Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Oct 04, 2021

Content may be subject to copyright.

Electricity Forecasting for Efﬁcient Energy

Management in Smart Grid

A Post Graduate Thesis submitted to the Department of Computer Science as partial

fulﬁllment of the requirements for the award of Degree of PhD (Computer Science).

Name Registration Number

Aqdas Naz CIIT/FA17-PCS-003/ISB

Supervisor

Dr. Nadeem Javaid

Associate Professor, Department of Computer Science

COMSATS University Islamabad (CUI), Islamabad

Co-Supervisor

Dr Sohail Iqbal

Associate Professor, Department of Computing

School of Electrical Engineering and Computer Science (SEECS), National

University of Sciences and Technology (NUST), Islamabad

iii

Certiﬁcate of Approval

This is to certify that the research work presented in this thesis, entitled “Electricity Forecast-

ing for Efﬁcient Energy Management in Smart Grid” was conducted by Ms. Aqdas Naz under

the supervision of Dr. Nadeem Javaid. No part of this thesis has been submitted anywhere

else for any other degree. This thesis is submitted to the Department of Computer Science,

COMSATS University Islamabad, Islamabad, in the partial fulﬁllment of the requirement for

the degree of Doctor of Philosophy in the ﬁeld of Computer Science.

Aqdas Naz Signature:

Examinations Committee:

..................................................... .....................................................

External Examiner 1: External Examiner 2:

(Designation and Ofﬁce Address) (Designation and Ofﬁce Address)

..................................................... .....................................................

Dr. Nadeem Javaid Dr. Majid Iqbal Khan

Supervisor, Head of Department,

Department of Computer Science Department of Computer Science

CUI, Islamabad CUI, Islamabad

..................................................... .....................................................

Dr. Ehsan Ullah Munir Prof. Dr. Zulﬁqar Habib

Chairperson, Dean,

Department of Computer Science Faculty of Information Science

CUI and Technology, CUI

iv

Author’s Declaration

I, Aqdas Naz, CIIT/FA17-PCS-003/ISB hereby state that my PhD thesis titled “Electricity

Forecasting for Efﬁcient Energy Management in Smart Grid” is my own work and has not

been submitted previously by me for taking any degree from this University i.e., COMSATS

University Islamabad or anywhere else in the country/world.

At any time if my statement is found to be incorrect even after I graduate the University has

the right to withdraw my PhD degree.

Date: 18-08-2021 Signature:

Aqdas Naz

CIIT/FA17-PCS-003/ISB

v

Plagiarism Undertaking

I solemnly declare that research work presented in this thesis titled, “Electricity Forecasting

for Efﬁcient Energy Management in Smart Grid” is solely my research work with no signiﬁ-

cant contribution from any other person. Small contribution/help wherever taken has been

duly acknowledged and that complete thesis has been written by me.

I understand the zero tolernace policy of HEC and COMSATS University Islamabad towards

plagiarism. Therefore, I as an author of the above titled thesis declare that no portion of my

thesis has been plagiarized and any material used as reference is properly referred/cited.

I undertake if I am found guilty of any formal plagiarism in the above titled thesis even after

award of PhD degree, the University reserves the right to withdraw/revoke my PhD degree

and that HEC and the university has the right to publish my name on the HEC/university

website on which names of students are placed who submitted plagiarized thesis.

Date: 18-09-2021 Signature:

Aqdas Naz

CIIT/FA17-PCS-003/ISB

vi

Certiﬁcate

It is certiﬁed that Aqdas Naz, CIIT/FA17-PCS-003/ISB has carried out all the work related

to this thesis under my supervision at the Department of Computer Science, COMSATS

University Islamabad, Islamabad and the work fulﬁlls the requirement for award of PhD

degree.

Date: 18-09-2021 Supervisor:

Dr. Nadeem Javaid

Associate Professor

Co-Supervisor:

Dr Sohail Iqbal

Associate Professor

Head of Department:

Dr. Majid Iqbal Khan

Associate Professor

Department of Computer Science

vii

DEDICATION

This research work is dedicated

to

My Beloved Parents,

My Supervisor: "Dr. Nadeem Javaid",

My Husband, Children and Family

viii

ABSTRACT

Electricity Forecasting for Efﬁcient Energy Management in Smart

Grid

Smart Grid (SG) is a modernized grid that provides efﬁcient, reliable and economic energy to

the consumers. Energy is the most important resource in the world and almost everything

relies on it. As smart devices are increasing dramatically with the rapid increase in population,

there is a need for an efﬁcient energy distribution mechanism. Furthermore, the forecasting

of electricity consumption is supposed to be a major constituent to enhance the performance

of the SG. Various learning algorithms have been proposed in the literature for efﬁcient load

and price forecasting. However, there exist some issues in the proposed work like increased

computational complexity. The sole purpose of the work done in this thesis is to efﬁciently

predict electricity load and price using different techniques with minimum computational

complexity. Chapter 1 provides an introduction of various concepts present in the power

grids. Afterwards, the uniﬁed system model, different sub-problems and the contributions

made in the thesis are also presented. Chapter 2 discusses the existing work done by different

researchers for performing electricity load and price forecasting. In Chapter 3, Enhanced

Logistic Regression (ELR) and Enhanced Recurrent Extreme Learning Machine (ERELM)

are proposed for performing short-term load and price forecasting. The former is an enhanced

form of Logistic Regression (LR); whereas, the weights and biases of the latter are optimized

using Grey Wolf Optimizer (GWO). Classiﬁcation And Regression Tree (CART), Relief-F

and Recursive Feature Elimination (RFE) are used for feature selection and extraction. On the

basis of selected features, classiﬁcation is performed using ELR. Moreover, cross validation

is done using Monte Carlo and K-Fold methods. In order to ensure optimal and secure

functionality of Micro Grid (MG), Chapter 4 focuses on coordinated energy management of

traditional and Renewable Energy Sources (RES). Users and MG with storage capacity are

taken into account to perform efﬁcient energy management. A two stage Stackelberg game is

formulated. Every player in the game tries to increase its payoff, and ensure user comfort and

system reliability. Furthermore, two forecasting techniques are proposed in order to forecast

