Content uploaded by Nadeem Javaid
Author content
All content in this area was uploaded by Nadeem Javaid on Jul 13, 2019
Content may be subject to copyright.
A Comparative Analysis of Neural
Networks and Enhancement of ELM
for Short Term Load Forecasting
Rahim Ullah1,2, Nadeem Javaid1(B
), Ghulam Hafeez1,3, Salim Ullah1,
Fahad Ahmad4, and Ashraf Ullah1
1COMSATS University Islamabad, Islamabad 44000, Pakistan
nadeemjavaidqau@gmail.com
2Higher Education Department, KP, Peshawar 25000, Pakistan
3Wah Engineering College, University of Wah, Wah Cantt 47040, Pakistan
4International Islamic University, Islamabad 44000, Pakistan
http://www.njavaid.com
Abstract. Smart grid is the evolved form of traditional power grid with
the integration of sensing, communication, computing, monitoring and
controlling technologies. These technologies make the power grid reliable,
efficient and economical. Smart grid enable users to make bidding on the
basis of demand side management models. Demand side management
can be made responsive and efficient by effective and accurate load fore-
casting. Accurate load forecasting is an important but challenging task
because of irregular and non-linear consumption of individual users and
industrial consumers. Different approaches have been proposed for load
forecasting, but artificial intelligent models, specifically ANNs perform
well for short, medium and long term load forecasting. The main focus
of this paper is to present a comparative analysis of NNs for short term
load forecasting. NYISO dataset is used for experiments. XGboost and
decision tree are used for features importance calculation and RFE is
used for features extraction on the basis of score assigned. Three basic
techniques (CNN, MLP and ELM) are used for forecasting. Further-
more, ELM is enhanced and E ELM is proposed for STLF. Moreover,
results are evaluated on four statistical measures (MAPE, MAE, MSE
and RMS).
Keywords: Smart grid ·Load forecasting ·Neural networks ·
Features ·Features engineering ·Classifiers ·
Machine learning techniques ·XGboost ·DT ·ELM ·
EELM ·MLP ·CNN
1 Introduction
Short term load forecasting is very important for the trust-able, secure and
efficient execution of electricity system. Forecasted load can help in improv-
ing important decisions for economic activities. Such decisions include flow of
c
Springer Nature Switzerland AG 2020
L. Barolli et al. (Eds.): CISIS 2019, AISC 993, pp. 73–86, 2020.
https://doi.org/10.1007/978-3-030-22354-0_7
74 R. Ullah et al.
load, transaction evaluation, generator unit commitment, coordination of ther-
mal units, fuel allocation, analysis of network, short and long term maintenance,
protection of the system, load balancing and contingency planning. As load
forecasting possesses a certain amount of error, which can cause greater loss.
Moreover, as industries have started relying on forecasting, the importance has
accuracy increased. An accurate forecasting technique can help in developing an
efficient action plan. Optimal action plan can help in the reduction of risk and
improvement of economical benefits of management [1].
Table 1. List of abbreviations used in this paper
Abbreviation Explanation
STLF Short Term Load Forecasting
LTLF Long Term Load Forecasting
ARIMA Autoregressive Integrated Moving Average
FWPT Flexible Wavelet Packet Transform
CMI Conditional Mutual Information
NLSSVM Non-linear Least Square Support Vector Machine
ABC Artificial Bee Colony
CFS Correlation-based Feature Selector
AI Artificial Intelligence
ANN ArtificialNeuralNetwork
RFE Recursive Features Elimination
NYISO New York Independent System Operator
RNN Recurrent Neural Network
ML Machine Learning
MLP Multi Layer Perceptron
CNN Convolutional Neural Network
ELM Extreme Learning Machine
EELM Enhance Extreme Learning Machine
MTR Model Tree Rule
RMS Root Mean Square
MSE Mean Squared Error
MAE Mean Absolute Error
MAPE Mean Absolute Percent Error
Accurate STLF is a challenging job. Electricity load has complex and non-
linear cycles on daily, weekly and annual basis. There can be some random
components as well because of irregular, unplanned and unpredicted use of elec-
tricity by individual users, building consumers and large industries. Irregularities
may also be caused by long term operation of industries, sudden weather change,
A Comparative Analysis of Neural Networks and Enhancement of ELM 75
special events and holidays which make load forecasting challenging [2]. List of
abbreviations is given in Table 1.
