Академический Документы
Профессиональный Документы
Культура Документы
2, APRIL 2018
Abstract—The smart grid (SG) has emerged as an important many actuators are deployed to enforce power scheduling,
form of the Internet of Things. Despite the high promises of protection, and security decisions. Because of the IoT-based
renewable energy in the SG, it brings about great challenges technology, SG and IoT are naturally inseparable. Recently,
to the existing power grid due to its nature of intermittent and
uncontrollable generation. In order to fully harvest its potential, numerous IoT technologies has been developed to fulfill the
accurate forecasting of renewable power generation is indispens- SG’s potential, including energy distribution and manage-
able for effective power management. In this paper, we propose ment [3]–[6], load balancing [7], security and privacy [8], and
a least absolute shrinkage and selection operator (LASSO)-based the future smart building framework [9].
forecasting model and algorithm for solar power generation The SG is characterized with the two-way flow of
forecasting. We compare the proposed scheme with two represen-
tative schemes with three real world datasets. We find that the power and information, microgrid, and distributed renew-
LASSO-based algorithm achieves a considerably higher accu- able energy resources (DRERs) [10]. In the meantime,
racy comparing to the existing methods, using fewer training the rise of new energy (e.g., photovoltaic power such as
data, and being robust to anomaly data points in the training solar power) has brought new challenges to such uncon-
data, and its variable selection capability also offers a convenient ventional power networks. Although integrating the power
tradeoff between complexity and accuracy, which all make the
proposed LASSO-based approach a highly competitive solution charge from solar power generators could reinforce the macro
to forecasting of solar power generation. grid, a large and uncertain amount of power generated by
micro solar grids could lead to severe energy management
Index Terms—Generation forecasting, Internet of Things
(IoT), least absolute shrinkage and selection operator (LASSO), problems [11].
machine learning, renewable energy. In order to fully harvest the potential of DRERs, two key
techniques, load forecasting (i.e., to predict the amount of
power needed to achieve the demand and supply equilib-
I. I NTRODUCTION rium) and power generation forecasting (i.e., to predict how
much power will be generated at a future time), are indis-
NTERNET of Things (IoT) is defined as uniquely
I identifiable objects that are organized in an Internet like
structure. With technology developments and evolution of the
pensable. Load forecasting has been well studied in [12]–[14],
with different statistics and machine learning approaches, such
as nonparametric functional time series analysis, state space
power grid, the concept of smart grid (SG) has emerged, which models, and artificial neural networks. Similarly, generation
is the next generation power grid, as well as an important forecasting has been investigated with various models and
part of the IoT [2]. An SG is an electricity network that methods as well.
can intelligently integrate the interactions of all users con- Since solar power generation is linked directly to solar inten-
nected to generators, consumers, and those that assume both sity, the solar power forecasting problem naturally translates
roles, in order to efficiently deliver sustainable, economic, and to a weather forecasting problem. In [15] and [16], support
secure electricity supplies [2]. Such capabilities are enabled by vector machine (SVM) and nonlinear time series are used to
the computation, communication, and control mechanisms that predict solar intensity, respectively. Other prior works such
are incorporated with the power grid, where a large amount as [17]–[19] also provided various effective solutions to the
of interconnected wireless sensors (e.g., phasor measurement solar intensity prediction problem. Although the prior works
units) are deployed to obtain real-time state information, while have done a good job on achieving a low error rate, there is
Manuscript received September 30, 2017; revised January 8, 2018 and always room for improvement for more accurate forecasting.
February 7, 2018; accepted March 1, 2018. Date of publication March 5, 2018; In addition, a deeper analysis will be helpful to gain a good
date of current version April 10, 2018. This work was supported in part by understanding of the problem. For example, the SVM-based
the U.S. NSF under Grant CNS-1702957 and Grant DMS-1736470 and in
part by the Wireless Engineering Research and Education Center at Auburn method [15] achieves a low error rate, but the selection of
University. This work was presented in part at IEEE GLOBECOM 2017, kernel is usually based on experience. For neural network-
Singapore, Dec. 2017 [1]. (Corresponding author: Shiwen Mao.) based technologies [19], [20], the neural network structure
N. Tang, S. Mao, and R. M. Nelms are with the Department of Electrical and
Computer Engineering, Auburn University, Auburn, AL 36849 USA (e-mail: needs to be predesigned and a quite complicated structure is
nzt0007@tigermail.auburn.edu; smao@ieee.org; nelmsrm@auburn.edu). needed to achieve a good precision, which, however, leads
Y. Wang is with the Department of Electrical Engineering, Nanjing to a high computational cost. In rainy or cloudy days, the
University of Aeronautics and Astronautics, Nanjing 210016, China (e-mail:
yuwang15@nuaa.edu.cn). time-series-based method [21] are usually not effective to cap-
Digital Object Identifier 10.1109/JIOT.2018.2812155 ture the high variations in data. In [16], we present a local
2327-4662 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
TANG et al.: SOLAR POWER GENERATION FORECASTING WITH LASSO-BASED APPROACH 1091
linear model for nonlinear time series, which leads to an accu- provides simultaneous confidence interval to making further
rate approximation and an analysis on the relationship between interpretation for each covariate. However, the nonlinear nature
the renewable power generation process and the weather vari- of the target problem could undermine the interpretation of the
able processes. However, the importance of each variable is results.
