’ energy management through Intrusive Load Monitoring (ILM) and Non-intrusive Load Monitoring (NILM), ILM is referred to as distributive sensing and NILM is based on single point sensing. In an effort to explore the recent trend in Energy Management for researchers in the area, various techniques of Home Energy Management (HEM) using ILM and NILM have been analyzed and categorized. Different contributions made by the researchers have been highlighted together with some techniques of reducing power consumption of a building for cost saving and greener environment. This study reveal that there are issues concerning load monitoring and load management that need the attention; issues such as more accurate recognition and the need for monitoring system that can recognize as many kind of loads as possible. More efforts to apply NILM in appliances’ energy management is also required. Finally there is need for promoting the energy management culture among the electricity consumers whether in homes, offices or industries. The study will help the researchers in the area to have a clear vision of the recent trends in the area.
Article
This paper describes a transient event classification scheme, system identification techniques, and implementation for use in nonintrusive load monitoring. Together, these techniques form a system that can determine the operating schedule and find parameters of physical models of loads that are connected to an AC or DC power distribution system. The monitoring system requires only off-the-shelf hardware and recognizes individual transients by disaggregating the signal from a minimal number of sensors that are installed at a central location in the distribution system. Implementation details and field tests for AC and DC systems are presented.