Photo-Voltaic Cell (PVC) generation for announcing optimal prices. Both the existence

and uniqueness of Nash Equilibrium (NE) for the energy management algorithm are also

considered. In Chapter 5, a novel forecasting model, termed as ELS-net, is proposed. It is a

x

combination of an Ensemble Empirical Mode Decomposition (EEMD) method, multi-model

Ensemble Bi Long Short Term Memory (EBiLSTM) forecasting technique and Support

Vector Machine (SVM). In the proposed model, EEMD is used to distinguish between linear

and non-linear Intrinsic Mode Functions (IMFs). EBiLSTM is used to forecast the non-linear

IMFs and SVM is employed to forecast the linear IMFs. The usage of separate forecasting

techniques for linear and non-linear IMFs decreases the computational complexity of the

model. In Chapter 6, a novel deep learning model, termed as Gated-FCN, is introduced for

short-term load forecasting. The key idea is to introduce automated feature selection and

a deep learning model for forecasting, which includes an eight layered FCN (FCN-8). It

ensures that hand crafted feature selection is avoided as it requires expert domain knowledge.

Furthermore, Gated-FCN also helps in reducing noise as it learns internal dependencies

as well as the correlation of the time-series. Enhanced Bidirectional Gated Recurrent Unit

(EBiGRU) model is dovetailed with FCN-8 in order to learn temporal long-term dependencies

of the time-series. Furthermore, weight averaging mechanism of multiple snapshot models is

adapted in order to take optimized weights of BiGRU. At the end of FCN-8 and BiGRU, a

fully connected dense layer is used that gives ﬁnal prediction results. The simulations are

performed and the results are provided at the end of each chapter. In Chapter 3, the simulations

are performed using UMass electric and UCI datasets. ELR shows better performance with

the former dataset; whereas, ERELM has better accuracy with the latter. The proposed

techniques are then compared with different benchmark schemes. The comparison is done

to verify the adaptivity of the proposed techniques. The simulation results show that the

proposed techniques outperform the benchmark schemes and increase the prediction accuracy

of electricity load and price. Similarly, in Chapter 4, simulations are performed using Elia,

Belgium dataset. The results clearly show that the proposed game theoretic approach along

with storage capacity optimization and forecasting techniques give beneﬁts to both users and

MG. In Chapter 5, simulations are performed to examine the effectiveness of the proposed

model using two different datasets: New South Wales (NSW) and Victoria (VIC). From the

simulation results, it is obvious that the proposed ELS-net model outperforms the benchmark

techniques: EMD-BILSTM-SVM, EMD-PSO-GA-SVR, BiLSTM, MLP and SVM in terms

of forecasting accuracy and minimum execution time. Similarly, the simulation results

of Chapter 6 depict that Gated-FCN gives maximum forecasting accuracy as compared to

xi

the benchmark techniques. For performance evaluation of the proposed work, different

performance metrics are used: Mean Absolute Percentage Error (MAPE), Mean Absolute

Error (MAE), Mean Squared Error (MSE) and Root Mean Square Error (RMSE). The overall

results prove that the work done in this thesis outperforms the existing work in terms of

electricity load and price forecasting, and computational complexity.

xii

Journal Publications

4

Aqdas Naz, Nadeem Javaid, Muhammad Asif, Muhammad Umar Javed, Abrar Ahmed,

Sardar Muhammad Gulfam, Muhammad Shaﬁq and Jin-Ghoo Choi. (2021). Electricity

Consumption Forecasting using Gated-FCN with Ensemble Strategy, IEEE Access,

Volume NN, Pages NN. ISSN: 2169-3536. [IF=3.367]. Download.

3

Nadeem Javaid, Aqdas Naz, Rabiya Khalid, Ahmad Almogren, Muhammad Shaﬁq and

Adia Khalid. (2020). ELS-net: A New Approach to Forecast Decomposed Intrinsic

Mode Functions of Electricity Load. IEEE Access, 8, 198935-198949. [IF= 3.367].

Download.

2

Aqdas Naz, Nadeem Javaid, Muhammad Babar Rasheed, Abdul Haseeb, Musaed

Alhussein and Khursheed Aurangzeb. (2019). Game Theoretical Energy Manage-

ment with Storage Capacity Optimization and Photo-Voltaic Cell Generated Power

Forecasting in Micro Grid. Sustainability, 1-22. [IF=3.251]. Download.

1

Aqdas Naz, Muhammad Umar Javed, Nadeem Javaid, Tanzila Saba, Musaed Alhussein

and Khursheed Aurangzeb. (2019). Short-Term Electric Load and Price Forecasting

Using Enhanced Extreme Learning Machine Optimization in Smart Grids, Energies,

1-30. [IF=3.004]. Download.

xiii

Conference Proceedings/Book Chapters

8

Aqdas Naz, Nadeem Javaid, Adia Khalid, Muhammad Shoaib, Muhammad Imran.

Electric Load Forecasting using EEMD and Machine Learning Techniques, in the 16th

International Wireless Communications and Mobile Computing Conference (IWCMC),

2020. Download.

7

Aqdas Naz, Nadeem Javaid, Abdul Basit Majeed Khan, Muhammad Mudassar Iqbal,

Muhammad Aqeel ur Rehman Hashmi and Raheel Ahmad Abbasi. Game-Theoretical

energy management for residential user and micro grid for optimum sizing of photo

voltaic battery systems and energy prices, in Workshops of the 33rd International

Conference on Advanced Information Networking and Applications (AINA), 2019.

Download

6

Aqdas Naz, Nadeem Javaid and Sakeena Javaid. Enhanced Recurrent Extreme Learning

Machine using Gray Wolf Optimization for Load Forecasting, in the 21st International

Multitopic Conference (INMIC), 2018. Download.