yet to be better identified. Other than these statistical methods, neural network is also a
In this paper, we investigate the solar power generation widely used approach to address the problem. Zhang et al. [18]
forecasting problem, aiming to develop an effective method proposed several different neural network models as competi-
that not only achieve a high forecasting accuracy but also tors. In this paper, the authors noticed the fact that weak
helps to reveal the significance of weather variables. To this stationarity and lack of continuity result in volatilities of solar
end, we propose a least absolute shrinkage and selection data. Thus, the proposed models use historical data similar to
operator (LASSO)-based method for solar power generation the target day to perform prediction. The accuracy for cer-
forecasting based on historical weather data. Based on a sin- tain models are great but we should be aware of the case
gle index model (SIM) and LASSO, we develop an effective when there are no historical, similar days. Also, the heavy
algorithm that maximizes Kendall’s tau coefficient to esti- computation of neural networks is always a concern.
mate the prediction model coefficients. The goal of variable Not limited to algorithms, data preprocessing could become
selection is achieved by the nature of LASSO, which auto- a useful means to reduce error. Chen et al. [32] and
matically reduce the weights of less important variables and Shi et al. [33] proposed to classify the data by weather con-
increase the sparsity of the overall coefficient vector. With the ditions before using their similar-day-based neural network
proposed algorithm, we can either maximize the prediction algorithms. Specifically, Chen et al. [32] classified histori-
accuracy using all the weather data/variables, or achieve a cal data by irradiance, total cloud, and cloud cover, while
tradeoff between accuracy and complexity by using a lim- Shi et al. [33] took the weather feature, such as sunny or
ited number of variables. The proposed scheme is evaluated cloudy for data classification. As certain tricks surely improve
with the real dataset collected from a weather station [22], the performance of model, we should still notice that the avail-
and comparison to two representative benchmark schemes. ability and accuracy of these features are highly dependent
The proposed LASSO-based scheme outperforms both exist- on location of dedicated datasets, making the schemes less
ing schemes with considerable reduction in prediction flexible.
error.
The remainder of this paper is organized as follows.
III. F ORECASTING M ODEL AND P ROBLEM S TATEMENT
We introduce LASSO and the forecasting problem in
Section III. Then we present our LASSO-based algorithm in A. LASSO Preliminaries
Section IV. Meanwhile, we also introduce the time-series solu- In machine learning and statistics, LASSO has become a
tion proposed in [16], which is used as a benchmark. We popular method for regression analysis, ever since it was first
validate the performance of our solution and compare it with introduced by Tibshirani [24] in 1996. By applying LASSO
two benchmark schemes in Section V. We conclude this paper to practical problems, we benefit from two main functions
in Section VI. that LASSO has: 1) regularization and 2) variable selection.
Due to the nature of LASSO, while a stronger l1 penalty is
used, LASSO is encouraged to shrink its coefficients of less
II. R ELATED W ORK important variables to 0. In other words, it performs variable
Power generation prediction is an essential issue in SG selection by dropping the corresponding variables from the
network, especially for system integrated with new energy, model and achieves a sparse solution in this case. On the other
such as wind or solar generation. In order to make accurate hand, while a weaker l1 penalty is used, the algorithm tends
prediction, considerable work has been done. In addition to to retain most variables and predict with better regularization.
those discussed in the introduction section, we review several The level of l1 penalty can be chosen by automatic techniques
additional key related work here. like cross-validation or by manually using the regularization
In [15], methods such as past-predicts-future (PPF) model, path. In recent years, LASSO has been successfully applied to
linear least square regression, and SVM regression are adopted various SIMs [25]–[27] due to the above-mentioned capability.
