Deep Learning based Smart Meter Data Analytics

for Electricity Load Prediction

1 2
J Jeyaranjani D Devaraj
Department of Computer Science and Engineering Department of Electrical and Electronic Engineering
Kalasalingam Academy of Research and Education Kalasalingam Academy of Research and Education
Virudhunagar, India Virudhunagar, India
[email protected] [email protected]

Abstract—In the smart grid era, Smart Meter data is be a suitable technique to be adopted. In this lane, in recent
playing a vital role. The residential load forecasting is one of years deep learning method achieves the state-of-art
the major issue that uses the Smart Meter data. It is high performance.
volatile load data due to its nature of variability in load The supervised learning algorithm of machine learning
pattern. The load pattern of every customer varies with
includes the Neural Network (NN) technology as one of
respect to certain factors like demographic information,
environmental factors, time stamp, etc. Load pattern-based the statistical learning using in various application for
customer clustering, aggregated load value based customers performing forecasting. The NN models the non-linearity
classification are certain conventional method used for in the input data. It approximates any complex function
addressing the load forecasting. It is less likely to address with arbitrary precision. Deep learning methods are ANN
forecasting issue using the conventional method. To (Artificial Neural Network) based representation learning.
overcome this downside, this research proposes the deep Deep learning architecture is composed of deep processing
learning based load forecasting method that uses the layers as in ANN. It composes non-linear modules that
aggregated smart meter dataset along with the demographic each module transforms into next level to obtain the final
dataset. The performance of the deep neural network (DNN)
result [5]. It learns non-linear and complicated relationship
is compared with shallow network that outperform with the
error calculation. with the number of layers. The conventional shallow
architecture models utilize clustering, aggregation, or
Keywords—Smart Meter, Load Forecasting, Deep spectral analysis to rule out the [6] problem.
Learning, RNN Umut Ugurlu et al. [7] propose the multi-layer Gated
Recurrent Units as a new technique for electricity price
I. INTRODUCTION forecasting. Weicong Kong et al. [8] uses real-world data
The renewable energy sources are widely recognized to prove LSTM based deep learning forecasting
electricity supply units. But it is less likely to be dependent framework with appliance consumption sequences
for Demand-side management. The reason for this is their improve the forecast accuracy. Muhammad Qamar Raza
variability and intermittency in the energy output. Also, et al. [AI load forecast] show the potential of AI
the major factor of electricity is storage. The way of techniques for effective load forecasting in order to
storing the produced electricity is difficult for utility achieve the concept of smart grid and buildings. Wan He
companies. Recent research focuses on electricity load [8] tried out the Convolution Neural Network components
forecasting for the efficient operation of the utility grid in to extract rich features from historical load sequence and
planning, managing, and marketing. The Forecasting is use Recurrent Components to model the implicit
done on three categories such as short term, medium term dynamics. Yayu Peng et al. [9] Short term load
and long term. The short term forecasting is the popular forecasting techniques, from the conventional linear
category where The outcome of the forecast is widely used regression to deep learning, are implemented for a
for DR (Demand Response) program, Grid communication detailed comparative analysis to verify the forecasting
Network management, Grid commitment, Price prediction, performances.
and customer load plan. In this paperwork, we implement DNN architecture
Most of the power utility companies are spotlighting for load forecasting. Various layers are tested to model the
that peak load reduction in dynamic time is a major network for identifying the better impact on forecasting.
situation that arises over a random period of time. The To improvise the performance, the aggregated smart
peak load reduction is Demand Side problem of utility meter reading is calculated from the historical data. As it
companies. The load estimation based on accurate load is a complex task to deal with time series data for load
prediction value will increase the system reliability in prediction during the peak time. The static data which
terms of time and cost. Peak load is criteria when there is influence more on the resident's load pattern is
an imbalance in load supplied and load utilized. The considered. The Deep neural network (DNN) with the
forecasting provides an increase in profit and reduction in required type and number of layers is modeled for
labor during peak is obtained. Conventional models like learning demographic data of the resident. The prediction
classical approaches and machine learning are available. of the future load is performed. This paper discusses the
The approaches are providing reliable results with features of residential smart meter reading and the
acceptable inferences. But the problem statement is demographic information of the resident. The patterns of
dealing with time series data. The algorithm/method that is residents load consumption that depends on demographic
capable of processing the time series data is considered to information.

