Received March 2, 2020, accepted March 14, 2020, date of publication March 18, 2020, date of current version

March 30, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.2981817

A Deep Learning Method for Short-Term

Residential Load Forecasting in Smart Grid
YE HONG 1 , YINGJIE ZHOU 1 , (Member, IEEE), QIBIN LI 2,

WENZHENG XU 1 , (Member, IEEE), AND XIUJUAN ZHENG 3

1 College of Computer Science, Sichuan University, Chengdu 610065, China
2 Key Laboratory of Low-Grade Energy Utilization Technologies and Systems, Ministry of Education, College of Energy and Power Engineering, Chongqing

University, Chongqing 400044, China

3 College of Electrical Engineering, Sichuan University, Chengdu 610065, China

Corresponding authors: Yingjie Zhou (yjzhou@[Link]) and Qibin Li (qibinli@[Link])

This work was supported by the National Natural Science Foundation of China under Grant 61801315 and Grant 51876015.

ABSTRACT Residential demand response is vital for the efficiency of power system. It has attracted
much attention from both academic and industry in recent years. Accurate short-term load forecasting is
a fundamental task for demand response. While short-term forecasting for aggregated load data has been
extensively studied, load forecasting for individual residential users is still challenging due to the dynamic
and stochastic characteristic of single users’ electricity consumption behaviors, i.e., the variability of the
residential activities. To address this challenge, this paper presents a short-term residential load forecasting
framework, which makes use of the spatio-temporal correlation existing in appliances’ load data through deep
learning. Multiple time series are conducted in the framework to describe electricity consumption behaviors
and their internal spatio-temporal relationship. And a method based on deep neural network and iterative
ResBlock is proposed to learn the correlation among different electricity consumption behaviors for short-
term load forecasting. Experiments based on real world measurements have been conducted to evaluate the
performance of the proposed forecasting approach. The results show that both the appliances’ load data
and iterative ResBlocks can help to improve the forecasting performance. Compared with existing methods,
measurements on Root Mean Squared Error, Mean Absolute Error and Mean Absolute Percentage Error
for the proposed approach are reduced by 3.89%-20.00%, 2.18%-22.58% and 0.69%-32.78%. In addition,
further experiments are conducted to evaluate the impact of using appliances’ load data, iterative ResBlocks
as well as other factors for the proposed approach.

INDEX TERMS Smart grid, short-term load forecasting, deep learning, residential load forecasting, iterative
ResBlocks.

I. INTRODUCTION demand will exceed the capacity of system in a residential

Energy is the one of the main drivers of human activity. community, the power company can incentivize residential
Demand response is of crucial importance for maintain- users to shift their power consumption by hiking up the
ing the reliable and efficient operation of the smart grid electricity price [5]. On the other hand, STLF can help to
system [1]–[4]. In residential power distribution, demand benefit the power company and residential users in economic
response can manage the power delivery from power sys- aspects. Based on the results of STLF, the power company can
tem to users and smooth the system load. Short-term load calculate optimal pricing strategies for residential electricity
forecasting (STLF) predicts users’ demand in the near future, usage [3]. For residential users, they can take appropriate
which provides the key information for making decisions in countermeasures if the STLF results indicate that there will
residential demand response. On one hand, STLF can help be a power shortage [6].
to satisfy users’ electricity demand and reduce the risk of Existing STLF methods could be categorized into two
outages. If the result of STLF indicates that users’ electricity kinds: aggregated load forecasting [7]–[9] and individual
users’ load forecasting [10]. Both of the forecasting methods
The associate editor coordinating the review of this manuscript and construct features based on historical records of load since
approving it for publication was Ahmed F. Zobaa . the value of short-term electricity consumption is related with

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see [Link]
VOLUME 8, 2020 55785
 Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

previous values. The aggregated load forecasting gives the More recently, researchers explored using deep learn-
estimation of the total electricity consumption for a group ing techniques to perform STLF for individual users, due
of users in a specific area, such as a city or a residential to its ability to extract the latent features of users’ elec-
community. tricity consumption behaviors and less domain knowledge
Short-term forecasting for aggregated load data has been requirements compared to traditional methods. Classical deep
extensively studied. Time series analysis has been applied on learning models are used for STLF, such as Deep Neu-
the STLF problem, including Auto Regressive Moving Aver- ral Network (DNN) [43], Extreme Learning Machine [36].
age (ARMA) based method [11]–[13], [37], [42] and Sup- Kong et al. [10] proposed a Long Short-Term Memory RNN
port Vector Regression (SVR) based method [14], [15], [45]. (LSTM-RNN) based framework for residential STLF. Exper-
Wei and Zhengang [11] combined the ARMA based method iment results showed that their method outperforms tradi-
and other statistical methods to improve the performance for tional machine learning methods. Shi et al. [21] proposed a
aggregated load forecasting. Pappas et al. [12] developed an pooling based deep RNN. The pooling stage uses load data of
offline model with enhanced the ARMA based method to neighbors to generate new features of inputs, which increases
predict the power usage provided by a power company in the data volume and helps to solve the over-fitting problem.
Greece. Huang and Shih [13] proposed a method that could Although the above methods made progresses in some
improve the forecasting accuracy through a modified ARMA aspects, the historical data they employed are the overall
based method with non-Gaussian process. Amini et al. [40] load data of single resident, which cannot include the spatio-
apply an ARIMA based model for electric vehicles’ demand temporal correlation existing among appliances’ load data.
forecasting. The method decouples conventional electrical The spatio-temporal correlation mentioned here is the spa-
load and charging demand of EV to forecast them indepen- tio correlation among electricity consumption behaviors of
dently, which can reduce forecasting errors. Yang et al. [14] different kind of appliances and the temporal correlation
presented an SVR based method to forecast power consump- between the historical electricity consumption behaviors and
tion at a city scale. They developed a grid search approach to future electricity consumption behaviors. For a single user,
automatically tune the model parameters, which can reduce the spatio correlation exists in the user’s electricity consump-
the difficulty in the parameter optimization phase. Velasco tion behaviors of different appliances. For instance, house-
et al. [15] presented a load forecasting model based on the hold members may have daily routines that using washing
SVR method for country-wide power usage. Ren [Link] [45] machine after taking a shower, or opening refrigerator before
employed an ensembled method for STLF, which consists of making a meal. The temporal correlation is the similarity
SVR, random forest and Xgboost. Other methods [16]–[20] of the historical electricity consumption behaviors and the
take extra conditions into account to improve the forecasting future consumption behaviors. More specifically, the time of
results. an electricity consumption behavior (e.g., electricity usage of
Load forecasting for individual users gives the estimation washing machine) happened in the future are probably close
of the total electricity consumption of an individual user, to the time that the same behavior happened in the past. These
e.g., a resident. However, this problem is still challenging. correlations exist inside the load profile, and is significant for
A major reason is that electricity consumption behaviors of individual users’ load forecasting.
single user is stochastic. The stochasticity is introduced by the Consequently, some researchers employed the load data
uncertainty of the time that electricity consumption activities of different appliances for load forecasting. These researches
occurred [21]. Another reason for the difficult of individ- employed different methods to explore the correlation among
ual users’ load forecasting is the consumption of electricity different appliances. Dinesh et al. [41] used nonintrusive
usage is dynamic, even for a specific application of the same appliance load monitoring techniques [26] to collect the
user. appliances’ load and forecast household load based on graph
In order to provide supporting accurate information for res- spectral clustering. Mohi Ud Din et al. [42] applied the
idential demand response, STLF for individual users attracts appliances-level load as the input features of neural network
increasing interests recently. Some methods [22]–[24] apply structure to forecast short-term load, and PCA technique is
clustering techniques to obtain the groups of users that have employed for feature reduction.
similar consumption behaviors. Teeraratkul et al. [22] pro- In this paper, we explore using the spatio-temporal cor-
posed a shape based clustering method for STLF. Based relation among different kinds of electricity consumption
on the profile of load curves, the method uses dynamic behaviors to improve the performance of STLF. We present
time warping technique to cluster the load curve and find a framework of STLF for individual users based on the cor-
a canonical shape for each set of curves. Then, a Markov relation information. Multiple time series are conducted in
model is used to conduct the individual users’ load fore- the framework to describe electricity consumption behaviors
casting. Quilumba et al. [23] grouped individual users based for different applications and their internal spatio-temporal
on similar consumption behaviors, which are represented by relationship. And a method based on Deep Neural Network
users’ load data. Based on the clustering results, a neural and iterative ResBlocks is proposed to learn the correlation
network employing weather and calendar features is used in among consumption behaviors for STLF. ResBlock that con-
the prediction phase. sists of few stacked layers and one skip connection is based