There are four classes of load forecasting on the basis of horizon: short term
is for days to weeks ahead forecasting, very short term is for minutes to hours
ahead forecasting, medium term is for weeks to months ahead and long term is
for years ahead load forecasting. This paper focus on short term load forecasting
for New York on the basis of NYISO dataset. It can be used by market operators
for setting demand requirements and by the consumer to make bids. Smart grid
further enhances the importance of STLF by demand response mechanism and
time varying price, as they require prediction at short intervals [3].
Load forecasting can be performed by three methods: traditional statistical
method, advanced machine learning models and hybrid models. As illustrated
earlier, short term load data has many non-linear and non-stationary compo-
nents which cannot be captured well by statistical models [4]. On the other
hand, a number of ML techniques have shown better performances in differ-
ent scenarios [5]. In ML, there are a number of techniques which can be used
for load forecasting; however, ANNs possess many advantages over the others
such as: capability to capture non-linear relationship between input features and
predictor variables, making of patterns instead of assumptions and tolerance to
noise.
Because of such advantages, ANNs are used in too many studies, but there are
a number of flavors of them. It is not evident which one is suitable for short term
load forecasting. The main focus of this papers is to give a comparative analysis
of three NNs (CNN, MLP and ELM) for STLF. Furthermore, we enhance the
performance of ELM by tuning the parameters iteratively and proposed E ELM
for STLF.
In this paper, a comparative study of NNs for STLF is given using NYISO
dataset. Dataset has sixteen features; however, all features are not contributive
to output. First we extract the important features. For features extraction, the
importance of features are calculated by XGboost and DT. Afterward, features
are extracted by RFE on the basis of score assigned. The dataset was then split
into training and evaluation sets and models were trained by training set while
evaluated on testing set. The results are evaluated on four statistical evaluators
(MAE, MAPE, RMS and MSE).
The remaining paper is organized as, Related Work is discussed in Sect.2,
Problem Statement and Contribution is briefly discussed in Sect. 3, Sect. 4
presents Proposed System while Sect. 5explains Simulation Setup. Section 6is
dedicated to Critical Analysis and Results and Discussion are elaborated in
Sect. 7. Paper is concluded in Sect. 8while References are presented at the end
of paper.
2 Related Work
With the introduction of competitive electricity market, the importance of STLF
is increased. Smart grid has boosted further the importance of STLF. It can be
76 R. Ullah et al.
performed with the help of three main approaches, traditional based on statistical
tools, advanced machine learning based and hybrid. ARIMA and exponential
growth are the two very prominent techniques in traditional approach [6]. There
are a bunch of techniques in second approach. Because of their advantages and
efficiency, these are used in many studies from different perspectives [5].
In [7], a hybrid model for load and price forecasting with demand side man-
agement has been given. Simultaneous price and cost forecasting is performed
by using different datasets in which the linear components are fed to its suit-
able model and non-linear are fed to the other. The proposed model has three
main components, the first one consists of FWPT which decomposes the sig-
nal into many components having distinct frequencies. CMI, a new model for
features selection is used for the selection of useful and valuable features. After-
ward, ARIMA is used to forecast the linear component while NLSSVM is used
for non-linear components. Moreover, to optimize the parameters of SVM, an
improved version of ABC based on time varying coefficient and stumble gener-
ator operator is used. The proposed system performed very well for various real
market datasets in forecasting price and load.