5

Talha Naeem Qureshi, Nadeem Javaid, Aqdas Naz, Waseem Ahmad, Muhammad Imran

and Zahoor Ali Khan. A Noval Meta-Heuristic Hybrid Enhanced Differential Harmony

Wind Driven (EDHWDO) Optimization Technique For Demand Side Management in

Smart Grid, in 32nd International Conference on Advanced Information Networking

and Applications Workshops (WAINA), 2018. Download.

4

Aqdas Naz, Nadeem Javaid, Talha Naeem Qureshi, Muhammad Imran, Mujahid Ali

and Zahoor Ali Khan. EDHBPSO: Enhanced Differential Harmony Binary Particle

Swarm Optimization, in 32nd International Conference on Advanced Information

Networking and Applications Workshops (WAINA) 2018. Download.

3

Isra Shaﬁ, Nadeem Javaid, Aqdas Naz, Yasir Amir, Israr Ishaq and Kashif Naseem.

Feature Selection and Extraction along with Electricity Price forecasting using Big

Data Analytics, in 12th International Conference on Innovative Mobile and Internet

Services in Ubiquitous Computing (IMIS), 2018. Download.

2

Aqdas Naz, Nadeem Javaid, Muhammad Mudassar Iqbal, Mujahid Ali, Muhammad Im-

ran and Zahoor Ali Khan. TBEENISH: Threshold Balanced Energy Efﬁcient Network

Integrated Super Heterogeneous Protocol for WSNs, in 32nd International Conference

xiv

on Advanced Information Networking and Applications Workshops (WAINA), 2018.

Download.

1

Muqaddas Naz, Nadeem Javaid, Urva Latif, Talha Naeem Qureshi, Aqdas Naz and

Zahoor Ali Khan. Efﬁcient Power Scheduling in Smart Homes using Meta Heuristic

Hybrid Grey Wolf Differential Evolution Optimization Technique, in 32nd International

Conference on Advanced Information Networking and Applications (AINA), 2018.

Download.

xv

TABLE OF CONTENTS

1 Introduction ..................................................................... 1

1.1 Problem Statement ............................................................... 4

1.1.1 Sub-problem 1: Short-term Electric Load and Price Forecast-

ing using Enhanced Extreme Learning Machine Optimization

in Smart Grids ............................................................ 4

1.1.2 Sub-problem 2: Game theoretical Energy Management with

Storage Capacity Optimization and Photo-Voltaic Cell Gener-

ated Power Forecasting in Micro Grids ............................... 4

1.1.3 Sub-problem 3: ELS-net: A New Approach to Forecast Decom-

posed Intrinsic Mode Functions of Electricity Load ................ 5

1.1.4 Sub-problem 4: Electricity Forecasting using Gated-FCN with

Ensemble Strategy ....................................................... 5

1.2 Contributions ...................................................................... 6

1.3 Proposed System Model ......................................................... 7

1.4 Thesis Organization .............................................................. 9

2 Related Work .................................................................... 12

3 Short-term Electric Load and Price Forecasting using Enhanced Ex-

treme Learning Machine Optimization in Smart Grids .................... 26

3.1 Summary of the Chapter......................................................... 27

3.2 Proposed Methodology .......................................................... 27

3.2.1 Classiﬁcation And Regression Technique ............................ 28

3.2.2 Recursive Feature Elimination ......................................... 28

3.2.3 Relief-F.................................................................... 29

3.2.4 Convolutional Neural Network ........................................ 30

3.2.5 Logistic Regression...................................................... 30

3.2.6 Enhanced Logistic Regression ......................................... 31

3.2.7 Grey Wolf Optimizer .................................................... 31

xvi

3.2.8 Recurrent Extreme Learning Machine ................................ 33

3.2.9 Enhanced Recurrent Extreme Learning Machine ................... 33

3.3 Proposed System Models........................................................ 34

3.3.1 Proposed System Model 1 .............................................. 34

3.3.2 Proposed System Model 2 .............................................. 35

3.4 Simulation Results and Discussion ............................................ 36

3.4.1 Simulation Results and Discussion of Proposed System Model 1 36

3.4.1.1 Data Description .............................................. 36

3.4.1.2 CART........................................................... 37

3.4.1.3 RFE ............................................................. 37

3.4.1.4 Relief-F ........................................................ 38

3.4.1.5 Load Forecasting.............................................. 38

3.4.1.6 Price Forecasting.............................................. 38

3.4.2 Simulation Results and Discussion of Proposed System Model 2 39

3.4.2.1 Data Description .............................................. 39

3.4.2.2 Results’ Discussion........................................... 42

3.5 Performance Metrics ............................................................. 44

3.6 Conclusion of the Chapter....................................................... 50

4 Game Theoretical Energy Management with Storage Capacity Opti-

mization and Photo-Voltaic Cell Generated Power Forecasting in Micro

Grid .............................................................................. 52

4.1 Summary of the Chapter......................................................... 53

4.2 Proposed and Existing Methodology .......................................... 53

4.2.1 Cost Model of Users..................................................... 54

4.2.2 MG Cost Model with Storage Capacity .............................. 56

4.2.3 Game Formulation and Analysis ...................................... 60

4.2.3.1 PVC Power Forecasting Algorithm ........................ 63

4.3 Simulation Results and Discussion ............................................ 66

4.3.1 Dataset description ...................................................... 67

4.3.2 Simulation Results ....................................................... 67

4.4 Conclusion of the Chapter....................................................... 72

xvii

5 ELS-net: A New Approach to Forecast Decomposed Intrinsic Mode

Functions of Electricity Load .................................................. 73