to solve the solar generation prediction problem. According to We propose to use LASSO for solar power generation
the authors, linear regression and SVM outperform PPF due with high accuracy. In addition, since weather data gathered
to the change of weather pattern. Also, as a conclusion, SVM from the local weather station can vary in different types of
shows a higher accuracy while linear least square achieves an weather parameters to monitor, it is important to find out
interpretable model. Although the idea is novel, the accuracy which variables are more important on solar power generation,
of these models are yet to be improved. especially when lacking of sufficient weather information, or
Inspired by this paper, we propose the time-series- when computation complexity is a concern. As discussed, lin-
based algorithm TLLE [16], which separates the his- ear regression, neural networks, and SVM-based algorithms
torical data into small neighborhoods and estimates the have already been applied to the solar power generation fore-
coefficient of each neighborhood with a linear solution. casting problem. To the best of our knowledge, this is the
While the algorithm improves the accuracy comparing to first application of LASSO to the problem, to achieve high
multilinear regression (MLR) and SVM, the model also prediction precision as well as variable selection.
1092 IEEE INTERNET OF THINGS JOURNAL, VOL. 5, NO. 2, APRIL 2018
B. System Model where sign(·) is the signum function. For the continuous form
The solar power forecasting problem is a good match for of β, Kendall’s tau coefficient is defined as
T
the SIM, which has the advantage of avoiding the so-called 1 Xi2 β − XTi1 β
∗
“curse of dimensionality” in fitting multivariate nonparametric τn (β) = sign Yi2 − Yi1 tanh
n(n − 1) c
regression functions by focusing on an index [23]. Specifically, 1≤i1 =i2 ≤n
Assume there are n data units, {X1 , X2 , . . . , Xn }, and the τn∗ β̂ (k−1) + β̂jk ejk − τn∗ β̂ (k−1) < . (8)
corresponding response values are {Y1 , Y2 , . . . , Yn }, respec-
tively. For discontinuous β, Kendall’s tau coefficient is Otherwise, we set β̂ (k) as
expressed as β̂ (k−1) + β̂jk ejk
β̂(k) = (9)
1 β̂ (k−1) + β̂jk ejk
τn (β) = sign Yi2 − Yi1 · sign XTi2 β − XTi1 β 2
n(n − 1)
1≤i1 =i2 ≤n and repeat the above steps until the stop condition (8) is
(3) satisfied. Then we obtain the estimated coefficient vector β̂.
TANG et al.: SOLAR POWER GENERATION FORECASTING WITH LASSO-BASED APPROACH 1093
j=1 Ker 1
b (t − Z( j) ) × Y( j)
f̂ (t) =
n
(11)
j=1 Ker 1
b (t − Z ( j) )
Fig. 7. Solar power generation prediction using the SVM-based method with
the UMass dataset. Fig. 10. Solar power prediction using SVM with the Diddington dataset.
Fig. 8. Solar power generation prediction using the TLLE-based method Fig. 11. Solar power prediction using TLLE with the Diddington dataset.
with the UMass dataset.
from [31], we predict the solar intensity for the period from
the 365th to 394th day.
With the Diddington dataset, we find the SVM method have
great difficulty with the small set of training samples, which
forces us to increase the amount of training samples to 100.
Here, we use the first 100 days’ weather data to train the SVM
model and finally obtain an acceptable result as presented in
Fig. 10. Meanwhile, the TLLE method still works better than
SVM. We used the first 60 days’ data as the model generating
data to achieve the best performance, which is illustrated in
Fig. 11. The prediction results with the proposed LASSO-
based approach is presented in Fig. 12, which is obtained
with a much smaller 30 training data size than SVM and
TLLE. The overall comparison of accuracy is presented in
Table I. For this dataset, the proposed LASSO-based approach
Fig. 9. Solar power generation prediction using the proposed LASSO-based
method with the UMass dataset. achieves reductions of 76.6763% and 66.1474% in RMSE over
SVM and TLLE, respectively, and reductions of 85.6069% and
80.0973% in MAPE over SVM and TLLE, respectively. Note
comprehensive evaluation. Unlike the weather in the U.S., the that such considerable gains are achieved with a much smaller
areas in Britain inevitably have less sunlight due to the much training data size.
more rainy and cloudy days. This different data feature can The simulation results with the Harnhill dataset are
be a practical test to our proposed method. For both datasets presented in Figs. 13–15 for the three schemes. The prediction
TANG et al.: SOLAR POWER GENERATION FORECASTING WITH LASSO-BASED APPROACH 1097
Fig. 12. Solar power prediction using the proposed method with the Fig. 14. Solar power prediction with TLLE of Harnhill dataset.