A load forecasting system based on data pre-analysis and weight co-efficient
optimization is presented in [8]. Experiments are performed with three datasets
having half-hourly updated from state of New South Wales, state of Victoria and
State of Queensland for almost three consecutive years. A number of algorithms
are used for data pre-processing, feature selection and forecasting such as, cuckoo
search algorithm is used for optimized weight assignment. A number of ANNs
are used for load forecasting. Obtained results are compared with benchmark,
ARIMA. The proposed system outperforms ARIMA for all the datasets and also
overcomes the problem of instability and unitary problem.
Koprinska et al. in [2], presented a study for load forecasting. The main focus
of this paper is on the selection of important features. Four methods are used for
the selection of appropriate features in order to give a comparative analysis of
these features selectors. Among the features selection methods, one is the tradi-
tional statistical method (autocorrelation), while, the other three are advanced
ML approaches (MI, RReliefF and CFS) for features selection integrated with
forecasting algorithms. Load is forecasted on three ML algorithms (NN, Linear
Regression and MTR) for two years Australian market data. Results show that
all features selectors select a subset from the features set; however, AC and ANN
performed very well. The proposed system is compared with existing work and
some industry benchmarks.
Electricity load forecasting using RNN is presented by Zheng et al. in [9].
According to author’s the load data of power system has non-linear, non-seasonal
and non-stationary patterns which cannot be predicted accurately by statistical
techniques and simple AI techniques. Authors proposed the use of RNN for
long short term load forecasting. Long STLF deals with the short term load
forecasting having long term dependencies. The system is evaluated for different
horizons and found comparably efficient.
A Comparative Analysis of Neural Networks and Enhancement of ELM 77
In [10], the authors proposed complex load decomposition for forecasting the
aggregated load of a campus. The main aim is that different components of
campus have different load profile and forecasting load as a whole is illogical.
The focus of this article is to decompose the overall load into its representative
components. Afterward, cluster the loads with the similar profiles and perform
the day ahead forecast for each cluster. It is proved that the system can performs
better for LTLF if data of long term is available. The system is evaluated on
different statistical measures and found well as compared to other state of the
art techniques.
Nowotarski et al. in [11] aimed to improve the accuracy of short term load
forecasting by combining sister forecasts. According to the authors, all forecast-
ing algorithms are sisters and combining them will improve load forecasting.
Eleven algorithms are combined with two features selection techniques. Sister
forecasts are demonstrated with two case studies. Sisters forecast outperforms
the benchmark algorithms in term of accuracy and MAPE. Authors suggest that
sisters load forecast has high accuracy for academic and practical values.
Beside the above mentioned studies, too many other studies can be found
using NNs for load forecasting; however, no study can be found for short term
load forecasting and comparing the mentioned techniques.
3 Problem Statement and Contribution
Electricity load forecasting is an important activity for supply markets as well
as consumers. Three approaches are in practice for load forecasting, elaborated
in next section. ML has significant benefits over traditional models [12]. ML has
a number of techniques each with some unique characteristics, but NNs have
shown better performance in too many studies. A number of flavors of NNs
are available and used in too many studies; however, it is not evident which
technique is suitable for STLF. Furthermore, a comparison of NNs for STLF is
not available.
The main contribution of this paper can be summarized as follows.
•The use of an integrated model for features importance calculation and fea-
tures selection.
•Presentation of a comparative analysis of NNs for STLF.
•Enhancement of ELM for short term load forecasting by iterative parameters
tuning. In other words, proposed E ELM for short term load forecasting which
beats ELM.
4 Proposed System Model
Short term load forecasting with high accuracy has prominent importance in
electricity market. It is fruitful for decisions regarding reliable, secure and opti-
mal use of electricity [13]. On the other hand, smart grid planning, transactions
and investment are also using load forecasting for optimal decisions. However,
78 R. Ullah et al.
because of diverse factors such as change in weather and change in social behav-
ior of consumers, load forecasting is a challenging task. A number of models have
been proposed for load forecasting in recent years. These models are categorized
as: AI models, time series models and hybrid models which has the good features
of both.