5.1 Summary of the Chapter......................................................... 74

5.2 Problem Formulation ............................................................ 74

5.3 Proposed System Model ......................................................... 75

5.3.0.1 Ensemble Empirical Mode Decomposition ............... 76

5.3.1 The Proposed Multi-model EBiLSTM................................ 78

5.3.1.1 Bi Long Short Term Memory ............................... 78

5.3.1.2 The Proposed Multi-model Ensemble Strategy........... 80

5.3.1.3 Support Vector Machine ..................................... 81

5.4 Simulation Results and Discussion ............................................ 81

5.4.1 Data Description ......................................................... 81

5.4.2 Performance Evaluators................................................. 82

5.4.3 Predictive Analytics of Electricity Consumption Data ............. 82

5.4.4 Computational Cost ..................................................... 90

5.4.5 Scalability Analysis ..................................................... 90

5.5 Conclusion of the Chapter....................................................... 91

6 Electricity Forecasting using Gated-FCN with Ensemble Strategy ....... 93

6.1 Summary of the Chapter......................................................... 94

6.2 Proposed Deep Learning Model................................................ 94

6.2.1 Fully Convolutional Network-8........................................ 95

6.2.1.1 Convolution Layer ............................................ 95

6.2.1.2 Pooling Layer ................................................. 97

6.2.2 Enhanced Bidirectional Gated Recurrent Unit ....................... 97

6.2.3 Cascaded FCN-8 and BiGRU .......................................... 99

6.3 Simulation results and discussion ..............................................101

6.3.1 Data Description .........................................................102

6.3.2 Validation Setting of the Proposed System Model ..................103

6.3.3 Exploratory Data Analysis..............................................103

6.3.4 Predictive Data Analysis ................................................104

xviii

I title 108

6.4 Conclusion of the Chapter.......................................................115

7 Conclusion and Future Work ..................................................119

7.1 Conclusions .......................................................................120

7.2 Future Recommendations .......................................................122

xix

LIST OF FIGURES

1.1 Uniﬁed system model .............................................................. 8

1.2 Steps taken during load and price forecasting .................................... 9

3.1 Grey wolf social hierarchy......................................................... 32

3.2 Functioning of RELM ............................................................. 33

3.3 Proposed system model 1 .......................................................... 35

3.4 Proposed system model 2 .......................................................... 35

3.5 One day load prediction using UMass dataset .................................... 39

3.6 One day load prediction using UCI dataset ....................................... 39

3.7 One week load prediction using UMass dataset .................................. 40

3.8 One week load prediction using UCI dataset ..................................... 40

3.9 One month load prediction using UMass dataset ................................. 40

3.10 One month load prediction using UCI dataset .................................... 41

3.11 One day price prediction using UMass dataset ................................... 41

3.12 One week price prediction using UMass dataset ................................. 41

3.13 One month price prediction using UMass dataset ................................ 42

3.14 Daily load consumption taken from 10 different smart meters .................. 44

3.15 ELM regression line plot for UMass dataset...................................... 46

3.16 ELM regression line plot for UCI dataset......................................... 46

3.17 RELM regression line plot for UMass dataset .................................... 46

3.18 RELM regression line plot for UCI dataset ....................................... 47

3.19 ERELM regression line plot for UMass dataset .................................. 47

3.20 ERELM regression line plot for UCI dataset ..................................... 47

4.1 Interaction between user and MG ................................................. 54

4.2 Flow diagram of CS algorithm .................................................... 65

4.3 Flow diagram of GWO algorithm ................................................. 66

4.4 PVC power generation in a certain day ........................................... 67

4.5 Hourly PVC generation ............................................................ 68

xx

4.6 Daily PVC generation.............................................................. 69

4.7 Consumption of users in 24 hours................................................. 69

4.8 Energy distribution without game theory ......................................... 69

4.9 Energy distribution with game theory ............................................. 70

4.10 PVC Energy distribution without game theory ................................... 70

4.11 PVC power generation with game theory ......................................... 71

4.12 MAPE of ﬁve different models with PVC power forecasting steps ............. 71

5.1 Input and output of deep learning model.......................................... 75

5.2 Proposed system model ............................................................ 76

5.3 Flow diagram of EBiLSTM ....................................................... 79

5.4 Ensemble strategy in BiLSTM .................................................... 80

5.5 Original NSW dataset .............................................................. 82

5.6 Original VIC dataset ............................................................... 82

5.7 Extracted IMFs - NSW dataset .................................................... 83

5.8 Extracted IMFS - VIC dataset ..................................................... 83

5.9 RMSE for NSW .................................................................... 84

5.10 MAE for NSW ..................................................................... 84

5.11 MAPE for NSW .................................................................... 84

5.12 Prediction results of electricity consumption for Summer season using NSW.. 85

5.13 Prediction results of electricity consumption for Autumn season using NSW .. 85

5.14 Prediction results of electricity consumption for Winter season using NSW . . . 85

5.15 Prediction results of electricity consumption for Spring season using NSW.. .. 86

5.16 RMSE for VIC ..................................................................... 86

5.17 MAE for VIC....................................................................... 87

5.18 MAPE for VIC ..................................................................... 87

5.19 Prediction results of electricity consumption for Summer season using VIC ... 87

5.20 Prediction results of electricity consumption for Winter season using VIC. ... . 88

5.21 Prediction results of electricity consumption for Spring season using VIC ... .. 88

5.22 Prediction results of electricity consumption for Autumn season using VIC .. . 88

6.1 Gated-FCN architecture............................................................ 95

6.2 Activation functions ............................................................... 96

xxi

6.3 Max pooling and average pooling ................................................. 98

6.4 BiGRU architecture ................................................................ 99

6.5

Ensemble strategy illustration - The number of snapshot models are selected

on the basis of minimum loss...................................................... 99

6.6 2011-2019 ISO-NE hourly load consumption ....................................104

