Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Jul 15, 2019

Content may be subject to copyright.

1

An Enhanced Very Short-Term Load Forecasting

Scheme Based on Activation Function

Adamu Sani Yahaya1, Nadeem Javaid1,∗, Kamran Latif2and Amjad Rehman3

1COMSATS University Islamabad, Islamabad 44000, Pakistan

2National Institute of Electronics, Islamabad 44000, Pakistan

3MIS Department COBA, Al Yamamah University, Riyadh 11512, Saudi Arabia

∗nadeemjavaidqau@gmail.com

Abstract—In this paper, we proposed a framework for ac-

curate load forecasting which consists of two stage processes;

feature engineering and classiﬁcation. Feature engineering con-

sists of feature selection and extraction. Relevant features are

selected by combining Decision Tree (DT) and Recursive Feature

Elimination (RFE) techniques. Moreover, Linear Discriminant

Analysis (LDA) technique is used to further improve the selected

features in terms of redundancy and dimensionality reduction. To

forecast the electricity load, an improved feedforward multilayer

perceptron classiﬁer is applied. Half a day ahead forecasting

experiment is conducted by using the proposed framework. At the

end, forecasting performance is examined by using Root Mean

Square Error, Mean Absolute Error, Mean Square Error and

Mean Absolute Percentage Error. Simulation results show higher

accuracy of our proposed scheme with 1.397% as compared to

the existing scheme.

Index Terms—Electricity load forecasting, Feature selection,

Features extraction, Feedforward multilayer perceptron, Accu-

racy.

I. INTRODUCTION

Intelligent power grid, also known as smart grid (SG) is a

power system that efﬁciently manages consumption, distribu-

tion and generation of energy through advanced technologies.

The communication between the utilities and consumers in

smart grid is done in two-way peer-peer i.e., from utility to

consumer and vice versa [1]. In this world, utilization of

energy is necessary and valuable asset. Increase in energy

consumptions from various sectors such as commercial, resi-

dential, industrial and transportation escalated burden on tradi-

tional power grids. Increase in load consumption forces utility

to migrate from traditional power grid to smart grid. With

the implementation of different techniques on utility, power

grid and demand side, Smart grid manages the consumption,

distribution and generation.

In smart grid, data analysis and information communication

technology are in use into every angle of the energy system

such as energy distribution, generation, transmission and also

in appliances of consumer. Smart grid technology is rapidly

increasing in order to meet the increasing demand for energy

delivering in a ﬂexible, robust and in cost effective manner. As

a result of the increase in data generated from smart grid, data

analysis becomes the topic of discussion. Valuable information

from the huge amount of data generated from smart grid can

easily be extracted by data analytics techniques. The valuable

TABLE I: List of Symbols

Symbols Description

f(x)improved Improved activations function

f(x)Logistic sigmoid function

g(x)Rectify linear unit function

PiThe forecasted electricity load value

RiThe actual electricity load value

TABLE II: Abbreviation Table

Abbreviation Description

AC AutoCorrelation

CFS Correlation-Based Selection

DE Differential evolution

DT Decision Tree

EMD Empirical mode decomposition

ELM Extreme learning machine

FOA Fruit Fly Optimization

FS Feature Selection

GPSO Global Best Particle Swam Optimization

GCA Grey Correlation Analysis

GWO Grey Wolf Optimization

IMF Intrinsic mode functions

KELM kernel extreme learning machine

KPCA Principle Component Analysis and Kernel function

LDA Linear Discriminant Analysis

LSTM Long Short-Term Memory

LTLF Long-Term Load Forecasting

MAE Mean Absolute Error

MAPE Mean Absolute Percentage Error

MTLF Medium-Term Load Forecasting

MSE Mean Square Error

PSO (SDPSO) new switching delayed PSO.

QLD New Queensland

RBFNN Radial basis function neural network

RF Random Forest

RF RRelief

RFE Recursive Feature Elimination

RMSE Root Mean Square Error

SDA Stacked denoising autoencoder

SG Smart Grid

STLF Short-Term Load Forecasting

SVM Support Vector Machine

VSTLF Very Short-Term Load Forecasting

information can be use in algorithms development and scheme

designed.