55786 VOLUME 8, 2020

Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

TABLE 1. Nomenclature. overview of the proposed STLF framework. As shown in

Figure 1, the proposed STLF framework consists of four
modules, including data acquisition, data preprocessing,
model training and load forecasting. The data acquisition
module collects measurements from household smart meters,
which report electricity consumption data of appliances for
each consumer. The spatio-temporal correlation is included
in the appliances’ consumption data, which is the output of
the data acquisition module. Data cleaning, data integration
and data transformation are conducted in the data prepro-
cessing module to improve the data quality for the input of
STLF model. In the model training module, a deep learning
model based on deep neural network and iterative ResBlock
is designed to learn the spatio-temporal correlation among
different electricity consumption behaviors. Also, a param-
eter optimization step is included in this module to further
enhance the learning ability of the proposed method. After
the preprocessing steps and the model training procedure,
the proposed model is able to calculate the predicted values
for an individual user. These components work together to
form a suitable solution for residential STLF.

A. DATA ACQUISITION MODULE

As shown in the Figure 2, we would like to employ appli-
ances’ load for residential STLF. This module’s output is
the appliances’ load data that contains the spatio-temporal
correlation among different kinds of electricity consump-
on the building block for ResNet [30]. The proposed method tion behaviors. There are two ways to obtain the appli-
based on iterative ResBlocks is able to learn both shallow and ances’ load for residential STLF. The first one is to install
deep features of the input vectors, which makes use of the load monitoring infrastructures for each appliance, which
correlation information among different kinds of electricity can report the electricity usage of the appliances. This way
consumption behaviors. The contributions of this paper are requires extra equipment, which will increase the cost of
summarized as follows: power system. The other way is to use nonintrusive appli-
1. We designed a framework of STLF for individual users, ance load monitoring techniques, which applies disaggrega-
which improves the performance of STLF by using the spatio- tion algorithms to decompose the entire household load into
temporal correlation among appliances’ load data. appliances’ load data [26] or separates certain appliance’s
2. We proposed a method based on DNN with iterative load from the entire household power consumption [25].
ResBlocks to learn the spatio-temporal correlation among The algorithms [26], [27] describe the characteristics of the
different kinds of electricity consumption behaviors in the appliances based on high frequency load data (1 Hz or even
framework. To the best of our knowledge, this is the first work higher). A typical disaggregation algorithm is sparse cod-
using iterative ResBlocks to learn the latent features of elec- ing [28], [29], which has high temporal resolution ability to
tricity consumption behaviors for the short-term residential separate different applications. At the end of this module,
load forecasting problem. appliances’ load that contains the spatio-temporal correlation
3. Experiments using real world data were conducted to are obtained.
evaluate the performance of the proposed approach.
B. DATA PREPROCESSING MODULE
Table 1 illustrates all symbols and abbreviations used in
this paper. The rest of this paper is organized as follows. The Usually, the load data acquired from the data acquisition
designed framework and the proposed Iterative Resblocks module are not suitable for the input of forecasting model
Based Deep Neural Network (IRBDNN) model are presented directly. This is because the original data may have missing
in Section 2 and 3. In Section 4, we evaluate the performance values, various formats and high computational needs in the
of the proposed forecasting approach. And in Section 5 is the real world. The residential load data need to be preprocessed
conclusion of this paper. and transformed into a suitable form for the STLF model.
The first step to preprocess the residential load is data
II. PRORPOSED FRAMEWORK cleaning. In the real world, the problem of missing values
In this section, we present the proposed STLF framework is usually the most popular problem, which may be caused
and the IRBDNN based method. Figure 1 is the schematic by hardware failures. We consider two ways to handle this

VOLUME 8, 2020 55787

 Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

FIGURE 1. Schematic overview of the short-term load forecasting framework.