Time series model is used by a number of studies [14–16]; however, time
series models do not capture the diverse factors of power system, that is why
cannot produce good accuracy [4]. In contrast, artificial intelligence models can
capture the diverse features and the non-linear behavior in a very good manner
and produce good result. Therefore, it is used by a number of studies for load
forecasting in so many studies [17–19]. That is why, this paper focus on second
approach for STLF.
The proposed system mainly consists of four main components, as shown in
Fig. 1. The first component at left most is the dataset which is NYISO dataset of
three years (2015, 2016 and 2017), publicly available. The details of using dataset
is given in section-IV, simulations setup. Second part is feature engineering in
which we try to extract the important features, which can help in good forecast-
ing. The main goal of features engineering is to increase classification accuracy,
reduce the dimensions of data so that complexity can be avoided and time of
processing can be reduced [20]. Three techniques are used for features engineer-
ing in the proposed system. Firstly, the importance of features are calculated by
XGboost and decision tree. Afterward, RFE is used to discard the low important
features from the dataset. The features with high importance are left, which can
help in achieving good accuracy.
Fig. 1. System model of the Proposed System where LD stands for Loaded Data, RFE
for Recursive Features Elimination, EFs for Extracted Features and PVs for Predicted
Values
The third part of the system model deals with load forecasting. Before feeding
data to classifiers, it is divided into training and testing sets with ratio 3:1. The
data of first nine months of each year is used as training set and the rest of the
three months is used as testing set. Four classifiers such as CNN, MLP, ELM
and E ELM are used. We have enhanced ELM and proposed E ELM for load
forecasting. A number of studies have used different ANN techniques, but it is
not evident which one is suitable for load forecasting. The core purpose of this
A Comparative Analysis of Neural Networks and Enhancement of ELM 79
paper is to present a comparative analysis of ANN techniques for short term
load forecasting and enhance ELM.
Fourth part is about the evaluation of the proposed system. Four statistical
measure are used for evaluation such as MAPE, MAE, RMS and MSE. The
details of theses errors are discussed in detail in Sect. 6.
5 Experimental Setup
Experiments are performed in Python. The three years data of NYISO is taken
for experiments [21]. Data has sixteen features and 1095 instances, consists of
system load recorded for each day and many other attributes described in Table2
below. As all the features are not supportive to load forecasting, therefore, in first
step we extract the important features. For features extraction, the importance
of features are calculated by XGboost and decision tree which are state of the
art and most efficient technique. A score is assigned to each feature separately
by each technique. The score assigned by XGboost is shown in Fig. 2, while that
of DT is shown in Fig. 3as follow. Afterward, features are selected by RFE on
the basis of score assigned to it. The combined score of DT and XGboost is
used for features extraction. Among sixteen features, System Load is used as
label, eight important features (DA Demand, RT Demand, DA LMP, DA EC,
RT EC, Dry Bulb, Dew Point, Reg Cap Price) are extracted and seven (DA CC,
DA MLC, RT LMP, RT CC, RT MLC, Reg Ser Price) are discarded.
Table 2. Features of dataset and its description
Attribute Description
Date Date of Data Recording
DA DEMAND The Day Ahead Demands
RT DEMAND The Day Locational Demands
DA LMP The Locational Marginal Price One Day Ahead
DA CC The Congestion Part of Day Ahead Price
DA EC The Energy Part of Day Ahead Price
DA MLC The Marginal Loss Part of the Day Ahead Price
RT LMP The Dynamic Locational Marginal Price
RT EC The Energy Part of the Dynamic Price
RT CC The Congestion Part of Dynamic Price
RT MLC The Marginal Loss Part of Dynamic Price
Dry Bulb The Dry Bulb Temperature (F◦)
Dew Point The Dew Point Temperature (F◦)
Reg Cap Price Regional Capacity Price
Reg Ser Price Regional Service Price
System Load The Per Day Load of the System
80 R. Ullah et al.
Fig. 2. Importance of features calculated by XGboost
Fig. 3. Importance of features calculated by decision tree
After features extraction, the next step is to feed the extracted features to
classifiers, but before giving data to classifiers, the data was split into training
and evaluation sets. Data of nine months of each year is kept in training set
while the rest of three months of each year is kept as evaluation set. Afterward,
the models are trained with training set. The primary justification of this paper
is to present a comparison of basic NNs and one enhanced ELM for STLF.