Machine learning and pattern recognition are among the

data analytics techniques that been developed and adopted

especially in the area of load and price forecasting, data

mining of consumer behaviors and fault detection in energy

networks [2]. The research in electricity load forecasting is

not new, is begin in earlier 1965 [3]. Forecasting electricity

2019 International Conference on Computer and Information Sciences (2019 ICCIS) 978-1-5386-8125-1 c

2019 IEEE

2

load accurately is paramount important in decision-making

process for economic load dispatch, power unit commitment,

contingency, power system operation, scheduling, security and

so on [3], [4].

Proper decision may results in minimization of cost of elec-

tricity and power loss [5]. According to [4] one percent (1%)

reduction of mean absolute percentage error in load forecasting

saved 10,000Mw electricity energy which may lead to saving

approximately $1.6 million per year. With these problems,

researchers paid more attention on how to resolve the power

scheduling problem. Many techniques had been used to opti-

mize power scheduling problem [6]–[10]. However, a precise

load forecasting model is required to properly plan on how

to effectively managed power grid resources. Redundancy and

nonlinearity in the historical electricity load data causes huge

challenges in load forecasting. The role of electricity load

forecasting is increasing in electricity generation, system oper-

ation, transmission and storage while forecasting accurately is

becoming a major concern in the management [11]. Therefore,

many researchers focused on how to promote the accuracy and

robustness of the electricity load forecasting. According to [3],

[12] and [13] electricity forecasting horizons are categorized

into Very Short-Term Load Forecasting (VSTLF), Short-Term

Load Forecasting (STLF), Medium-Term Load Forcasting

(MTLF) and Long-Term Load Forecasting (LTLF) as shown

in Table III. In this research, very short-term load forecasting

horizon will be the main focus. Modiﬁed feedforward mul-

tilayer perceptron will be used as classiﬁer to forecast the

electricity load.

A. Motivation

A lot of researches had been conducted in electricity load

forecasting as in [3], [11]–[14]. Although, optimal result from

each model is obtained. However, forecasting accuracy of load

and price needs further improvement. As consequence, this

motivated us to propose a model that provides better accuracy

as well as to improve the efﬁciency of the existing models in

literature.

B. Problem Statement

In [13], recurrent deep neural network and feedforward deep

neural network models performance are compared on the basis

of accuracy and computational performance in which they con-

sider short-term forecasting. Also [14], Global Best Particle

Swam Optimization (GPSO) is proposed to optimized Neural

Network bias value in training processes which decrease the

error and increase accuracy.

In this paper, very short-term load forecasting is considered

to predict the variations in load and activation function is

modiﬁed to increase accuracy and decrease error which are

not considered in [13], [14]. Although, the dataset used in

this research is not the same as that of the benchmark models,

However a reasonable result is obtained. The neural network

is improved by combining logistics sigmoid and rectify linear

unit activation functions.

C. Contribution

The main contributions of this research is as follows:

•Proposing an improved electricity load forecasting sys-

tem.

•Combination of Decision tree and Recursive feature elim-

ination are used in features selection to ﬁnd the most

relevant features.

•To reduce features redundancy and dimensionality, linear

discriminant analysis is applied.

•Activation function of Feedforward multilayer perceptron

is improved by combining logistics sigmoid and rectify

linear unit functions which enhanced the accuracy and

decreases the errors in the forecasted result.

The term prediction and forecasting are alternatively used

throughout this paper. The remaining part of this paper is

planned as follows: Section 2 contained the related works,

section 3 gives details of the techniques (Methodology) used

to build up the model and section 4 describes system model,

dataset and their experimental setup. Section 5 describes the

performance estimation while the simulations result obtained

are discussed in section 6. Finally, The conclusion of the

research is discussed in Section 7.