FIGURE 2. Disaggregation of household electricity usage. FIGURE 3. An example of daily energy consumption for a house in the
greater Boston area of the U.S.

problem in the residential load. 1) If the length of the missing

values’ duration is acceptable, the missing values can be not suitable for STLF due to the following reasons. Firstly,
estimated according to the values before and after the missing load forecasting at the second level is impossible due to the
durations and the values in other days in the corresponding dynamic and stochastic characteristic of electricity consump-
time. 2) If the length of the missing values’ duration is too tion behaviors. Secondly, there is too much noise in the load
long that is not acceptable, the load data of this day will be records at a second level, which may increase the training
ignored in the model training phase. difficulties of forecasting models. Therefore, it is necessary to
The next step is data integration. When appliances’ conduct the data transformation step, which aims to obtain a
load monitoring infrastructures report electricity usage, simplified representation of the original dataset. The record-
the recording frequency of each monitor may be different. ing frequency of the processed data set after the data trans-
Therefore, data integration needs to be conducted to provide formation step will be lower than the original dataset. For
a uniform format for the input of forecasting model. One of example, the recording frequency will be transformed into
the data integration approaches is calculating the sum for load one measurement each hour from one measurement every
data in a duration to form a dataset that has lower recording three seconds. An example of daily energy consumption for
frequency than the original dataset. For instance, in order to a house in the greater Boston area of the U.S. is illustrated
get time series which records 1 value each minute, load data in Figure 3.
of appliances need to be integrated by calculating the sum of
load data in every minute. Also, normalization needs to be C. MODEL TRAINING MODULE AND
conducted in each appliance’s load data, which aims to equal FORECASTING MODULE
the influence of different kinds of appliances’ load data on the The forecasting model should learn the spatio-temporal cor-
forecasting results. relation among different kinds of electricity consumption
The recording frequency of load monitoring infrastruc- behaviors. The model is expected to have the ability that cap-
tures are usually at a second level, such as one measurement tures the characteristics among appliances’ load in both space
every three seconds. However, the second level load data is and time. In the model training module, a specific forecasting

55788 VOLUME 8, 2020

Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

model that can learn the spatio-temporal correlation needs to

be built. Then the model needs to be optimized and updated
to obtain an effective solution. After these steps, the training
of the forecasting model is completed.
The last module of the framework is the load forecasting
module. Based on the model output in the model training
module and the historical records, a predicted load value will
be calculated this step.
FIGURE 4. Comparison of stacked layers and ResBlock. (a) Structure of 2
III. PROPOSED MODEL stacked layers; (b) Structure of a ResBlock.
A. THE IRBDNN MODEL
In this subsection, we introduce the proposed the IRBDNN
model, which implements the forecasting model building step
in the model training module of Figure 1. The proposed model
employs iterative ResBlocks in a DNN and the model is able
to learn the spatio-temporal correlation among different kinds
of user’s electricity consumption behaviors. The input of the
IRBDNN model includes both the spatio correlation and the
temporal correlation of a residential user, in terms of resi-
dential user’s electricity consumption behaviors of different
appliances and historical electricity consumption behaviors,
respectively. The structure of the IRBDNN model enables the
model to learn both the deep features and the shallow features
from the input. Moreover, the structure of iterative ResBlocks
FIGURE 5. An illustration of the IRBDNN structure.
enables the learning ability of the deep IRBDNN model is no
worse than the learning ability of a shallow IRBDNN model.
Detailed analysis for the learning ability of the proposed
model is discussed in Section IV. C, which shows that the
employment of iterative ResBlocks can help to improve the
forecasting performance.
The proposed model is employed to learn the nonlinear
relationship between the input features and the output value.
Generally, the increase of model depth strengthens the learn-
ing ability of the neural network. However, the performance
of deep learning model may degrade in practice. The possible
reasons for the degraded performance may be the intrinsic
characteristic of data or the optimization difficulties of deep FIGURE 6. The iterative ResBlocks in the IRBDNN.
learning model. He et al. [30] proposed a method using Res-
Blocks to obtain a better performance than compared meth-
ods only using stacked hidden layers. Inspired by their work, stacked layers is y = F(x). Figure 4(b) shows the structure
we propose a model based on DNN and iterative ResBlocks of a ResBlock, which consists of two stacked layers and one
for the STLF task. The iterative ResBlocks are able to learn skip connection. When the input and the output of the skip
the spatio-temporal correlation in the STLF task and ensure connections have the same dimension, the skip connection
the learning ability of the model. is a typical identical mapping. Therefore, the output of the
ResBlock is a structure that is different to stacked layers. corresponding ResBlock is y = F(x) + x. When the input
It is similar to the building block for ResNet [30] that is and output of the skip connection are in different dimensions,
widely used in the image classification problem, but with a the skip connection performs as a linear projection to match
twist. the changes of dimensions. The output of the corresponding
The input and output of the skip connection in a ResBlock ResBlock is y = F(x) + Wx, where W is a linear projection.
can be in different dimensions, while in the building block When the stacked layers and the ResBlocks contain the same
for ResNet they are usually the same. The basic structure number of hidden layers, the skip connection ensures that the
of ResBlock consists of two components, i.e., few stacked learning ability of the ResBlock is no worse than the learning
layers and one skip connection. Stacked layers consist of ability of the stacked layers.
few hidden layers and adjacent layers are directly connected. The IRBDNN model is constructed by the stacked lay-
A structure of two stacked layers are shown in Figure 4(a). ers and the iterative ResBlocks. The structure is shown in
The input of the stacked layers is x and the output of the Figure 5, and Figure 6 illustrates the insight for the structure