These three techniques (MLP, CNN and ELM) are being used frequently for
load forecasting, but it is not clear which one is the most suitable for short term
load forecasting. After training, the performance of classifiers were evaluated by
testing set.
A Comparative Analysis of Neural Networks and Enhancement of ELM 81
For evaluating the accuracy of classifiers, four statistical measures such as
RMS, MSE, MAPE and MAE are used. Built in functions of MSE and MAE are
available while functions are defined for RMS and MAPE.
6 Simulation Results and Discussion
In this section, we give a discussion about the results obtained. In the first phase,
features engineering is performed to extract best features which contributes a lot
in good accuracy. Two features scoring techniques such as XGboost and DT are
used which assign score to each feature of the dataset. Afterward, RFE eliminates
the unimportant features and extracts the important features. Eight features are
kept and the rest are eliminated. This makes the system simple, accurate and
less time consuming. A number of techniques are available for features selection
and extraction; however, in our case, the above mentioned techniques performed
very well.
Electricity load can be forecast by a number of techniques which are catego-
rized as traditional statistical methods, AI models and hybrid models. As power
data consists of non-linear behavior, which cannot be well captured by statistical
model alone. On the other hand ML algorithms and hybrid model have shown
good results presented in many studies. Moreover, a number of AI techniques
are available, but ANNs has many advantages over other techniques.
Neural networks are the most popular load and price forecasting techniques.
These techniques are widely used by both industries and research forecasters
[2]. NNs possess many advantages over traditional statistical methods (LR and
ARIMA) such as (1) capable to model non-linear relationship between features
and predictor variables. (2) Capable to make patterns from given examples
instead of making assumptions. (3) NNs are tolerant to noise. There are a num-
ber of techniques available in NNs; however, it is not evident which technique is
suitable for short term load forecasting.
In this study we focus to use three basic NNs such as CNN, MLP, ELM
and one enhanced version of ELM(E ELM) for short term load forecasting and
find which one can perform well. Here we present the comparative analysis of the
afore mentioned techniques for NYISO dataset. The trend of prediction is shown
in Fig. 4as under. The trend of forecasting is compared with actual trend and
it can be seen clearly that CNN and E ELM follow the actual trend very well.
It means these two algorithms are performing very well. Furthermore, E ELM
beats the ELM in following trend.
Moreover, the comparison is shown on the basis of different statistical mea-
sures as well. Mean absolute error is shown in Fig.5. As shown in figure, MAE
of CNN is very low, it means it produces very low error as compared to other
techniques. Furthermore, the processing time of CNN is also very short in our
scenario. MAE for E ELM is ranked on second and is better than ELM. Mean
absolute percent error is also calculated for the forecasting of short term load
as shown in Fig. 6. The error rate of CNN in MAPE is a bit unpredictable, it is
higher than E ELM. Here E ELM performs well and outperforms all the tech-
niques. The performance of CNN may increases with increase in the number of
82 R. Ullah et al.
Fig. 4. Trend of forecasting by different techniques
Fig. 5. Mean absolute error score
Fig. 6. Mean absolute percent error score
instances. E ELM works well and has a very short emergence time. The perfor-
mance of RMS is shown in Fig. 7. Root meas square has a bit same pattern as
MAPE. The performance of E ELM is better than others. ELM stands second
A Comparative Analysis of Neural Networks and Enhancement of ELM 83
Fig. 7. Root mean square score
Fig. 8. Mean square error score
here. For calculating RMS, a function is defined using the standard formula of
RMS. In last, the score given by mean squared error is shown in Fig. 8.The
built-in function of sklearn are used for MAE and MSE. The MSE for CNN
is about zero, which indicates that it has very low mean squared error. Again
EELM stands second. The score of ELM is also a bit identical to E ELM, but
EELM performs outstanding here as well. From the above detailed discussion
about results, it is clear that CNN and E ELM are very good for short term
load forecasting because of easy use, fast emergence and very good accuracy.