6.7 Weekend and Weekday load consumption ........................................104

6.8 Load forecasting using MAPE ....................................................106

6.9 Daily load consumption forecasting...............................................109

6.10 Weekly load consumption forecasting ............................................110

6.11 Load consumption - Monday ......................................................110

6.12 Load consumption - Tuesday ......................................................110

6.13 Load consumption - Wednesday................................................... 111

6.14 Load consumption - Thursday .....................................................111

6.15 Load consumption - Friday ........................................................111

6.16 Load consumption - Saturday .....................................................112

6.17 Load consumption - Sunday .......................................................112

6.18 Loss of CNN-LSTM ...............................................................112

6.19 Loss of Gated-FCN ................................................................113

xxii

LIST OF TABLES

1.1 Comparison of different levels of the proposed system model................... 10

2.1 Summary of related work .......................................................... 18

3.1 Features in UMass electric dataset ................................................ 37

3.2 Results of CART for UMass electric dataset ..................................... 37

3.3 RFE features for UMass electric dataset .......................................... 38

3.4 Relief-F features for UMass electric dataset ...................................... 38

3.5 Obtained RMSE using ELM, RELM and ERELM............................... 44

3.6

Obtained RMSE for half-yearly testing data using ELM, RELM and ERELM

by Monte Carlo and K-Fold cross validation...................................... 45

3.7

Obtained RMSE for yearly testing data using ELM, RELM and ERELM by

Monte Carlo and K-Fold cross validation. ........................................ 45

3.8 Computational time comparison of ERELM, RELM and ELM execution. .. .. . 45

3.9

Load performance metrics comparison for one day using the UMass electric

dataset............................................................................... 48

3.10

Load performance metrics comparison for one week using the UMass electric

dataset............................................................................... 48

3.11

Load performance metrics comparison for one month using the UMass elec-

tric dataset .......................................................................... 48

3.12

Price performance metrics comparison for one day using the UMass electric

dataset............................................................................... 49

3.13

Price performance metrics comparison for one week using the UMass electric

dataset............................................................................... 49

3.14

Price performance metrics comparison for one month using the UMass elec-

tric dataset .......................................................................... 49