VOLUME 8, 2020 55789

 Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

of the iterative ResBlocks. We denote the number of iterations C. SEQUENTIAL GRID SEARCH METHOD FOR
as t. When t = 0, the ResBlocks module degenerates into a HYPER-PARAMETER OPTIMIZATION
non-iterative structure, and the IRBDNN model degenerates This subsection introduces the hyper-parameters optimiza-
into a DNN model. When t = 1, the input of the first ResBlock tion method for the proposed the IRBDNN model. As it
(denoted as ResBlock 1) is added to the output of the ResBlock is described in Section 3.1, the IRBDNN architecture con-
1 by a skip connection. The ResBlock 1 consists of three parts: sists of the stacked layers and the iterative ResBlocks. The
m stacked layers, the ResBlock 2 and a skip connection. When number of hidden neurons in each layer of IRBDNN model
t = t0 , the iterations are repeated t0 times. The ResBlock are the same, which is denoted as N . The weight matrix
t0 consists of m stacked layers, ResBlock (t0 + 1) and a that connects the hidden neurons in (l − 1)th layer and
skip connection. If iteration t0 is the last iteration, ResBlock the neurons in l th layer is W l , which is a N × N matrix.
(t0 + 1) degenerate into n stacked layers ( m > 0, n > 0). bl is a vector that contains N elements which represents
For each ResBlock in the IRBDNN model, the input of the bias of the hidden neurons in lth layer. The output of
ResBlock is linked to the output directly by skip connection, (l − 1)th layer is denoted as al−1 . Then the output of layer
which ensures the learning capability of the current ResBlock lth is al = σ (W l al−1 + bl ), where σ denotes the activation
with deeper embedded ResBlock is no worse than the learning function.
capability of the ResBlock without deeper embedded Res- The loss function applied in the model is given by (5):
Block. The structure enables the model to make full use of the r
spatio-temporal correlation among the different consumption 1 X
behaviors. Loss = (t(n) − p(n))2 (5)
N
As it is described above, when the number of iterations
is 0, the structure of the IRBDNN model degenerates into a where t(n) and p(n) denote the truth load data and the
structure of DNN. When the number of iterations is no less predicted load data in the n-th time interval, and N is
than 1, the iteration procedures of the IRBDNN model are the number of predicted time intervals in the training
presented as below: set.
In deep learning methods, numbers of hyper-parameters
y = F(x1 , 21 ) + W0 (x0 ) (1) need to be optimized. An exhaustive grid search for all
F(x1 , 21 ) = F(x2 , 22 ) + W1 (x1 ) (2) hyper-parameters is time-consuming. To address this prob-
lem, we design a sequential grid search approach to opti-
...
mize hyper-parameters for the proposed the IRBDNN model,
F(xt , 2t ) = F(xt+1 , 2t+1 ) + Wt (xt ) (3) which is inspired by Ismail et al. work [31]. In this
paper, the following hyper-parameters are optimized. The
where F(xt+1 , 2t+1 ) is the output of (m + n) staked layers firsthyper-parameter is the number of neuron N in each layer,
in the ResBlock (t + 1) with the input xt , (t + 1) is the searched in (100, 150, 200, 300, 400, 450, 500). The second
number of iterations, y is the output of the IRBDNN model, hyper-parameter is the learning rate LR, searched in (0.001,
2 denotes weights, biases associated with the model and Wt 0.0001, 0.00001). And the last one is the initializer I that can
denotes the linear projection to match the possible changes of be used for the IRBDNN parameters, searched in (Normal,
dimensions. Uniform, Glorot Normal, Glorot Uniform). The sequential
grid search process is illustrated in Figure 7 and Figure 8.
B. INPUT OF THE IRBDNN MODEL The Algorithm 1 described in Figure 7 is the main program
In this subsection, we construct multiple time series based and the Algorithm 2 described in Figure 8 is its subprogram.
on appliances’ load data to form the input of the IRBDNN The sequential grid search can be divided into three parts.
model. The time series consist of a number of load values In the first part, the hyper-parameters are initialized to build
and each load value represents the energy consumption for a the initial forecasting model. N and LR have significant
duration. We use ‘‘time interval’’ to denote a duration in the impact on the learning ability of the forecasting model. There-
rest of this paper. The preprocessed load data for the appliance fore, we optimize N and LR synchronously to define Model
i in time interval t is represented as Ei (t), and the overall 1 in the second part. I is set an initialized value first, then it is
load in time interval t is denoted as E0 (t). The sequence for optimized in the third part. After the above process, Model 2
forecasting the overall load E0 (t0 + 1) is shown as below: is defined with the optimal N , LR and I .

X (t) = {E1 (t0 − T + 1), E1 (t0 − T + 2), . . . , E1 (t0 ), . . . , IV. RESULTS

Ei (t0 − T + 1), Ei (t0 − T + 2), . . . , Ei (t0 ), . . . , To evaluate the performance of the proposed framework and
E0 (t0 − T + 1), E0 (t0 − T + 2), . . . , E0 (t0 )} (4) the IRBDNN based method, we compare the forecasting
performance of the proposed method with existing algorithms
where X (t) is the input vector consisting of two parts, includ- including the ARMA based method, the ELM based method
ing the appliances’ load data and the overall load; the length and the SRX [42] based method. The above experiments are
of the historical records for forecasting is denoted as T . conducted based on the Redd dataset [32].

55790 VOLUME 8, 2020

Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

FIGURE 8. The flow diagram for the subprogram of the sequential grid
search approach.

appliances’ load data in each 30 minutes to get a time series

that consist of 48 values each day, then apply the min-max
normalization for the time series.
The Redd dataset that we use contains ten appliances
(oven, refrigerator, dishwasher, kitchen outlet, washer dryer,
bedroom, lighting, electric heat, microwave and stove).
We obtained 32 days’ load data for one user and separate them
into three subsets: 20 days’ data, 4 days’ data and 7 days’
data for the training set, the validation set and the testing set,
respectively. Table 2 shows the descriptive statistics of the
overall load data in the three subsets, including total size,
mean value, maximum value, minimum value and standard
deviation.
An example of the daily load for all appliances is illus-
trated in Figure 9. Bars with different colors illustrate the
FIGURE 7. The flow diagram for the main program of the sequential grid appliances’ load and a red line presents the overall load.
search approach.
As shown in the figure, the time series of appliances’ load are
quite dynamic. For example, the load record of oven is higher
A. DATASETS than 2500kWh around 7:30, while the load data is lower than
In this section, we introduce the Redd dataset and the pre- 1500kWh around 20:00. Moreover, the spatial correlation
processing steps for it. The Redd dataset is a public available among appliances is obvious in the figure. For instance,
data set, which records the consumption data of appliances the oven and the microwave are used together around 7:30 and
in residential users’ houses from March, 2011 to July, 2011. around 20:00. Also, lighting’ load record value will increase
In our experiments, we employ the data that the recording when the oven and microwave are used.
frequency is every three seconds for each record.
Data preprocessing for the Redd dataset includes data B. EXPERIMENT RESULTS
cleaning, data integration and data transformation. The data In the experiments, we use Root Mean Square Error (RMSE),
cleaning step is conducted according to the description in the Mean Absolute Error (MAE) and Mean Absolute Percentage
Section 2.2. After the data cleaning steps, we sum up the Error (MAPE) as the metrics of performance evaluation.