978-1-7281-4407-8/19/$31.00 ©2019 IEEE

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 27,2020 at 02:31:11 UTC from IEEE Xplore. Restrictions apply.
 The resident behavior in electricity consumption and C. Demographic Data
its correlation to the demographic information is The lifestyle of the resident has more influence over
discussed. the electricity consumption value. The socio-economic
The rest of the paper is organized as follows. Section nature of the resident is revealed by their demographic
II discusses the components of the model. Section III load data. This data is obtained as survey answers from the
curve of the resident. Section IV represents the proposed smart meter residents. This data provide the information
DNN based load prediction. Section V presents results about the resident attributes like no of people at home,
and discussion. Finally, Section VI includes a conclusion. occupation, family status, no of appliances, the age
category of people at home, gender, family income, etc.
II. PROPOSED MODEL DESCRIPTION The frequency of change in demographic data will be less.
A. Block Diagram of the Proposed Model So this data is considered as static data. The utility
The artificial neural network is trained with the sample company may involve in updating the required changes in
dataset and it is tested with the generated load data’s static data. The Irish Social Science from smart meter
obtained from the smart meter. The stream of input data is achieves to provide the survey dataset comprised of
fed into the prediction module of neural network and the residents demographic information. It is obtained as a
prediction output is obtained as the error rate and accuracy questioner from the residents whose smart meter readings
of prediction. are available in the archive database. It includes 235
feature values and 3427 records. Each record is the
demographic data of a smart meter resident. The features
Resident’s raw dataset
are the questionnaires related to the electricity load
Electricity Load Demographic consumption
dataset dataset
D. DNN
Deep Neural Network refers to a deep learning process
that works with the Artificial Neural Network (ANN)
Data preprocessing Data
preprocessing
background. It is an instance of machine learning. The
ANN is widely implemented for prediction problem that
works in the pattern of the human brain. The ANN works
Processed dataset with 3 or more layers whereas DNN works with 5 and
even thousands of layers. The layers are increased in DNN
Electricity Load data Demographic dataset
to perform deep learning process with high-level feature
Aggregate load Clustering dataset. The network design is more complex in DNN
value calculated the dataset
rather than being shallow in ANN. The estimate of layers
in a shallow network is much lower than that of a deep
network. In shallow network, the features are the non-
ANN DNN linear combinations of input which are learned by hidden
neurons. In a deep network, the layers learn the non-linear
combinations of features. Thus it has a high non-linear
Compare error and
correlation.
accuracy The feed-forward neural network model performs a
sequence of operation where the output depends on its
Fig. 1 Proposed model
previous input only. It has no memory. The Recurrent
B. Smart Meter Data Neural Network popularly known as Long Short Term
The smart grid has three components. Automated Memory has internal memory. It uses the intermediate
Metering infrastructure is one of the components, also output for latter input. It provides a fully connected
named as Smart Meter. It is an electronic device that network also called as multilayer perceptron.
records electricity consumption and allows bi-direction E. Load Prediction
communication between the utility company and the smart Load forecasting is the major energy management
meter resident for monitoring and billing. The utility is the technique for the power system. A large number of
company that generates electricity and supply to the researches focus on load forecasting to predict the future
consumers. The meter records the energy in a periodic requirement of power from the utilities. This may prevent
time like hourly, every 30 minutes or 15 minutes. The the adverse effect of unpredictable change in lifestyle due
number of electricity consumption value received for per to the cut-down or scarcity of electricity. The Demand
resident per day is 48 values. The smart meter data is the Response is the program that utility company executes to
time series variety of data. The most dependent factors for manage various electricity issues like peak load reduction,
electricity consumption pattern of the resident are load prediction, manage tariff based on usage, etc.
environmental factors, socioeconomic factors. The
III. LOAD CURVE OF RESIDENTS
environmental factors include weather, temperature, and
A. Load Curve
humidity of the resident’s location. The aggregated smart
To achieve the best forecast output from the model, the
meter data values are calculated from two consecutive
load data should be analyzed for dynamic feature variation.
years. It uses two consecutive years smart meter dataset
Based on the analysis, the required operation is applied for
provided by Irish Social Science from smart meter
acquiring better forecast output. The data treatment like
achieve.
data preprocessing is applied to the load data and
demographic data.