II. RE LATE D WORK

Lots of researches had been done in STLF in order to

improve forecasting accuracy. Kalman ﬁltering, linear regres-

sion methods, exponential smoothing, gray forecasting and

ARIMA [16]–[18] were proposed in the primal stage. These

forecasting models are good in forecasting linear problem

however are not sufﬁcient when processing more complex

non-linear load problem forecasting. Due to the limitations

of these models, accurate forecasting cannot be achieved. The

aforementioned models are based on mathematical statistics.

Automated and more intelligent artiﬁcial neural network tech-

niques were developed to overcome the complex relationship

and non-linearity in electricity time series. When an ANN was

developed, fewer parameters were needed to step up the model.

This directly affects the forecasting performance. This reason

inﬂuence the researchers to come up with hybrid models [19]

which leads to the addition of numerous smart optimizations

techniques like Fruit Fly Optimization Algorithm (FOA), Grey

Wolf Optimization Algorithm (GWO) e.tc [20], [21]. In [5]

cuckoo search is incorporated with Support Vector Machine

(SVM) to improve its parameters. The performance of the

model signiﬁcantly increase in the experimental results. In [22]

New Switching Delayed PSO (SDPSO) is used to enhance

the Extreme Learning Machine (ELM) techniques. The exper-

imental result of the model outperforms Radial Basis Function

Neural Network (RBFNN) model. However, experiment focus

only in STLF. Finding shows that combination of two or more

models can obtain better forecasting result than the single

ANN model. A robust combination model is proposed by

many researchers to integrate more than two forecasting mod-

els [23]. This approach balance the distribution of forecasting

risk and gain beneﬁts of individual models.

Some researchers observed that best forecasting result is

not guarantee with models that has unavoidable deﬁciency.

3

TABLE III: Forecasting Horizons

S/N Category Period range Aim

1 VSTLF The range is between Few minutes, 30 minutes to approximately one day. To control and adjust demand load and price in real-time

2 STLF Started with one day to a week and upto a month To makes a real-time plan for optimal generator unit commitment

and economic dispatch

3 MTLF start with one month upto approximately one year To maintain the balance between generation and demand for few

months to one year

4 LTLF More than one year To plan for the future electricity network conditions

As a result of that combining two or more forecasting model

is proposed to resolve the limitations. A hybrid model of

Boosting algorithm and a multistep forecast are combined to

produce accurate forecasting results in electricity market [24].

In [25], Auto Regressive Moving average and Self-adapting

Particle Swarm Optimization improve the hybrid combina-

tion of the Kernel Extreme Learning Machine (KELM) and

wavelet transform. The result produce by the model was

remarkable. however, computational time is not considered.

A short-term electricity price was forecasted in [26] using

Stacked Denoising Autoencoder (SDA). The result of the

model was compared with multivariate adaptive regression

splines, classical NN, SVM and least absolute shrinkage and

selection operators and better result is obtained. The price

forecasting accuracy decreases in [27] when forecasting wider

range i.e. weeks, month. Four deep learning techniques were

used to predict the electricity price. The techniques used are

DNN, LSTM-DNN, GRU-DNN and CNN model.

Features Selection (FS) contributed immensely in machine

learning to establish a powerful and reliable model. Feature

selection is a process whereby a set of important features with

high correlation with output are selected [28]. Researchers

used different set of feature selection method. For example, in

[29] four feature selection techniques: RRelief (RF), Mutual

Information, Correlation-Based Selection (CFS), and Auto-

correlation (AC) were evaluated. The corresponding experi-

mental results show that AC-NN and RF-NN gives a better

performance. In [30] Grasshopper and evolutionary population

dynamics techniques are used to select more relevant features.

This method performed very well when handling a larger set

of features and on the other hand, its performance reduced

drastically with a smaller set of input features. Forecasting

performance increasing with more relevant features and vice

versa. In [31], K-means and gradient boosting-based weighted

techniques are used to ﬁnd the similarities between the actual

and predicated days. Empirical Mode Decomposition (EMD),

decompose the similar days to residual and several Intrinsic

Mode Functions (IMFs) and LSTM were used to forecast each

decomposed result. New forcasting model is established in

[32]. a hybrid feature selection is proposed by combining

Relief-F algorithm and Random Forest (RF) together with

Grey Correlation Analysis (GCA) in order to minimized

the redundancy. Principle Component Analysis and Kernel

function (KPCA) are integrated to reduced dimensionality. In

the end, SVM is improved by using differential evolution (DE)

to forecast the electricity price.