Diddington dataset.
TABLE I
P REDICTION ACCURACY W ITH THE D IDDINGTON DATASET
Fig. 15. Solar power prediction with proposed method of Harnhill dataset.
TABLE II
P REDICTION ACCURACY W ITH THE H ARNHILL DATASET
TABLE III
C ORRELATION M ATRIX OF THE UM ASS DATASET
TABLE IV
O PTIMIZED β W ITH T HREE VARIABLES
and 2) simplifying the prediction model. Due to the struc- VI. C ONCLUSION
ture of our proposed algorithm, historical data will also be
used in the prediction stage. So when less parameters are In this paper, we proposed an LASSO-based algorithm
used in the prediction model, the computational cost will that accurately predict solar power generation with a small
be greatly reduced. In addition, a simplified model can pro- amount of historical data. After presenting the detailed algo-
vide us a clearer understanding of the relationship between rithm design, we compared the proposed scheme with two
solar power generation and the weather parameters. We can representative existing schemes using three datasets with
use the reduced model to estimate solar power generation different features. We found that the LASSO-based algo-
when the dataset is incomplete, or to reduce the computation rithm achieved considerably higher accuracy comparing to
time when necessary (i.e., to tradeoff between complexity and the existing methods, using fewer training data, and being
accuracy). robust to anomaly data points in the training data. In addi-
As an example, we use the UMass dataset to illustrate the tion, the variable selection capability offered a nice tradeoff
variable selection procedure. Table III shows the correlation between complexity and accuracy. These features all made it
matrix computed with the dataset. Although the problem can- a highly competitive solution to forecasting of solar power
not be simply defined as a linear one, we could still use generation.
the correlation matrix to obtain an intuitive observation. As
the matrix shows, temperature and humidity are more closely R EFERENCES
correlated to solar intensity, while precipitation is quite inde- [1] N. Tang, S. Mao, Y. Wang, and R. M. Nelms, “LASSO-based single
pendent with most other parameters. After adjusting the λ index model for solar power generation forecasting,” in Proc. IEEE
value to reduce the model to a 3-variable model, we have GLOBECOM, Singapore, Dec. 2017, pp. 1–6.
the optimized β values listed in Table IV. [2] Y. Wang, S. Mao, and R. M. Nelms, Online Algorithms for Optimal
Energy Distribution in Microgrids, 1st ed. New York, NY, USA:
From Table IV, we find that both dewpoint and windspeed Springer, 2015.
are seen as less important variables. The result for windspeed [3] Y. Wang, S. Mao, and R. M. Nelms, “Distributed online algorithm for
coincides with the correlation matrix but dewpoint and pre- optimal real-time energy distribution in the smart grid,” IEEE Internet
Things J., vol. 1, no. 1, pp. 70–80, Feb. 2014.
cipitation have a conflict. However, noticing from Table III [4] L. Yu, T. Jiang, and Y. Zou, “Distributed online energy management for
that dewpoint is tightly correlated with temperature, while data centers and electric vehicles in smart grid,” IEEE Internet Things
precipitation is quite independent to temperature, the result J., vol. 3, no. 6, pp. 1373–1384, Dec. 2016.
[5] Y. Wang, S. Mao, and R. M. Nelms, “Online algorithm for optimal real-
for β becomes reasonable. Fig. 16 provides the prediction time energy distribution in the smart grid,” IEEE Trans. Emerg. Topics
result with the 3-parameter model. The RMSE in this case Comput., vol. 1, no. 1, pp. 10–21, Jun. 2013.
TANG et al.: SOLAR POWER GENERATION FORECASTING WITH LASSO-BASED APPROACH 1099
[6] Y. Huang, S. Mao, and R. M. Nelms, “Adaptive electricity scheduling [30] O. Burdakov, A. Grimvall, and M. Hussian, “A generalised PAV
in microgrids,” IEEE Trans. Smart Grid, vol. 5, no. 1, pp. 270–281, algorithm for monotonic regression in several variables,” in Proc.
Jan. 2014. COMPSTAT, Prague, Czech Republic, 2004, pp. 761–767.