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 27,2020 at 02:31:11 UTC from IEEE Xplore. Restrictions apply.
 The variation in the load curve of the resident, and similar data. Table I represents the cluster formed by a
Small and Medium Enterprise (SME) is not the same. A grouping of residents based on their data.
single resident load data itself will have a variety of load
curve patterns based on the season, weekday/ weekend/
vacation, etc. Whereas there will not be a drastic difference
in the load curve of SME. So the further research focus is
required for the residential load. So for our problem, the
residential load curve is analyzed. Fig.1and 2 represents
the load utilization of a sample resident and SME. The
smart meter collects the load value for every 15 minutes.
It is inferred from the curve that the variation load
utilization is static in the resident pattern of utilization. The
SME load curve infers the load utilization is gradual and it
is predictable over time without the requirement of any Fig. 4 Elbow curve to find the number of clusters
analytics overhead.

Fig. 5 Elbow curve to find the number of clusters

Fig.2 Sample resident load curve of a day TABLE I RESIDENT’S GROUPED INTO 12 CLUSTER
cluster Number 1 2 3 4 5 6 7 8 9 10 11 12
cluster Size 321 259 239 309 350 304 225 301 225 326 310 254

It is noted that the load data of residents of a cluster

is nearer to each. Also, the mean load value of every
cluster is satisfactory when compared to its resident load
value. From this, it is proved that the load curve of the
resident is influenced by their demographic information.
Fig. 3 Sample SME load curve of a day
IV. DNN BASED LOAD PREDICTION
B. Load Curve Influenced by Demographic Data DNN is designed for time series forecasting. It is
The load data of individual residents play a vital role in used to perform long short term memory forecasting
load prediction problem. The environmental data like (LSTM). Typical architecture designs of deep learning
temperature, humidity, wind speed, etc will also influence including,
the load value. The short term load forecasting problem Convolutional Deep Neural Networks (CNN), Deep
predicts the near future load (eg. Next 1 week or next day). Sparse
So the environmental factors are not influencing much on Autoencoder (DSA), Deep Recurrent Neural Networks
the prediction result. Whereas the impact of the (DRNN), Multi-Layer Perceptrons (MLP), Deep
demographic data of the corresponding individual resident Restricted
such as no of people, income, area of the home, no of Boltzmann Machines (DRBM), etc. [20]. The LSTM
rooms, etc. are high when compared to environmental approach is employed to obtain the best performance of
information. To obtain a better prediction result, the DRNN.
demographic data is considered along with the aggregated The network is trained with 80% of processed
smart meter data. The accuracy value of predicted output dataset and tested with remainig 20% of dataset to achieve
has a major role in smart grids and smart buildings. good accuracy percentage. The DNN is built by
To show the proof of concept, the residents are combining multiple types of components. There are
clustered based on their demographic data. Then the various toolkits available for quick learning capability
aggregated smart meter readings are compared within each through deep learning concept such as tensorflow, Keras,
group. The optimal number of clusters is specific to the H2O, etc. These toolkits are used to learn from different
dataset. An ideal method that works along with kmean input features to perform the load forecast task. Based on
clustering algorithm is used to s determine the number of this fact, instead of considering the aggregated load data
clusters. From Fig.3 it is inferred that fitting the model alone for the LSTM based forecast, we propose to input
above 12 clusters will not influence the result. So it is all available resident lifestyle data to train and predict.
decided to fix the cluster size as 12. Now, kmeans
unsupervised algorithm is applied to cluster the V. RESULTS AND DISCUSSIONS
demographic dataset with the cluster size of 12. Fig.4 Here, the assessment is done to exploit the potential
represents the dendrogram plot of the dataset. The of deep learning with respect to forecasting. The realistic
hierarchal clustering represents the relations between load data and its corresponding demographic data of the
domestic residents are used for the investigation. The