Moreover, CNN and E ELM can capture the non-linear and noisy data of load
forecasting very well. Furthermore, with the course of time, the data of load will
increase, but the performance of CNN will not decrease, rather will increase.
So, on the basis of our experiments, we suggest the use of CNN and E ELM for
short term load forecasting among before mentioned NNs.
84 R. Ullah et al.
7 Critical Analysis
Beside performance, NNs have certain advantages over other AI and traditional
model for load forecasting such as: tolerance to noise, making of pattern instead
of assumption, handling non-linearity and ease of use [22]. Specifically, we used
three techniques (CNN, ELM and MLP). Reasoning for the use and performance
of each individual technique is briefly elaborated here.
CNN has a multi layered architecture which allows the learning structure to
learn complex relationships between input and predicting variables and extracts
complex patterns from data. Thus, CNN has revolutionized too many fields
including load forecasting [23]. Although CNN is basically for images but have
shown significant performance for load forecasting in too many studies.
MLP is suitable for mapping N-dimensional input to M-dimensional output
with high accuracy. It has multi layers and as the learning process proceeds, it
learns implicit and salient features from data, identifies error signal and improves
accuracy. It is appropriate algorithm for data having high dimensions and load
data has very high dimensions [24].
ELM is a single layer feed forward neural network which chose the
input weights randomly and analytically determines the output weights which
extremely enhance the accuracy and is very fast [25]. In our scenario, it gives
good results in a very short time. Furthermore, we have enhanced the basic ELM
andhaveproposedEELM which further improves the performance and beats
ELM.
8 Conclusion and Future Work
This paper simulated short term load forecasting on NYISO dataset with four
machine learning techniques. XGboost and decision tree are used for features
importance calculation and RFE is used to extract important features. Three
ML techniques (CNN, MLP and ELM) are used for short term load forecasting.
Moreover, ELM is enhanced (E ELM) for short term load forecasting which has
beaten ELM in all cases. Results are evaluated on four statistical measures. It
is found that CNN and ELM performed well in all the cases and we suggest the
use of these two techniques for STLF.
In future, we plan to use the above mentioned techniques for combined load
and price forecasting for short term. Furthermore, it is planned to evaluate the
system on some other datasets as well as compare the system with some indus-
trial bench marks.
References
1. ¨
Onkal, D., Sayim, K.Z., Lawrence, M.: Wisdom of group forecasts: does role-playing
play a role? Omega 40(6), 693–702 (2012)
2. Koprinska, I., Rana, M., Agelidis, V.G.: Correlation and instance based feature
selection for electricity load forecasting. Knowl.-Based Syst. 82, 29–40 (2015)
A Comparative Analysis of Neural Networks and Enhancement of ELM 85
3. Chan, S.-C., Tsui, K.M., Wu, H.C., Hou, Y., Wu, Y.-C., Wu, F.F.: Load/price
forecasting and managing demand response for smart grids: methodologies and
challenges. IEEE Sig. Process. Mag. 29(5), 68–85 (2012)
4. Che, J., Wang, J., Wang, G.: An adaptive fuzzy combination model based on self-
organizing map and support vector regression for electric load forecasting. Energy
37(1), 657–664 (2012)
5. Alfares, H.K., Nazeeruddin, M.: Electric load forecasting: literature survey and