VOLUME 8, 2020 55791

 Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

TABLE 2. Descriptive statistics of the three subsets. TABLE 3. Range of hyper-parameters for grid search.

TABLE 4. The comparison for results of forecasting methods.

FIGURE 9. An example of energy consumption over a day for one user in

the Redd dataset.

RMSE is root mean square error, which is sensitive to

errors and suitable to evaluate the accuracy of forecasting
results [33]. MAE is mean absolute error, which is a useful
metric widely used in model evaluation [34]. MAPE is mean
absolute percentage error between the predicted value and
the truth value, which is able to avoid the offset problem of
errors [35]. The definitions of the metrics are as follows:
r
1 X
RMSE = (y − p)2 (6)
N
1X
MAE = |(y − p)| (7) method and the ARMA based method, which only employ
m
1 X |(y − p)| the overall load. For the ELM based method, the search range
MAPE = ∗ 100% (8) of the hidden neurons is the same with the IRBDNN based
m y
method. And for the DNN based method, the number of
where m is the number of values that represent the load hidden layers is searched in range (3, 4, 6, 8), and the search
data for a duration of 30 minutes in the testing set, y is the ranges of other hyper-parameters are the same as that of the
truth value and p is the predicted value. We implemented IRBDNN based method described in Table 2.
an 8 layers IRBDNN with 3 iterations. The number of the Table 4 shows the forecasting results for each day in the
stacked layers in each ResBlock is 1 layer, 1 layer, and testing set and the forecasting results for the whole testing set.
2 layers, respectively. The sequential grid search algorithm It can be observed that the IRBDNN based method performs
is applied in the hyper-parameter tuning step, and the search the best and the second best on Day 2, Day 3, Day 5 Day 6
ranges of hyper-parameters are presented in Table 3. and Day 7. The last row illustrates the overall results of
In order to evaluate the effectiveness of our framework five methods. The IRBDNN based method outperforms other
and the proposed IRBDNN based method, we conducted two methods, and the DNN based method performs the second
groups of experiments. In the first group, the experiments best, better than the SRX based method, the ELM based
employ both the appliances’ load data and the overall load. method and the ARMA based method. We can observe from
The DNN based method, the SRX based method, and the pro- the table that the IRBDNN based method performs better
posed IRBDNN based method are conducted in this group. than other methods on the three metrics, which indicates
The experiments in another group includes the ELM based that the proposed method using deep neural network and

55792 VOLUME 8, 2020

Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

FIGURE 10. Error distributions of the results of the forecasting methods. (The suffix ‘AO’ indicates that the model employs both appliances’
load and the overall load data, and the suffix ‘O’ indicates that the model only employs the overall load data).

for the other methods that is larger than 210 is larger than 25.
The IRBDNN based method’s forecasting results have less
extreme large RSEs (>210) than other methods and has
smaller RSEs (0-140) than other methods. Considering the
overall performance in Table 4, we can draw the conclusion
that the IRBDNN based method generally performs better
than other methods.

C. DISCUSSION
To obtain a more comprehensive understanding for the per-
formance of the proposed approach, four additional sets of
experiments are conducted. The training data, validation data
FIGURE 11. Forecasting Performance of the proposed method for a week. and the testing data in these experiments are the same as the
data used in the Section IV, B.
iterative ResBlock can effectively learn the spatio-temporal
correlation among consumption behaviors. The forecasting 1) PERFORMANCE ANALYSIS OF SPATIO-TEMPORAL
results of a residential user using the proposed method as well CORRELATION AND ITERATIVE ResBlocks
as the actual load are illustrated in Figure 11. In the first set, three groups of experiments are conducted to
Figure 10 is the error distributions of the results of the fore- verify the impact of spatio-temporal correlation among dif-
casting methods. We can observe from the figure that the error ferent appliances and iterative ResBlocks. The three groups
distributions of the IRBDNN-OA, DNN-OA and SRX-OA are the IRBDNN group, the DNN group and the SRX group.
perform better than ELM-O and ARMA-O that only employ Each group includes two cases with different historical data:
the overall load data. More specifically, the number of RSEs 1) the overall load data; 2) the overall load data and the
between 0 and 140 (the first two bars) in the IRBDNN-OA’s appliances’ load data. The experiments results are shown
forecasting result is over 255, which is more than DNN-OA, in Figure 12.
SRX-OA, ARMA-O and ELM-O. In other words, the number The results from Figure 12 show that both the appliances’
of RSEs larger than 140 in the IRBDNN-OA’s forecasting load data and iterative ResBlocks can help to improve the
results is the fewest among the forecasting result of the five forecasting performance. Detailed analysis is presented as
methods, because the RSE is calculated for each predicted follows. Firstly, comparing two experiments in each group,
value and the number of RSEs for each method is equal. The the case employs both appliances’ load and the overall load
number of RSEs larger than 210 in the forecasting result of performs better than the case only employs the overall load
the IRBDNN-OA is around 25, while the number of RSEs data. We can conclude that the employment of the appliances’

VOLUME 8, 2020 55793

 Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

FIGURE 13. RMSE of the IRBDNN model with different number of

iterations.

FIGURE 12. Impact of employing the appliances’ load data. (a) RMSE;
(b) MAE.