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 27,2020 at 02:31:11 UTC from IEEE Xplore. Restrictions apply.
 datasets utilized for this paper are from the Smart
Metering Electricity Customer Behavior Trials (CBTs)
initiated by Commission for Energy Regulation (CER) in
Ireland. The period during 2009 and 2010 with over 5000
Irish residential consumers and small and medium
enterprises (SMEs) involved. It has load data collected for
every half an hour in Kilo Watt per hour (kWh) from each
meter. Questionnaires and corresponding answers named
as survey dataset obtained from the meter residents is the
demographic dataset. These two dataset are used as the
Fig. 6 MSE of DNN
primary data sources of the experiment carried out.
To investigate Deep network performance, the
results are compared with the shallow network. The deep
network architecture is implemented with 5 layers. The
number of layers varies according to size of the dataset
and the expected accuracy percentage. The configuration
is adopted in training process, i.e., hidden units, activation
function, return sequence, batch size, verbose, training
epochs, metrics, loss, and optimizer. The network
architecture varies with differing network layers, from 1
layer shallow network to 5 layers deep network, to reveal
the improvement from ‘shallow’ to ‘deep’.
Layer (type) Output Shape Param # Fig. 6 MSE of A NN
dense_1 (Dense) (None, 10000) 2370000
dense_2 (Dense) (None, 1000) 10001000 The performance for disaggregation level is evaluated
dense_3 (Dense) (None, 1000) 1001000 with Root Mean Square Error value. The performance is
dense_4 (Dense) (None, 1000) 1001000 shown in following Fig.6 and Fig 7. From the graph, it is
dense_5 (Dense) (None, 1) 1001 inferred that the error value of ANN is comparatively
higher than DNN.
TABLE I ERROR RATE ON EACH EPOCH
Epoch Time RMSE VI. CONCLUSION
1 9s 3ms/step loss: 274.2872 In this paper, the feasibility of DNN is validated with
2 5s 2ms/step loss: 0.3228
3 5s 2ms/step loss: 0.3228
two datasets. The influence of demographic data with
4 5s 2ms/step loss: 0.3228 respect to load forecasting is proved using the experiment
5 5s 2ms/step loss: 0.3228 result. The results obtained are also compared with the
6 5s 2ms/step loss: 0.3228 shallow network. The results reveal the great potentials
7 6s 2ms/step loss: 0.3228 and possibilities of deep learning in the following aspect:
8 6s 2ms/step loss: 0.3228
9 6s 2ms/step loss: 0.3228
i) Deep learning address large scale forecasting problem
10 6s 2ms/step loss: 0.3228 with number of deep learning layers ii) Deep learning has
11 6s 2ms/step loss: 0.3228 the capability to handle volume of data. The DNN is
12 6s 2ms/step loss: 0.3228 found to be the efficient technique to handle other load
13 6s 2ms/step loss: 0.3228 problems such as Demand Response, price prediction, and
14 6s 2ms/step loss: 0.3228 fault identification.
15 6s 2ms/step loss: 0.3228
16 6s 2ms/step loss: 0.3228
17 6s 2ms/step loss: 0.3228 REFERENCES
18 6s 2ms/step loss: 0.3228 [1] F. Javed, N. Arshad, F. Wallin, I. Vassileva, and E. Dahlquist,
19 6s 2ms/step loss: 0.3228 "Forecasting for demand response in smart grids: An analysis on
20 6s 2ms/step loss: 0.3228 use of anthropologic and structural data and short term multiple
loads forecasting," Applied Energy, vol. 96, pp. 150-160, 2012.
[2] A. Borghetti, M. Bosetti, S. Grillo, S. Massucco, C. A. Nucci, M.
The batch size is represented as ‘400 per epoch’, Paolone, et al., "Short-term scheduling and control of active
verbose is ‘1’, activation function is ‘relu’, distribution systems with high penetration of renewable resources,"
kernel_initializer is ‘uniform’, optimizer is ‘adam IEEE Systems Journal, vol. 4, pp. 313-322, 2010.
[3] A. Botterud, Z. Zhou, J. Wang, J. Sumaili, H. Keko, J. Mendes, et
optimizer’ and the network is compiled with 20 epochs. al., "Demand dispatch and probabilistic wind power forecasting in
Table II represent the loss value which is considered as unit commitment and economic dispatch: A case study of Illinois,"
Mean Square Error and the time taken by the network in IEEE Transactions on Sustainable Energy, vol. 4, pp. 250-261,
each epoch. The considered metrics is categorical 2013.
[4] B. F. Hobbs, S. Jitprapaikulsarn, S. Konda, V. Chankong, K. A.
entropy. Loparo, and D. J. Maratukulam, "Analysis of the value for unit
commitment of improved load forecasts," IEEE Transactions on
Power Systems, vol. 14, pp. 1342-1348, 1999.
[5] Stefan Hosein, Patrick Hosein, “Load Forecasting using Deep
Neural Networks” 978-1-5386-2890-4/17 IEEE xplore 2017. DOI:
10.1109/ISGT.2017.8085971
[6] Heng Shi, Minghao Xu, Ran Li, “Deep Learning for Household
Load Forecasting – A Novel Pooling Deep RNN” IEEE xplore,
DOI 10.1109/TSG.2017.2686012, 2016.