3.15 Load performance metrics comparison for one day using the UCI dataset . .. .. 49

3.16 Load performance metrics comparison for one week using the UCI dataset. .. . 50

3.17 Load performance metrics comparison for one month using the UCI dataset .. 50

xxiii

3.18 Accuracy of ERELM using RMSE, MSE and MAE for half-yearly data ....... 50

3.19 Accuracy of ERELM using RMSE, MSE and MAE for yearly data ............ 51

4.1 Pricing model parameters.......................................................... 68

4.2 PVC generation storage battery parameters....................................... 68

5.1 Performance results ................................................................ 89

5.2

Obtained MAPE for NSW and VIC testing datasets using multistep against

ELS-net, EMD-BiLSTM-SVM and BiLSTM by K-Fold cross validation. ..... 89

5.3 Computational time (sec) ......................................................... 90

6.1 Validation setting of FCN model ..................................................103

6.2 Validation setting of BiGRU model ..............................................103

6.3 Exploring activation functions.....................................................105

6.4 Performance results - Monday.....................................................114

6.5 Performance results - Tuesday.....................................................114

6.6 Performance results - Wednesday .................................................115

6.7 Performance results - Thursday ...................................................115

6.8 Performance results - Friday.......................................................116

6.9 Performance results - Saturday ....................................................116

6.10 Performance results - Sunday......................................................117

6.11

Comparison of Gated-FCN with existing models using noise added consumption

117

6.12 Computational time (sec) ..........................................................118

6.13 Accuracy of Forecasting Techniques ..............................................118

xxiv

LIST OF ABBREVIATIONS

ACF Auto Correlation Function

ADF Augmented Dickey Fuller

AEMO Australian Electricity Market Operator

AI Artiﬁcial Intelligence

AIC Akaike Information Criteria

ANN Artiﬁcial Neural Network

AR Auto regressive

ARIMA Auto Regressive Integrated Moving Average

ARMAX Auto Regressive Moving Average with Exogenous variables

BIC Bayesian Information Criteria

BiGRU Bidirectional Gated Recurrent Unit

BiLSTM Bidirectional Long Short Term Memory

BP Back Propagation

CARIMA Cuckoo Search Optimized ARIMA

CART Classiﬁcation And Regression Technique

CNN Convolutional Neural Network

CS Cuckoo Search

DAE Deep Auto Encoders

DBN Deep Belief Net

DEM Distributed Energy Management

DE-SVM Differential Evolution Support Vector Machine

DE-ELM Differential Evolution Extreme Learning Machine

DLSTM Dilated LSTM

DNN Deep Neural Network

DR Demand Response

DRN Deep Residual Network

DSM Demand Side Management

DT Decision Tree

DWT Discrete Wavelet Transform

EBiLSTM Ensemble Bi Long Short Term Memory

xxv

EEMD Ensemble Empirical Mode Decomposition

EEMD-DBN Ensemble Empirical Mode Decomposition Deep Belief Network

EPEX European Power Exchange

ELR Enhanced Logistic Regression

ELM Extreme Learning Machine

ELS-net EEMD BiLSTM SVM Network

EMD Empirical Mode Decomposition

ERELM Enhanced Recurrent Extreme Learning Machine

ES Exponential Smoothing

ESAE Enhanced Stacked Auto Encoder

EUNITE EUropean Network of excellence on Intelligent TEchnologies for smart

adaptive systems

FCN-8 Fully Convolutional Network

FFNN Feed Forward Neural Network

Gated FCN Gated Fully Convolutional Network

GA Genetic Algorithm

GANN Generalized Artiﬁcal Neural Network

GARIMA Grey Wolf Optimized ARIMA

GASVM Genetic Algorithm Support Vector Machine

GCA Grey Correlation Analysis

GRU Gated Recurrent Unit

GSAE Genetic Stacked Auto encoder

GWO Grey Wolf Optimization

HHT Hilbert-Huang Transform

ICT Information and Communication Technology

IMF Intrinsic Mode Function

ISONE Independent System Operators New England

ISO NECA Independent System Operator New England Control Area

KELM Kernel Extreme Learning Machine

KPCA Kernel Principal Component Analysis

LR Logistic Regression

xxvi

LSTM Long Short Term Memory

MAE Mean Absolute Error

MAPE Mean Absolute Percentage Error

MG Micro Grid

MISO Midcontinent Independent System Operator

MLP Multi Layer Perceptron

MLR Multi Linear Regression

MSE Mean Square Error

NLS-SVM Nonlinear Least Square Support Vector Machine

NE Nash Equilibrium

NN Neural Network

NSW New South Wales

NYISO New York Independent System Operator

OS-ELM Online Sequential Extreme Learning Machine

PACF Partial Auto Correlation Function

PAR Peak Average Ratio

PJM Pennsylvania–New Jersey–Maryland Interconnection

PSO Particle Swarm Optimization

PVC Photo-Voltaic Cell

QLD Queensland

RBM Restricted Boltzmann Machine

RELM Recurrent Extreme Learning Machne

ReLU Rectiﬁed Linear Unit

RES Renewable Energy Sources

RFE Recursive Feature Elimination

RMSE Root Mean Square Error

RNN Recurrent Neural Network

RTP Real Time Pricing

SARIMA Seasonal Auto Regressive Integrated Moving Average

SBELM Sparse Bayesian Extreme Learning Machine

SDA Stacked De-noising Autoencoders

xxvii

SG Smart Grid

SLFN Single Layer Feedforward Network

SM Smart Meters

SMI Smart Meter Infrastructure

SVM Support Vector Machine

SVR Support Vector Regression

TG Traditional Grid

ToU Time of Use

TVC-ABC Time Varying Coefﬁcients Artiﬁcial Bee Colony

VIC Victoria

xxviii

NOMENCLATURE

AvActual value

AvActual time-series

ak,bkFixed parameters

bl(k)Total energy required to discharge battery

CkxkCost function of user

Cbat

m(y)Daily depreciation cost function

Csp

nTotal cost of solar power generation

Cn,kTotal cost of Nusers

enEnergy demanded from users

eb

m(k)Energy required to charge battery

FPenalty factor

FvPredicted time-series

fmList of features

ftForget gate

h Step size

hk

ch Pattern of battery charging

hk

disch Pattern of battery discharging

htHidden layer

itInput gate

KTime slots in a period

lnDaily Energy usage of user

m Mean

MNumber of ensembles

mNumber of snapshot models

N Total number of samples

NMultiple residential users

nNumber of hidden nodes

nt(i)White noise

otOutput gate

PkReal time price of each slot

xxix

P

vPredicted time-series

Q∗Micro Grid

Q∗Test static

ri,sFinal residue

sState of battery

TTotal number of observations

T Total time duration

t Time slot

UnProﬁt of MG

Xaverage Average user consumption

XkSum of total energy

Xpeak Peak consumption

x Actual value

xn(k)Energy demand in a particular slot

x’ Predicted value

y Output

ym

lLower limit of battery capacity

yu

mUpper limit of battery capacity

αFittest wolf 1

βFittest wolf 2

δFittest wolf 3

ωRemaining wolves

ηch Charging efﬁciency of a battery

ηdisch Discharging efﬁciency of a battery

σ2

εConstant variance

UCost function of users

BCost function of a MG

εtError at time t

σ2Mean square error

δPrediction error

ςStrategy form

xxx

λsSolar power selling price

φiVector based on rules of AR

θiVector based on rules of MA

xxxi

Chapter 1

Introduction

1

Conventionally, Traditional Grids (TGs) are being used for electricity generation and distribu-

tion. As the time passes, the infrastructure of TG is getting obsolete, which results in energy

loss and less efﬁcient output. Due to the usage of outdated infrastructure, intensive power loss

is being faced, which further leads to various issues like load shedding, blackouts, voltage

ﬂuctuations, etc., [

1

]-[

5

]. TGs use fossil fuels like coal, petrol, diesel, etc., for the combustion

process of turbines. The extensive burning of fossil fuels leads to natural resource depletion

and increase in pollution. To tackle the aforementioned issues, authors and researchers have

suggested to use Renewable Energy Sources (RES) and modify the existing TGs by incorpo-

rating the latest technologies and modern infrastructure. The new and modiﬁed form of TG is

the Smart Grid (SG) [

6

]-[

12

], which involves bi-directional communication between users

and the utility [

13

]-[

17

]. It also includes forecasting of electricity load and price. The accurate

and efﬁcient forecasting helps the power companies to perform enhanced grid construction

planning. It assists in the improvement of economic and social beneﬁts of the power systems.

Moreover, forecasting is also done to schedule the load consumption from on-peak hours to

off-peak hours for the next day, week or month to reduce the electricity cost and enhance the

user comfort [

18

]-[

25

]. Furthermore, the accurately forecasted data is also very effective in

predicting the potential faults of the system. Thus, security of the power grids can be ensured

using predicted data. Load forecasting also helps to ensure the balance between supply and

demand of electricity. Consequently, energy generation can be managed without wasting or

generating surplus energy.

A detailed analysis is performed on the time-series data, which aims to extract meaningful

statistics and other important features from it [

26

]. The time-series data is captured and

analyzed in two main forms: uni-variate and multi-variate [

27

]. Time-series data has cyclical

or seasonal characteristics that are inﬂuential in the context of its analysis. In order to perform

analysis on cyclic characteristics, different models have been proposed in the literature, e.g., a

pattern recognition framework is proposed to locate the changes in cyclic time-series, which

combines linear discriminant as well as K-means clustering method [

28

]. In time-series load

forecasting paradigm, there exist four types of forecasting horizons: long-term, medium-term,

short-term and very short-term [

29

]. Forecasting of electricity consumption is a challenging

task. Climate changes and social activities are considered as external factors that highly affect

the linearity and predictability of the data [

30

]. Several statistical linear methods are used for

time-series forecasting. The goal of these methods is to extrapolate the energy requirements

of future. Besides, the most successful techniques are based on Holt winters exponential

smoothing [

31

] as well as Auto Regressive Integrated Moving Average (ARIMA) [

32

]-[

35

].