[7] M. Li, P. He, and L. Zhao, “Dynamic load balancing applying water- [31] The Detection of Archaeological Residues using Remote-sensing
filling approach in smart grid systems,” IEEE Internet Things J., vol. 4, Techniques (DART) Project. DART Weather Data. Accessed:
no. 1, pp. 247–257, Feb. 2017. Mar. 2018. [Online] Available: https://dartportal.leeds.ac.uk/dataset/
[8] Z. Guan et al., “Achieving efficient and secure data acquisition for cloud- dart_monitoring_weather_data/
supported Internet of Things in smart grid,” IEEE Internet Things J., [32] C. Chen, S. Duan, T. Cai, and B. Liu, “Online 24-h solar power forecast-
vol. 4, no. 6, pp. 1934–1944, Dec. 2017. ing based on weather type classification using artificial neural network,”
[9] J. Pan et al., “An Internet of Things framework for smart energy in Solar Energy, vol. 85, no. 11, pp. 2856–2870, Nov. 2011.
buildings: Designs, prototype, and experiments,” IEEE Internet Things [33] J. Shi, W.-J. Lee, Y. Liu, Y. Yang, and P. Wang, “Forecasting power out-
J., vol. 2, no. 6, pp. 527–537, Dec. 2015. put of photovoltaic systems based on weather classification and support
[10] Y. Wang, S. Mao, and R. M. Nelms, “On hierarchical power schedul- vector machines,” IEEE Trans. Ind. Appl., vol. 48, no. 3, pp. 1064–1069,
ing for the macrogrid and cooperative microgrids,” IEEE Trans. Ind. May/Jun. 2012.
Informat., vol. 11, no. 6, pp. 1574–1584, Dec. 2015.
[11] M. J. E. Alam, K. M. Muttaqi, and D. Sutanto, “A comprehensive Ningkai Tang received the B.E. degree in infor-
assessment tool for solar PV impacts on low voltage three phase distribu- mation science and engineering from Southeast
tion networks,” in Proc. IEEE ICDRET, Dhaka, Bangladesh, Jan. 2012, University, Nanjing, China, in 2010, and the M.S.
pp. 1–5. degree in electrical and computer engineering from
[12] M. Chaouch, “Clustering-based improvement of nonparametric func- Auburn University, Auburn, AL, USA, in 2014,
tional time series forecasting: Application to intra-day household-level where he is currently pursuing the Ph.D. degree
load curves,” IEEE Trans. Smart Grid, vol. 5, no. 1, pp. 411–419, at the Department of Electrical and Computer
Jan. 2014. Engineering.
[13] V. Dordonnat, S. J. Koopman, and M. Ooms, “Dynamic factors in He was a Technology Support Engineer with
periodic time-varying regressions with an application to hourly elec- the ZTE Corporation, Shenzhen, China. His current
tricity load modelling,” Comput. Stat. Data Anal., vol. 56, no. 11, research interests include machine learning, smart
pp. 3134–3152, Nov. 2012. grids, cognitive radio, full-duplex transmission, and control theory.
[14] H. S. Hippert, C. E. Pedreira, and R. C. Souza, “Neural networks for
short-term load forecasting: A review and evaluation,” IEEE Trans. Shiwen Mao (S’99–M’04–SM’09) received the
Power Syst., vol. 16, no. 1, pp. 44–55, Feb. 2001. Ph.D. degree in electrical and computer engineering
[15] N. Sharma, P. Sharma, D. Irwin, and P. Shenoy, “Predicting solar gen- from Polytechnic University, Brooklyn, NY, USA,
eration from weather forecasts using machine learning,” in Proc. IEEE in 2004.
SmartGridComm, Brussels, Belgium, Oct. 2011, pp. 528–533. He is the Samuel Ginn Distinguished Professor
[16] Y. Wang, G. Cao, S. Mao, and R. Nelms, “Analysis of solar generation and the Director of the Wireless Engineering
and weather data in smart grid with simultaneous inference of non- Research and Education Center, Auburn University,
linear time series,” in Proc. IEEE INFOCOM WKSHPS, Hong Kong, Auburn, AL, USA. His current research interests
Apr./May 2015, pp. 600–605. include wireless networks, multimedia communica-
[17] R. J. Bessa, A. Trindade, and V. Miranda, “Spatial-temporal solar power tions, and smart grids.