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 27,2020 at 02:31:11 UTC from IEEE Xplore. Restrictions apply.
 [7] Salah Bouktif, Ali Fiaz, Ali Ouni, Mohamed Adel Serhani, [14] H. S. Hippert, C. E. Pedreira, R. C. Souza, Neural networks for
“Optimal Deep Learning LSTM Model for Electric short-term load forecasting: A review and evaluation, IEEE
Load Forecasting using Feature Selection and Genetic Transactions
Algorithm: Comparison with Machine on power systems 16 (1) (2001) 44–55.
Learning Approac”, Energies 2018, 11, 1636; [15] J. G. Jetcheva, M. Majidpour, W.-P. Chen, Neural network model
doi:10.3390/en11071636 ensembles for building-level electricity load forecasts, Energy and
[8] Weicong Kong, Student Member, IEEE, Zhao Yang Dong, David Buildings 84 (2014) 214–223.
J. Hill, Fengji Luo, Yan Xu,” Short-Term Residential Load [16] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT
Forecasting based on Resident Behavior Learning” 0885-8950 Press, 2016, http://www.deeplearningbook.org.
2016 IEEE. [17] W. He, Deep neural network based load forecast, Comput. Model.
[9] Yayu Peng, Yishen Wang, Xiao Lu, Haifeng Li, Di Shi, Zhiwei New Technol 18 (3) (2014) 258–262.
Wang, Jie Li, “Short-term Load Forecasting at Different [18] G. M. U. Din, A. K. Marnerides, Short term power load forecasting
Aggregation Levels with Predictability Analysis” using deep neural networks, in: Computing, Networking and
arXiv:1903.10679 Communications (ICNC), 2017 International Conference on, IEEE,
[10] J. D. Gooijer and R. Hyndman, “25 years of time series 2017, pp. 594–598.
forecasting,”International Journal of Forecasting, vol. 22, pp. [19] Y. Bengio, I. J. Goodfellow, and A. Courville, "Deep learning," An
443–473, 2006. MIT Press book in preparation. Draft chapters available at
[11] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, http://www. iro. umontreal. ca/∼ bengioy/dlbook, 2015
and S. Bengio, “Why does unsupervised pre-training help deep [20] "Electricity Smart Metering Customer Behaviour Trials (CBT)
learning?” J. Mach. Learn. Res., vol. 11, Mar. 2010. Findings Report," The Commission for Energy Regulation, Dublin,
[12] S. Mishra, “Long short-term memory in recurrent neural Information Paper2011.
networks,” [21] D. L. Marino, K. Amarasinghe, M. Manic, Building energy load
Master’s thesis, National Institute of Technology Rourkela, 2008. forecasting using deep neural networks, in: Industrial Electronics
[13] T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, R. J. Society, IECON 2016-42nd Annual Conference of the IEEE, IEEE,
Hyndman, Probabilistic energy forecasting: Global energy 2016, pp. 7046–7051.
forecasting competition 2014 and beyond, International Journal of
Forecasting 32 (3) (2016) 896–913.

Authorized licensed use limited to: CALIFORNIA INSTITUTE OF TECHNOLOGY. Downloaded on September 27,2020 at 02:31:11 UTC from IEEE Xplore. Restrictions apply.