III. METHODOLOGY

A. Decision Tree

Decision tree aim to ﬁnd the purest child possible nodes

with minimum split in order to classiﬁed the instances. In this

model “Information gain” is used as the attribute selection

measures to ﬁnd purest child nodes.

Algorithm 1 Decision Tree

1: S, where S= set of classiﬁed instances

2: Require:S6=∅, numattribute > 0

3: procedure DECISION TREE :

4: Repeat:

5: maxGain ←0

6: splitA ←null

7: e←Entropy(attributes)

8: for all a in S do

9: gain ←InformationGain(a,e)

10: if gain >maxGain then

11: maxGain ←gain

12: splitA ←a

13: end if

14: end for

15: Partition(S, splitA)

16: Untill all partition processed

B. Recursive Feature Elimination

Description on how iterative Recursive Feature Elimination

(RFE) works is provided as follows:

1) The classiﬁer will be trained.

2) The ranking criterion for all features will be calculated.

3) Eliminate all the features with the lowest ranking crite-

rion.

The algorithm of SVM-RFE is described in algorithm 2

Algorithm 2 Recursive Feature Elimination

1: procedure RFE:

2: classiﬁer training :α=train −SV M (x, y)

3: weight vector is computed of the demension legth :

4: w=Pkαkxkyk

5: ranking criterias is computed : ci= (wi)2

6: Find the feature with the smallest ranking criteria :

7: f=argmain(c)

8: Eliminate feature with c=f

The stop condition can be based on the number of features

or the classiﬁcation accuracy.

4

C. Linear Discriminant Analysis

The pseudocode for LDA is depicted in Algorithm 3. Where

1nis a vector of all ones and 1n∈R, with the appropriate

dimension for each class i = 1, 2. Linear Discriminant Analysis

proceeds to compute the in-between and within class scatter

matrices, B and S, after The data is dividing into two groups

G1 and G2. The best LD vector is obtained as the dominant

eigenvector of S−1B.

Algorithm 3 Linear Discriminant Analysis

1: procedure LDAG={(xT

i, yi)}n

i=1:

2: speciﬁc-class subsets :

3: Gi={(xT

j|yj=ci, j = 1, .., n}, i = 1,2

4: class means :

5: βi=mean(Gi), i = 1,2

6: scatter in-between matrix class :

7: B= (β1−β2)(β1−β2)T

8: Centeric matrices class :

9: zi=Di−1nβT

i, i = 1,2

10: scatter Class matrices :

11: Pi=zT

izn, i = 1,2

12: class within scatter matrix :

13: P=P1+P2, i = 1,2

14: computed eigenvector is obtained :

15: λ1,w=eigen(P−1B)

D. Feed Forward Neural Network

The Multilayer feedforward NN is divided into three layers.

The layers are input layer, hidden layers and output layers

(Figure 1) [14].

Furthermore, relevant features with lowest error forecast

were used as input of feedforward neural networks. In this

research, an enhanced multilayer perceptron was used to

train the network. The activation function of the model is

improved by combining the logistic sigmoid function (equ.

1) and rectify linear unit (equ. 2). Both activations function

happened to produced good result compared to the remaining

functions. The new activation function f(x)improved is used

to improve accuracy of electricity load forecasting.

f(x) = 1

1 + e−x(1)

g(x) = (0, if x<0

1, if x≥0(2)

Combining equation 1 and 2 we have the improved function

f(x)improved = 2(f(x)∗g(x)) (3)

where the constant value 2 is a scaling factor

Fig. 1: Feedforward Neural Network [15]

IV. SYS TE M MO DE L

Split load dataset and apply

normalization

Historical

electrical load

Feature selection using decision

tree and recursive feature

elimination

Training feature vectors

Feature extraction

using linear

discriminant analysis

Prediction algorithm using

enhanced FF-NN

Prediction result

Fig. 2: Proposed System Model

The electricity load demand data set was collected from

New Queensland (QLD) Australia [33]. The collected load

data are used to verify the authenticity of the proposed model

performance. The load data are sampled for every half an

hour, which means there are 336 observations per week and 48

observations per day. The data set is covered from 11-10-2018

21:00–14-10-2018 4:00 in QLD.