forecasting for smart grids,” IEEE Trans. Ind. Informat., vol. 11, no. 1, Dr. Mao was a recipient of the 2017 IEEE
pp. 232–241, Feb. 2015. ComSoc ITC Outstanding Service Award, the 2015 IEEE ComSoc TC-CSR
[18] Y. Zhang, M. Beaudin, R. Taheri, H. Zareipour, and D. Wood, Distinguished Service Award, the 2013 IEEE ComSoc MMTC Outstanding
“Day-ahead power output forecasting for small-scale solar photo- Leadership Award, and the NSF CAREER Award in 2010. He was a co-
voltaic electricity generators,” IEEE Trans. Smart Grid, vol. 6, no. 5, recipient of the Best Demo Award from IEEE SECON 2017, the Best
pp. 2253–2262, Sep. 2015. Paper Awards from IEEE GLOBECOM in 2015 and 2016, IEEE WCNC
[19] C. Wan et al., “Photovoltaic and solar power forecasting for smart 2015, and IEEE ICC 2013, and the 2004 IEEE Communications Society
grid energy management,” CSEE J. Power Energy Syst., vol. 1, no. 4, Leonard G. Abraham Prize in the Field of Communications Systems. He
pp. 38–46, Dec. 2015. is a Distinguished Lecturer of the IEEE Vehicular Technology Society. He is
[20] Y. Wang, Y. Shen, S. Mao, G. Cao, and R. M. Nelms, “Adaptive learning on the Editorial Board of the IEEE T RANSACTIONS ON M ULTIMEDIA, the
hybrid model for solar intensity forecasting,” IEEE Trans. Ind. Informat., IEEE I NTERNET OF T HINGS J OURNAL (Area Editor), IEEE M ULTIMEDIA
to be published, doi: 10.1109/TII.2017.2789289. and ACM GetMobile.
[21] L. Martín et al., “Prediction of global solar irradiance based on time
series analysis: Application to solar thermal power plants energy pro- Yu Wang received the B.E. degree in measur-
duction planning,” Solar Energy, vol. 84, no. 10, pp. 1772–1781, ing and control technology and instrumentation and
Oct. 2010. the M.E. degree in instrument science and technol-
[22] University of Massachusetts. The UMass Trace Repository. ogy from Southeast University, Nanjing, China, in
Accessed: Mar. 2018. [online] Available: http://traces.cs.umass.edu/ 2008 and 2011, respectively, and the M.S. degree in
index.php/Sensors/Sensors/ probability and statistics and Ph.D. degree in elec-
[23] R. Zhang, Z. Huang, and Y. Lv, “Statistical inference for the index trical computer engineering from Auburn University,
parameter in single-index models,” J. Multivariate Anal., vol. 101, no. 4, Auburn, AL, USA, in 2015.
pp. 1026–1041, Apr. 2010. Since 2015, he has been an Assistant Professor
[24] R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. Roy. with the Department of Electrical Engineering,
Stat. Soc. B Methodol., vol. 58, no. 1, pp. 267–288, Jun. 1996. Nanjing University of Aeronautics and Astronautics,
[25] P. Zeng, T. He, and Y. Zhu, “A lasso-type approach for estimation and Nanjing. His current research interest includes smart grids and optimization.
variable selection in single index models,” J. Comput. Graph. Stat.,
vol. 21, no. 1, pp. 92–109, Apr. 2012. R. M. Nelms (F’04) received the B.E.E. and
[26] W. Y. Hwang, H. H. Zhang, and S. Ghosal, “FIRST: Combining for- M.S. degrees in electrical engineering from Auburn
ward iterative selection and shrinkage in high dimensional sparse linear University, Auburn, AL, USA, in 1980 and 1982,
regression,” Stat. Interface, vol. 2, no. 3, pp. 341–348, Jun. 2009. respectively, and the Ph.D. degree in electrical engi-
[27] S. Luo and S. Ghosal, “Forward selection and estimation in high dimen- neering from the Virginia Polytechnic Institute and
sional single index models,” Stat. Methodol., vol. 33, pp. 172–179, State University, Blacksburg, VA, USA, in 1987.
Dec. 2016. He is currently a Professor and the Chair of the
[28] M. G. Kendall, “A new measure of rank correlation,” Biometrika, vol. 30, Department of Electrical and Computer Engineering,
nos. 1–2, pp. 81–89, Jun. 1938. Auburn University. His current research interests
[29] R. E. Barlow, D. J. Bartholomew, J. Bremner, and H. D. Brunk, include power electronics, power systems, and elec-
Statistical Inference Under Order Restrictions: The Theory and tric machinery.
Application of Isotonic Regression. New York, NY, USA: Wiley, 1972. Dr. Nelms is a registered Professional Engineer in the State of Alabama.