TABLE 5. Details for the network structure of the IRBDNN model with
different number of iterations. FIGURE 14. MAE of the IRBDNN model with different number of
iterations.

optimal the hyper-parameters, and the search ranges of the

hyper-parameters are illustrated in Table 3. Figure 13 and
Figure 14 illustrate the RMSE and MAE of the experiment
results.
From Figure 13, we can observe that the IRBDNN model
with 3 iterations performs the best among all models. The
RMSE decreases when the number of iterations (i.e., the num-
load data can help to improve the forecasting performance. ber of ResBlocks) increases. The same trend can be observed
Secondly, the case ‘IRBDNN-AO’ performs better than the from the MAE in Figure 14. There is no experiment of the
case ‘DNN-AO’ and the case ‘SRX-AO’. The same trend IRBDNN model with 4 iterations because the performance
can be observed from the cases that only employ the overall of the IRBDNN model with 3 iterations is close to the per-
load. Case ‘IRBDNN-O’ performs better than other case formance of the IRBDNN model with 2 iterations, which
‘DNN-O’ and case ‘SRX-O’. These results indicate that the only improves 1.14% on RMSE and 0.81% on MAE. When
employment of iterative ResBlocks can better capture the the number of iterations continuously increases, the improve-
spatio-temporal correlation among different appliances than ment on the RMSE and MAE will not be obvious. The results
compared methods and help to improve the forecasting per- indicate that the IRBDNN based method intends to have
formance. However, the improvement of forecasting perfor- improved performances with the increment of the iteration
mance is at the cost of computation efficiency. We further number.
analysis the cost of computation time in the fourth part of
this subsection. 3) IMPACT OF HIDDEN NEURON NUMBERS ON THE
IRBDNN BASED METHOD
2) IMPACT OF ITERATION TIMES ON THE In the third group of experiments, we explore the influence of
IRBDNN BASED METHOD the hidden neuron numbers for the 8-layer IRBDNN model
In the second set of experiments, we test the IRBDNN with 3 iterations, and the results are illustrated in Figure 15
models with 0 iteration, 1 iteration, 2 iterations and 3 iter- and Figure 16.
ations, respectively. Table 5 shows the details for the net- According to the results in Figure 15 and Figure 16, we can
work structure of the IRBDNN model with different number observe that optimal hidden neuron number is 300. The
of iterations. The sequential grid search method is used to changes on RMSE and MAE for different number of hidden

55794 VOLUME 8, 2020

Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

TABLE 6. Running time for the IRBDNN models.

FIGURE 15. Influence on MAE of hidden neuron number.

computation cost on the proposed method, the overall training
time as well as the overall testing time are acceptable for the
power company since they are relatively short and the training
process is usually off-line.

V. CONCLUSION
In this paper, we explored using the spatio-temporal correla-
tion among different kinds of appliances to predict the short-
term electricity demand for individual residential users.
An effective STLF framework that includes the data acqui-
sition module, the data preprocessing module, the model
FIGURE 16. Influence on RMSE of hidden neuron number.
training module and the load forecasting module was pro-
posed. Multiple time series were conducted in the frame-
work to describe electricity consumption behaviors and their
neurons are similar. Both of them decrease when the hidden internal spatio-temporal relationship. In order to fully exploit
neuron number increases from 100 to 300, while both of the correlation of user behaviors and characteristics of users’
them increase when the number of hidden neurons is larger consumption patterns, a method based on DNN and iterative
than 300. ResBlocks was proposed to learn the correlation. A grid
search method was employed in the hyper-parameter opti-
4) ANALYSIS OF COMPUTATION TIME ON THE IRBDNN
mization phase. The proposed method and several existing
BASED METHOD
forecasting methods were evaluated on a real world dataset.
The results show that the IRBDNN based method outper-
In order to analysis the computational cost, we compared
forms other compared methods. Moreover, we demonstrated
the training time and the testing time of the IRBDNN based
that both the appliances’ load data and iterative ResBlocks
method and the based DNN method, since both of them
can help to improve the forecasting performance. In addition,
are implemented on neural network. The experiments are
experiment results indicate that the IRBDNN based method
conducted on a personal computer equipped with a 2.5GHz
intends to have improved performances with the increment of
Intel i5 Core Processer and 8GB RAM. We compared the
the iteration number. In the future work, we intend to employ
training time and the testing time for both methods with two
the correlation defined in communication networks [38], [39]
types of historical data, i.e., 1) only the overall load (denoted
to express the spatio-temporal correlation among differ-
as input type ‘O’); 2) both the overall load and the appliances’
ent residential users to improve the performance of STLF.
load (denoted as input type ‘OA’).
Also, we will explore predicting thermophysical properties of
Table 6 shows the computational cost for both methods.
matter [44].
The training time is recorded by running 200 epochs for the
whole training set. And the testing time is recorded by the
REFERENCES
whole testing set, which has 7 day’s data. We can observe
[1] H. J. Monfared, A. Ghasemi, A. Loni, and M. Marzband, ‘‘A hybrid price-
from the table that the employment of the appliance’s load based demand response program for the residential micro-grid,’’ Energy,
increases the computation cost for the training process and vol. 185, pp. 274–285, Oct. 2019.
the testing process for both methods, which indicates that the [2] J. A. Gomez-Herrera and M. F. Anjos, ‘‘Optimal collaborative demand-
response planner for smart residential buildings,’’ Energy, vol. 161,
improvement of the forecasting performance by employing
pp. 370–380, Oct. 2018.
the appliances’ load is at the cost of computation time. Also, [3] M. H. Albadi and E. F. El-Saadany, ‘‘A summary of demand response
we can observe from the table that the employment of iterative in electricity markets,’’ Electric Power Syst. Res., vol. 78, no. 11,
ResBlocks does not evidently increase the computation cost pp. 1989–1996, Nov. 2008.
[4] W. Yuan, J. Huang, and Y. J. A. Zhang, ‘‘Competitive charging station
of the IRBDNN based method that adopting skip connections pricing for plug-in electric vehicles,’’ IEEE Trans. Smart Grid, vol. 8, no. 2,
in the network structure. Thus, although there is additional pp. 627–639, Dec. 2017.