The collected data are used as input variable of the forecast

model. After carrying out some preprocessing, predictive anal-

ysis is performed with enhanced feedforward neural network

method as shown in ﬁgure 2.

In the preprocessing step, data is normalized and then

divided into three parts: train, test and validation. Following

the preprocessing, the data is then given to hybrid features

5

selection. The hybrid features selection is the combination of

decision tree and recursive feature elimination, which are used

to select the relevant features. The features selected by the

hybrid feature selection are assumed to have fewer irrelevant

features; however, dimensionality and redundancy are further

reduced to improve the forecasting accuracy. Linear discrimi-

nant analysis (LDA) is applied to remove the redundancy and

reduce dimensionality of the dataset.

Low weighted and less redundant features have been ﬁltered

after the processes with feature selector and feature extraction.

Finally, the ﬁltered data will be sent to the enhanced feed

forward neural network to forecast the electric load.

V. PERFORMANCE ESTIMATION

In order to measure the forecasting performance of the

model, four standard evaluating metrics are used: Root

Mean Square Error (RMAE),Mean Absolute Percentage Error

(MAPE), Mean Absolute Error (MAE) and Mean Square Error

(MSE). They are deﬁned in equations (1)-(4) [3], [12], [13].

M AP E =1

w

w

X

i=1

|Ri−Pi

Ai

|(4)

MAE =1

w

w

X

i=1

|Ri−Pi|(5)

RM SE =v

u

u

t

w

X

i=1

(Ri−Pi)2

Ri

(6)

MSE =

w

X

i=1

(Ri−Pi)2

Ri

(7)

where Piis the forecasted electric load value at point i,Riis

the actual electric load value at point i;wis the total number

of data; .

VI. SIMULATIONS RESU LT

Numerical results are provided in this section to validate the

new forecasting model. Figure 4 depicted the normalized loads

of QLD for three days (half an hour basis). The normalization

graph shows different variations across the hours. In the

implementation, data are divided in the ratio 75:25. Where

training data has 41.5 hours and testing data has 14 hours. The

processed data from feature selection and extraction techniques

are then forwarded to the forecasting engine for predictions

as shown in ﬁgure 3. The data of features selection and

dimensionality reduction is given in Table IV.

TABLE IV: Attribute of Data

S/N Features Status

1 Date R

2 Net Import (MW) R

3 Spot Price ($/MWh) S

4 Scheduled Generation (MW) S

5 Semi Scheduled Generation

(MW)

S

6 Type R

NOTE: The ‘S’ means that the feature is relevant and should

be selected, the ‘R’ means the feature is irrelevant and should

be rejected.

Fig. 3: Comparison of the Forecasted and Actual Load in QLD.

Fig. 4: Normalized load of QLD from 11-10-2018 21:00–14-

10-2018 4:00.

Fig. 5: Comparison of Performances Evaluators.

TABLE V: Errors evaluation metrics for the QLD market for

the days 11-10-2018 21:00–14-10-2018 4:00

Method MSE MAE RMSE MAPE

Existing

Model

24.505 14.289 17.502 3.745

New Model 10.455 9.05 11.432 2.366

6

The actual and forecasted loads for the enhanced and existing

model of the QLD is presented in ﬁgure 3. In this ﬁgure,

blue curves represent the actual load, the red curve is the

forecasted load for the existing model while the yellow curve

is for the enhanced model. As we can see from ﬁgure 3, the

performance of the enhanced model shows good result, since

the load curve of the enhanced model is more closely to the

actual value as compared to the existing model. Error analysis

performance result can be obtained from table V. Figure 5

shows the comparison within performance evaluators. Table

V records four forecasting errors about the proposed and

existing model in the experiment. As shown in table V, the

result of performance apparently shows that the proposed

model is reasonably good because it has the lowest error in

MAPE, MAE, MAE, RMSE and MSE. For example, Value

of MAPE is 2.366%, which is apparently smaller than that

of the existing model. This indicates that the new model

outperforms its counterpart model.