classification of methods. Int. J. Syst. Sci. 33(1), 23–34 (2002)
6. Taylor, J.W.: An evaluation of methods for very short-term load forecasting using
minute-by-minute British data. Int. J. Forecast. 24(4), 645–658 (2008)
7. Ghasemi, A., Shayeghi, H., Moradzadeh, M., Nooshyar, M.: A novel hybrid algo-
rithm for electricity price and load forecasting in smart grids with demand-side
management. Appl. Energy 177, 40–59 (2016)
8. Xiao, L., Wang, J., Hou, R., Wu, J.: A combined model based on data pre-analysis
and weight coefficients optimization for electrical load forecasting. Energy 82, 524–
549 (2015)
9. Zheng, J., Xu, C., Zhang, Z., Li, X.: Electric load forecasting in smart grids using
long-short-term-memory based recurrent neural network. In: 2017 51st Annual
Conference on Information Sciences and Systems (CISS), pp. 1–6. IEEE (2017)
10. Park, K., Yoon, S., Hwang, E.: Hybrid load forecasting for mixed-use complex
based on the characteristic load decomposition by pilot signals. IEEE Access 7,
12297–12306 (2019)
11. Nowotarski, J., Liu, B., Weron, R., Hong, T.: Improving short term load forecast
accuracy via combining sister forecasts. Energy 98, 40–49 (2016)
12. Zhang, J., Wei, Y.-M., Li, D., Tan, Z., Zhou, J.: Short term electricity load fore-
casting using a hybrid model. Energy 158, 774–781 (2018)
13. Hu, R., Wen, S., Zeng, Z., Huang, T.: A short-term power load forecasting model
based on the generalized regression neural network with decreasing step fruit fly
optimization algorithm. Neurocomputing 221, 24–31 (2017)
14. Papalexopoulos, A.D., Hesterberg, T.C.: A regression-based approach to short-
term system load forecasting. IEEE Trans. Power Syst. 5(4), 1535–1547 (1990)
15. Mbamalu, G.A.N., El-Hawary, M.E.: Load forecasting via suboptimal seasonal
autoregressive models and iteratively reweighted least squares estimation. IEEE
Trans. Power Syst. 8(1), 343–348 (1993)
16. Chen, J.-F., Wang, W.-M., Huang, C.-M.: Analysis of an adaptive time-series
autoregressive moving-average (ARMA) model for short-term load forecasting.
Electric Power Syst. Res. 34(3), 187–196 (1995)
17. Rahman, S., Bhatnagar, R.: An expert system based algorithm for short term load
forecast. IEEE Trans. Power Syst. 3(2), 392–399 (1988)
18. Pai, P.-F., Hong, W.-C.: Forecasting regional electricity load based on recurrent
support vector machines with genetic algorithms. Electric Power Syst. Res. 74(3),
417–425 (2005)
19. Pandian, S.C., Duraiswamy, K., Rajan, C.C.A., Kanagaraj, N.: Fuzzy approach
for short term load forecasting. Electric Power Syst. Res. 76(6–7), 541–548 (2006)
20. Chitsaz, H., Zamani-Dehkordi, P., Zareipour, H., Parikh, P.P.: Electricity price
forecasting for operational scheduling of behind-the-meter storage systems. IEEE
Trans. Smart Grid 9(6), 6612–6622 (2018)
21. NYISO: NYISO Electricity Market Data. http://www.nyiso.com/. Accessed 8 May
2019
22. Hayati, M., Shirvany, Y.: Artificial neural network approach for short term load
forecasting for Illam region. World Acad. Sci. Eng. Technol. 28, 280–284 (2007)
86 R. Ullah et al.
23. Dong, X., Qian, L., Huang, L.: Short-term load forecasting in smart grid: a com-
bined CNN and K-means clustering approach. In: 2017 IEEE International Con-
ference on Big Data and Smart Computing (BigComp), pp. 119–125. IEEE (2017)
24. Mori, H., Yuihara, A.: Deterministic annealing clustering for ANN-based short-
term load forecasting. IEEE Trans. Power Syst. 16(3), 545–551 (2001)
25. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K.: Extreme learning machine: a new learning
scheme of feedforward neural networks. Neural Netw. 2, 985–990 (2004)