2

In the past few years, rapid development in the ﬁeld of computational intelligence, which in-

cludes Artiﬁcial Neural Network (ANN) [

36

]-[

39

], Support Vector Machine (SVM) [

40

]-[

43

]

and fuzzy comprehensive evaluation [

44

], has been observed. ANN is extensively used in

literature for regression and classiﬁcation purposes. However, it has the limitation that it often

gets trapped in local optima

[45]

. Apart from the convergence issue, short-term load forecast-

ing is also a challenge due to its non-stationary load and long-term dependency forecasting

horizon [

46

]-[

50

]. Thus, deep forecasting techniques are given due importance these days

[

51

]. These techniques use multiple stacked hidden layers to achieve the level of abstraction

and give better learning ability. Moreover, the extension of Recurrent Neural Network (RNN),

i.e., Long Short Term Memory (LSTM), is gaining attention in the ﬁeld of machine learning

due to its dependency learning ability and memory storage [

52

]-[

56

]. However, LSTM only

learns the past dependencies in the training data. On the other hand, Bidirectional LSTM

(BiLSTM) learns both past and future dependencies. Deep learning model varies from data

to data; therefore, it is required to increase the generalization capabilities of the model.

Ensemble learning methods are proven to have better forecasting capabilities. These methods

strategically combine multiple forecasting algorithms and perform regression and classiﬁca-

tion on uni-variate and multi-variate data. There are three basic reasons behind the success

of ensemble methods: statistical, computational and representational [

57

]. Many different

ensemble methods are in practice nowadays like ensemble of Deep learning Belief Networks

(DBNs) and ensemble of Extreme Learning Machines (ELMs) [

58

,

59

]. Among different

ensemble methods, divide and conquer rule performs the best using time-series data [

60

].

Forecasting techniques are also combined with the time-series decomposition techniques in

the literature. Empirical Mode Decomposition (EMD) is one of the decomposition methods

that is used for time-series decomposition. It is a part of the Hilbert-Huang Transform (HHT)

method. With the latest technological advancements made in the power sector, efﬁcient

load forecasting has become a necessity. Keeping this in mind, different electricity load

forecasting methods have been developed over the past few years. Starting from conventional

neural network techniques like ANN, CNN, etc.,

[61]-[66]

, the focus has been shifted to-

wards hybrid and other meta heuristic techniques like CNN-LSTM, EMD, etc.,

[67]-[72]

.

Conventional techniques perform well up to a certain limit. However, these techniques have

some major limitations like overﬁtting, premature convergence, mode mixing, etc. Due to

these limitations, no technique is able to completely solve the trade-off between forecasting

accuracy and computational complexity. Therefore, the work provided in the thesis is carried

out to accurately forecast electricity consumption and proceed with minimum computational

3

complexity. Simulation results given at the end of each chapter prove the efﬁcacy of the work

done in the thesis.

1.1 Problem Statement

This section provides the problem statement of the thesis in the form of multiple sub-problems.

1.1.1 Sub-problem 1: Short-term Electric Load and Price Forecasting using

Enhanced Extreme Learning Machine Optimization in Smart Grids

The data of SGs is increasing drastically with the increase in population, so an efﬁcient tech-

nique is required to predict electricity load and price [

73

]-[

77

]. Authors in [

78

] use Recurrent

Extreme Learning Machine (RELM) to predict the electricity load. However, in RELM,

weights and biases are randomly assigned, which leads to drastic variations in prediction

results. An enhanced technique is proposed to solve the aforesaid issue. In [

79

], authors use

Convolutional Neural Network (CNN) for predicting the energy demand. However, CNN

involves tuning of a number of layers, which makes it spatio-temporal complex. In this thesis,

two enhanced techniques are proposed to increase the forecasting accuracy of electricity load

and price. Uni-variate and multi-variate datasets are used for both techniques. Furthermore,

analysis of both residential and utility data is performed collectively.

1.1.2 Sub-problem 2: Game theoretical Energy Management with Storage Ca-

pacity Optimization and Photo-Voltaic Cell Generated Power Forecast-

ing in Micro Grids

To make efﬁcient use of renewable resources, different techniques are proposed in the past.

However, Distributed Energy Management (DEM) problem is not tackled well. Moreover, the

authors try to achieve the objective function without simultaneously dealing with user comfort

and utility proﬁt. On the contrary, the work done in this thesis aims to maximize the objective

function of each player while satisfying users’ demand of electricity and guaranteeing reliable

energy system operations. Due to uncontrollable and uncertain nature of energy generation

by RES [

80

]-[

89

], different techniques are used for short-term electricity load forecasting.

However, the conventional techniques do not perform efﬁcient load forecasting. In addition,

the problem of energy management in a distributed Micro Grid (MG) also needs to be tackled

[

90

]-[

94

]. The problem is addressed in the proposed work using non-cooperative game theory

as it ignores common commitment of players and has minimum communication overhead

4

[

95

]. The optimization of storage capacity in MG plays a vital role in context of efﬁcient

management of electric load [96], which is also addressed in this thesis.

1.1.3 Sub-problem 3: ELS-net: A New Approach to Forecast Decomposed In-

trinsic Mode Functions of Electricity Load

With the advancements made in the technological sector and the drastic increase in population

over the past few decades, the electricity demand has also increased tremendously. However,

electricity generation has not increased with the same ratio, which leads to an imbalance

between electricity demand and supply. This imbalance further leads to many other issues like

load shedding, blackouts, increased electricity price, burden on power grids, etc. Electricity

theft detection is also one of the major issues, which occurs due to expensive electricity

prices [

104

]-[

108

]. To mitigate these issues, accurate load forecasting and Demand Side

Management (DSM) have become the need of the hour [

109

]-[

118

]. Keeping this in view,

researchers have proposed various methods for accurately forecasting electricity demand

like ANN

[99]

, LSTM

[119]

, combination of Ensemble Empirical Mode Decomposition

(EEMD) and Support Vector Regression (SVR)

[120]

, etc. ANN suffers from overﬁtting

and premature convergence problems. It does not perform well for time-series data. The

authors in

[119]

use a variant of RNN, i.e., LSTM, to forecast electricity consumption using

time-series data. Observation of the electricity demand at equal time intervals is termed as

electricity data, also known as the time-series data. There exist two types of time-series

data: linear and non-linear. Using same forecasting method for both types of data series

increases the computational complexity. The authors in

[120]

use EEMD-SVR to decompose

time-series into sub-series. SVR is used to forecast electricity demand. However, it does not

perform well with non-linear data. To overcome the aforementioned limitations, a hybrid

model is proposed in the underlying work with emphasis on denoising of time-series data.