VII. CONCLUSION

This research proposed a modiﬁed electricity load forecasting

model by improving the activation function of the existing

model. The corresponding results of the experiment show that

the new model outperforms the existing one with 1.397%. Fea-

ture selection and extraction are used to reduce dimensionality

and redundancy in order to give a more relevant features to

predictor.

REFERENCES

[1] Neupane, Bijay, Wei Lee Woon, and Zeyar Aung. ”Ensemble prediction

model with expert selection for electricity price forecasting.” Energies

10.1 (2017): 77.

[2] Ma, Zhanyu, et al. ”The role of data analysis in the development of

intelligent energy networks.” IEEE Network31.5 (2017): 88-95.

[3] Yang, Ailing, Weide Li, and Xuan Yang. ”Short-term electricity load

forecasting based on feature selection and Least Squares Support Vector

Machines.” Knowledge-Based Systems (2018).

[4] Dong, Yongquan, Zichen Zhang, and Wei-Chiang Hong. ”A Hybrid

Seasonal Mechanism with a Chaotic Cuckoo Search Algorithm with

a Support Vector Regression Model for Electric Load Forecasting.”

Energies 11.4 (2018): 1009.

[5] X. Zhang, J. Wang, K. Zhang, et al., Short-term electric load forecasting

based on singular spectrum analysis and support vector machine op-

timized by Cuckoo search algorithm, Electr. Power Syst. Res. (2017)

270285.

[6] Moon, Seokjae, and Jang-Won Lee. ”Multi-residential demand response

scheduling with multi-class appliances in smart grid.” IEEE transactions

on smart grid 9.4 (2018): 2518-2528.

[7] Ahmad, Adnan, et al. ”An optimized home energy management system

with integrated renewable energy and storage resources.” Energies 10.4

(2017): 549.

[8] Ahmad, Ashfaq, et al. ”An accurate and fast converging short-term load

forecasting model for industrial applications in a smart grid.” IEEE

Transactions on Industrial Informatics 13.5 (2017): 2587-2596.

[9] Moon, Seokjae, and Jang-Won Lee. ”Multi-residential demand response

scheduling with multi-class appliances in smart grid.” IEEE transactions

on smart grid 9.4 (2018): 2518-2528.

[10] Javaid, Nadeem, et al. ”A new heuristically optimized Home Energy

Management controller for smart grid.” Sustainable Cities and Society

34 (2017): 211-227.

[11] Jiang, Huaiguang, et al. ”A short-term and high-resolution distribution

system load forecasting approach using support vector regression with

hybrid parameters optimization.” IEEE Transactions on Smart Grid 9.4

(2018): 3341-3350.

[12] Jiang, Ping, Feng Liu, and Yiliao Song. ”A hybrid forecasting model

based on date-framework strategy and improved feature selection tech-

nology for short-term load forecasting.” Energy 119 (2017): 694-709.

[13] Din, Ghulam Mohi Ud, and Angelos K. Marnerides. ”Short term power

load forecasting using deep neural networks.” Computing, Networking

and Communications (ICNC), 2017 International Conference on. IEEE,

2017.

[14] Raza, Muhammad Qamar, et al. ”An intelligent hybrid short-term load

forecasting model for smart power grids.” Sustainable Cities and Society

31 (2017): 264-275.

[15] Ylmaz, Ali Can, Cigdem Inan Aci, and Kadir Aydin. ”MFFNN and

GRNN Models for Prediction of Energy Equivalent Speed Values of

Involvements in Trafﬁc Accidents/Traﬁk Kazalarnda tutulumunun Enerji

Edeer Hz Deerleri Tahmininde MFFNN ve GRNN Modelleri.” Interna-

tional Journal of Automotive Engineering and Technologies 4.2 (2015):

102-109.