VOLUME 8, 2020 55795

 Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

[5] M. Muratori and G. Rizzoni, ‘‘Residential demand response: Dynamic [28] C.-N. Yu, P. Mirowski, and T. K. Ho, ‘‘A sparse coding approach to
energy management and time-varying electricity pricing,’’ IEEE Trans. household electricity demand forecasting in smart grids,’’ IEEE Trans.
Power Syst., vol. 31, no. 2, pp. 1108–1117, Mar. 2016. Smart Grid, vol. 8, no. 2, pp. 738–748, Jan. 2016.
[6] T. Chen, ‘‘A collaborative fuzzy-neural approach for long-term load [29] Y. Wang, Q. Chen, C. Kang, Q. Xia, and M. Luo, ‘‘Sparse and redundant
forecasting in taiwan,’’ Comput. Ind. Eng., vol. 63, no. 3, pp. 663–670, representation-based smart meter data compression and pattern extrac-
Nov. 2012. tion,’’ IEEE Trans. Power Syst., vol. 32, no. 3, pp. 2142–2151, May 2016.
[7] L. Han, Y. Peng, Y. Li, B. Yong, Q. Zhou, and L. Shu, ‘‘Enhanced deep [30] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image
networks for short-term and medium-term load forecasting,’’ IEEE Access, recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR),
vol. 7, pp. 4045–4055, 2019. Jun. 2016, pp. 770–778.
[8] Z. Yu, Z. Niu, and W. Tang, ‘‘Deep learning for daily peak load [31] M. Ismail, M. Shahin, M. F. Shaaban, E. Serpedin, and K. Qaraqe, ‘‘Effi-
forecasting—A novel gated recurrent neural network combining dynamic cient detection of electricity theft cyber attacks in AMI networks,’’ in Proc.
time warping,’’ IEEE Access, vol. 7, pp. 17184–17194, 2019. IEEE Wireless Commun. Netw. Conf. (WCNC), Apr. 2018, pp. 1–6.
[32] J. Z. Kolter and M. J. Johnson, ‘‘REDD: A public data set for energy
[9] M. R. Haq and Z. Ni, ‘‘A new hybrid model for short-term electricity load
disaggregation research,’’ in Proc. Sustkdd, vol. 25, 2011, pp. 59–62.
forecasting,’’ IEEE Access, vol. 7, pp. 125413–125423, 2019.
[33] X. Gao, X. Li, B. Zhao, W. Ji, X. Jing, and Y. He, ‘‘Short-term electric-
[10] W. Kong, Z. Y. Dong, Y. Jia, D. J. Hill, Y. Xu, and Y. Zhang, ‘‘Short-term
ity load forecasting model based on EMD-GRU with feature selection,’’
residential load forecasting based on LSTM recurrent neural network,’’
Energies, vol. 12, no. 6, p. 1140, 2019.
IEEE Trans. Smart Grid, vol. 10, no. 1, pp. 841–851, Jan. 2019.
[34] T. Chai and R. R. Draxler, ‘‘Root mean square error (RMSE) or mean
[11] L. Wei and Z. Zhen-gang, ‘‘Based on time sequence of ARIMA model absolute error (MAE),’’ Geoscientific Model Develop. Discuss., vol. 7,
in the application of short-term electricity load forecasting,’’ in Proc. Int. no. 1, pp. 1525–1534, 2014.
Conf. Res. Challenges Comput. Sci., Dec. 2009, pp. 11–14. [35] S. Kim, G. Lee, G.-Y. Kwon, D.-I. Kim, and Y.-J. Shin, ‘‘Deep learning
[12] S. S. Pappas, L. Ekonomou, D. C. Karamousantas, G. E. Chatzarakis, based on multi-decomposition for short-term load forecasting,’’ Energies,
S. K. Katsikas, and P. Liatsis, ‘‘Electricity demand loads modeling using vol. 11, no. 12, p. 3433, 2018.
AutoRegressive moving average (ARMA) models,’’ Energy, vol. 33, no. 9, [36] G. Huang, S. Song, J. N. D. Gupta, and C. Wu, ‘‘Semi-supervised and
pp. 1353–1360, Sep. 2008. unsupervised extreme learning machines,’’ IEEE Trans. Cybern., vol. 44,
[13] S.-J. Huang and K.-R. Shih, ‘‘Short-term load forecasting via ARMA no. 12, pp. 2405–2417, Dec. 2014.
model identification including non-Gaussian process considerations,’’ [37] P. J. Brockwell, R. A. Davis, and M. V. Calder, Introduction to Time
IEEE Trans. Power Syst., vol. 18, no. 2, pp. 673–679, May 2003. Series and Forecasting, vol. 2. New York, NY, USA: Springer, 2002.
[14] Y. Yang, J. Che, C. Deng, and L. Li, ‘‘Sequential grid approach based [Online]. Available: [Link]
support vector regression for short-term electric load forecasting,’’ Appl. 2526-1#authorsandaffiliationsbook
Energy, vol. 238, pp. 1010–1021, Mar. 2019. [38] Z. Qin, Y. Wang, H. Cheng, Y. Zhou, Z. Sheng, and V. C. M. Leung,
[15] L. Clark, D. Lou, D. Michelle, G. T. Alegata, and G. C. Luna, ‘‘Day-ahead ‘‘Demographic information prediction: A portrait of smartphone appli-
load forecasting using support vector regression machines,’’ Int. J. Adv. cation users,’’ IEEE Trans. Emerg. Topics Comput., vol. 6, no. 3,
Comput. Sci. Appl., vol. 9, no. 3, pp. 22–27, 2018. pp. 432–444, Jul. 2018.
[16] L. Hernández, C. Baladrón, J. M. Aguiar, B. Carro, [39] H. Huang, H. Yin, G. Min, H. Jiang, J. Zhang, and Y. Wu, ‘‘Data-driven
A. Sánchez-Esguevillas, and J. Lloret, ‘‘Artificial neural networks information plane in software-defined networking,’’ IEEE Commun. Mag.,
for short-term load forecasting in microgrids environment,’’ Energy, vol. 55, no. 6, pp. 218–224, 2017.
vol. 75, pp. 252–264, Oct. 2014. [40] M. H. Amini, A. Kargarian, and O. Karabasoglu, ‘‘ARIMA-based decou-
[17] X. Cao, S. Dong, Z. Wu, and Y. Jing, ‘‘A data-driven hybrid optimization pled time series forecasting of electric vehicle charging demand for
model for short-term residential load forecasting,’’ in Proc. IEEE Int. Conf. stochastic power system operation,’’ Electr. Power Syst. Res., vol. 140,
Comput. Inf. Technol., Ubiquitous Comput. Commun., Dependable, Auto- pp. 378–390, Nov. 2016.
nomic Secure Comput., Pervasive Intell. Comput., Oct. 2015, pp. 283–287. [41] C. Dinesh, S. Makonin, and I. V. Bajic, ‘‘Residential power forecasting
[18] Y. Liang, D. Niu, and W.-C. Hong, ‘‘Short term load forecasting based on using load identification and graph spectral clustering,’’ IEEE Trans. Cir-
feature extraction and improved general regression neural network model,’’ cuits Syst. II, Exp. Briefs, vol. 66, no. 11, pp. 1900–1904, Nov. 2019.
Energy, vol. 166, pp. 653–663, Jan. 2019. [42] G. Mohi Ud Din, A. U. Mauthe, and A. K. Marnerides, ‘‘Appliance-level
short-term load forecasting using deep neural networks,’’ in Proc. Int. Conf.
[19] M. Tucci, E. Crisostomi, G. Giunta, and M. Raugi, ‘‘A multi-objective
Comput., Netw. Commun. (ICNC), Maui, HI, USA, Mar. 2018, pp. 53–57.
method for short-term load forecasting in European countries,’’ IEEE
[43] V. Sze, Y.-H. Chen, T.-J. Yang, and J. S. Emer, ‘‘Efficient processing of
Trans. Power Syst., vol. 31, no. 5, pp. 3537–3547, Sep. 2016.
deep neural networks: A tutorial and survey,’’ Proc. IEEE, vol. 105, no. 12,
[20] K. M. Powell, A. Sriprasad, W. J. Cole, and T. F. Edgar, ‘‘Heating, cooling, pp. 2295–2329, Dec. 2017.
and electrical load forecasting for a large-scale district energy system,’’ [44] Y. Zhou, Q. Li, and Q. Wang, ‘‘Energy storage analysis of UIO-66 and
Energy, vol. 74, pp. 877–885, Sep. 2014. water mixed nanofluids: An experimental and theoretical study,’’ Energies,
[21] H. Shi, M. Xu, and R. Li, ‘‘Deep learning for household load forecasting— vol. 12, no. 13, p. 2521, 2019.
A novel pooling deep RNN,’’ IEEE Trans. Smart Grid, vol. 9, no. 5, [45] L. Ren, L. Zhang, H. Wang, and L. Qi, ‘‘An ensemble model based on
pp. 5271–5280, Mar. 2017. machine learning methods for short-term power load forecasting,’’ in Proc.
[22] T. Teeraratkul, D. O’Neill, and S. Lall, ‘‘Shape-based approach to house- IOP Conf. Ser., Earth Environ. Sci., vol. 186, Oct. 2018, Art. no. 012042.
hold electric load curve clustering and prediction,’’ IEEE Trans. Smart
Grid, vol. 9, no. 5, pp. 5196–5206, Sep. 2017.
[23] F. L. Quilumba, W.-J. Lee, H. Huang, D. Y. Wang, and R. L. Szabados,
‘‘Using smart meter data to improve the accuracy of intraday load fore-
casting considering customer behavior similarities,’’ IEEE Trans. Smart
Grid, vol. 6, no. 2, pp. 911–918, Mar. 2015.
[24] J. D. Rhodes, W. J. Cole, C. R. Upshaw, T. F. Edgar, and M. E. Webber,
‘‘Clustering analysis of residential electricity demand profiles,’’ Appl.
Energy, vol. 135, pp. 461–471, Dec. 2014. YE HONG received the [Link]. degree in com-
[25] K. X. Perez, W. J. Cole, J. D. Rhodes, A. Ondeck, M. Webber, puter science from Sichuan University, China,
M. Baldea, and T. F. Edgar, ‘‘Nonintrusive disaggregation of residential air- in 2017, where she is currently pursuing the
conditioning loads from sub-hourly smart meter data,’’ Energy Buildings, master’s degree in computer science. Her research
vol. 81, pp. 316–325, Oct. 2014. interests include behavioral data analysis and deep
[26] G. W. Hart, ‘‘Nonintrusive appliance load monitoring,’’ Proc. IEEE, learning applications.
vol. 80, no. 12, pp. 1870–1891, 1992.
[27] J. Z. Kolter, S. Batra, and A. Y. Ng, ‘‘Energy disaggregation via discrim-
inative sparse coding,’’ in Proc. Adv. Neural Inf. Process. Syst., 2010,
pp. 1153–1161.

55796 VOLUME 8, 2020

Y. Hong et al.: Deep Learning Method for Short-Term Residential Load Forecasting in Smart Grid

YINGJIE ZHOU (Member, IEEE) received the WENZHENG XU (Member, IEEE) received
Ph.D. degree from the School of Communication the [Link]., M.E., and Ph.D. degrees in com-
and Information Engineering, University of Elec- puter science from Sun Yat-sen University,
tronic Science and Technology of China (UESTC), Guangzhou, China, in 2008, 2010, and 2015,
China, in 2013. He was a Visiting Scholar with the respectively. He was a Visitor with The Australian
Department of Electrical Engineering, Columbia National University and The Chinese University
University, New York. He is currently an Assistant of Hong Kong. He is currently an Associate
Professor with the College of Computer Science, Professor with Sichuan University. His research
Sichuan University (SCU), China. His current interests include wireless ad hoc and sensor net-
research interests include network management, works, mobile computing, approximation algo-
behavioral data analysis, and resource allocation. rithms, combinatorial optimization, online social networks, and graph theory.

XIUJUAN ZHENG received the [Link]. degree

in electronic information engineering from
QIBIN LI received the Ph.D. degree from the Col- Northwestern Polytechincal University, Xi’an,
lege of Energy and Power Engineering, Chongqing China, in 2004, the [Link]. degree in biomedical
University (CQU), Chongqing, China, in 2014. He engineering from Zhejiang University, Hangzhou,
was a Visiting Scholar with the Department of China, in 2006, and the Ph.D. degree in elec-
Earth and Environmental Engineering, Columbia tronic and information engineering from The
University, New York. He is currently an Associate Hong Kong Polytechnic University, in 2011. She
Professor with the College of Energy and Power was a Postdoctoral Fellow with the School of
Engineering, CQU. His current research interests Medicine, Shanghai Jiao Tong University, from
include molecular dynamics (MD) and experimen- 2011 to 2013. Since 2014, she has been an Associate Professor with the
tal study of thermodynamic properties of working College of Electrical Engineering, Sichuan University, Chengdu, China. Her
fluid, thermal circulation system and its thermal management, and the anal- research interests include quantitative methods for biomedical signal and
ysis of thermal and mechanical properties of new materials in thermal-force image analysis and their interdisciplinary applications.
coupling field.

VOLUME 8, 2020 55797