[16] Raﬁei, Mehdi, Taher Niknam, and Mohammad-Hassan Khooban. ”Prob-

abilistic forecasting of hourly electricity price by generalization of ELM

for usage in improved wavelet neural network.” IEEE Transactions on

Industrial Informatics 13.1 (2017): 71-79.

[17] Singh, Nitin, Soumya Ranjan Mohanty, and Rishabh Dev Shukla.

”Short term electricity price forecast based on environmentally adapted

generalized neuron.” Energy 125 (2017): 127-139.

[18] Yu, Chun-Nam, Piotr Mirowski, and Tin Kam Ho. ”A sparse coding

approach to household electricity demand forecasting in smart grids.”

IEEE Transactions on Smart Grid 8.2 (2017): 738-748.

[19] W. Li, X. Yang, H. Li, et al., Hybrid forecasting approach based

on GRNN neural network and SVR machine for electricity demand

forecasting, Energies 10 (1) (2017) 117.

[20] Mirjalili, Seyedali. ”Dragonﬂy algorithm: a new meta-heuristic optimiza-

tion technique for solving single-objective, discrete, and multi-objective

problems.” Neural Computing and Applications 27.4 (2016): 1053-1073.

[21] Saremi, Shahrzad, Seyedali Mirjalili, and Andrew Lewis. ”Grasshopper

optimisation algorithm: theory and application.” Advances in Engineering

Software 105 (2017): 30-47.

[22] Zeng, Nianyin, et al. ”A switching delayed PSO optimized extreme

learning machine for short-term load forecasting.” Neurocomputing 240

(2017): 175-182.

[23] Xiao, Liye, et al. ”Research and application of a combined model based

on multi-objective optimization for electrical load forecasting.” Energy

119 (2017): 1057-1074.

[24] Jiang, Yu, et al. ”Short-term wind power forecasting using hybrid

method based on enhanced boosting algorithm.” Journal of Modern Power

Systems and Clean Energy 5.1 (2017): 126-133.

[25] Yang, Zhang, Li Ce, and Li Lian. ”Electricity price forecasting by a

hybrid model, combining wavelet transform, ARMA and kernel-based

extreme learning machine methods.” Applied Energy 190 (2017): 291-

305.

[26] Wang, Long, Zijun Zhang, and Jieqiu Chen. ”Short-term electricity price

forecasting with stacked denoising autoencoders.” IEEE Transactions on

Power Systems 32.4 (2017): 2673-2681.

[27] Lago, Jesus, Fjo De Ridder, and Bart De Schutter. ”Forecasting spot

electricity prices: Deep learning approaches and empirical comparison of

traditional algorithms.” Applied Energy 221 (2018): 386-405.

[28] Abedinia, Oveis, Nima Amjady, and Hamidreza Zareipour. ”A new

feature selection technique for load and price forecast of electrical power

systems.” IEEE Transactions on Power Systems 32.1 (2017): 62-74.

[29] Koprinska, Irena, Mashud Rana, and Vassilios G. Agelidis. ”Correlation

and instance based feature selection for electricity load forecasting.”

Knowledge-Based Systems 82 (2015): 29-40.

[30] Mafarja, Majdi, et al. ”Evolutionary population dynamics and grasshop-

per optimization approaches for feature selection problems.” Knowledge-

Based Systems 145 (2018): 25-45.

[31] Zheng, Huiting, Jiabin Yuan, and Long Chen. ”Short-term load fore-

casting using EMD-LSTM neural networks with a Xgboost algorithm for

feature importance evaluation.” Energies 10.8 (2017): 1168.

[32] Wang, Kun, et al. ”Robust big data analytics for electricity price

forecasting in the smart grid.” IEEE Transactions on Big Data (2017):

1-12.

[33] https://www.aemo.com.au/Electricity/National-Electricity-Market-

NEM/Data-dashboard. accessed date: 10/11/2018.