Moreover, the electricity demand time-series data is decomposed into two sub-series: linear

and non-linear. The non-linear data is forecasted using Ensemble Bi Long Short Term

Memory (EBiLSTM); whereas, the linear data is forecasted using SVM.

1.1.4 Sub-problem 4: Electricity Forecasting using Gated-FCN with Ensemble

Strategy

Many variants of Deep Artiﬁcial Neural Network (DANN) have been proposed in literature

for electricity load forecasting [

122

]-[

125

], such as Restricted Boltzmann Machine (RBM)

[

121

] and ANN with multiple layers [

126

]. However, an increase in the number of layers

5

increases the complexity of the network in terms of execution time and memory requirement.

Furthermore, forecasting of residential electricity load is performed using ANN in [

127

]

where the effectiveness of ANN increases as per the availability of exogenous variables such

as weather data, temperature, wind speed and humidity. The non-availability of dependent

variables adversely affects the proposed model’s performance. Besides, long-term dependen-

cies are hard to learn by traditional RNN as it faces vanishing gradient issue due to time depth

and single hidden layer. To tackle this issue, DANN is used in [

128

], which uses variables’

data to detect latent features. The extracted latent features are fed to Multi Layer Perceptron

(MLP) for forecasting the time-series data. However, MLP performance is very poor for

detecting temporal dependencies. To deal with temporal dependencies, LSTM is used. In

[

129

], authors use LSTM as it is efﬁcient in learning temporal trends in time-series data.

However, the correlation of exogenous and dependent variables plays a vital role in learning

the time-series data, which is not considered in the proposed work. Moreover, LSTM also

requires excessive memory during its execution.

1.2 Contributions

This section provides the contributions made in this thesis.

•

Data pre-processing is performed after acquiring data from different datasets like

UMass Electric, UCI, Elia, VIC, NSW, ISO-NE, etc. Data pre-processing includes

different steps like data normalization, outliers removal, linear interpolation, etc.

•

After data pre-processing, feature engineering is performed using different techniques

like Recursive Feature Elimination (RFE), Classiﬁcation And Regression Technique

(CART), Relief-F, Fully Convolutional Network (FCN), etc. Also, the hand crafted

feature extraction process is replaced with an automated deep learning model, which

automatically selects the features.

•

Once, the pre-dominant and less redundant features are obtained, electricity load and

price forecasting is performed using the newly proposed techniques like Enhanced

Logistic Regression (ELR), Enhanced Recurrent Extreme Learning Machine (ERELM),

GARIMA, CARIMA, etc. The electricity forecasting is performed both at the SG and

the MG levels.

•

To increase the forecasting accuracy, game theory is applied in the form of a two

stage Stackelberg game, which ensures Nash Equilibrium (NE). Moreover, for reducing

6

computational complexity, the time-series data is decomposed into linear and non-linear

sub-series, which are forecasted individually.

•

Simulations are performed to prove the efﬁcacy of the solutions proposed in this thesis

and different performance metrics are used for evaluation. The results clearly show

that the work done in this thesis has the tendency to outperform similar existing works.

1.3 Proposed System Model

In this section, the uniﬁed system model of the thesis is presented along with the comparison

of the solutions proposed for the problems identiﬁed in Section 1.1. Figure 1.1 shows

the uniﬁed system model of the thesis. Whereas, Figure 1.2 shows the major steps taken

while performing load and price forecasting: data acquisition, data pre-processing, feature

engineering (feature selection and extraction) and classiﬁcation (forecasting). Moreover,

Table 1.1 gives the comparison of the solutions proposed for sub-problems 1.1.1–1.1.4 and

provided in Chapters 3–6. Each solution is considered as an individual level of the uniﬁed

proposed system model and labeled as L.1–L.4. Level 1 deals with the short-term load

and price forecasting. At this level, the electricity is generated using fossil fuels at the

SG side. For feature engineering, three different methods are used: RFE, Relief-F and

CART. After feature engineering, forecasting is done through two newly proposed techniques:

ELR and ERELM. Both the techniques perform efﬁciently and outperform the benchmark

techniques: LR, ELM and RELM. However, electricity generation at the SG side causes

increase in carbon emission, electricity cost and power losses. To tackle the issues, the

focus is shifted towards the MG, where electricity is generated using RES. Level 2 of the

proposed model deals with the forecasting of power generated through PVC and its storage

optimization. Efﬁcient energy management is also performed at this level. A Stackelberg

game is formulated between the users and the MG, which ensures NE. Two techniques are

proposed to perform forecasting: GARIMA and CARIMA. No doubt, the issues faced in the

SGs are tackled at this level. Still, forecasting load and price using continuous time-series

data is a time-consuming task. Therefore, there is a need to devise such a mechanism that

will deal with this issue. At level 3 of the proposed system model, the time-series data is

decomposed into linear and non-linear sub-series using EEMD, which helps in forecasting

the electricity load in minimum time. Electricity load forecasting using linear sub-series is

performed using SVM; whereas, EBiLSTM is used for forecasting non-linear time-series

data. The splitting of a bigger time-series data into smaller sub-series forecasts electricity

7

load in a minimum time. However, there is still a scope of improvement if manual feature

extraction is replaced with automated feature extraction. At level 4, Enhanced Bidirectional

Gated Recurrent Unit (EBiGRU) is used in combination with FCN to learn the temporal

dependencies and automatically extract the important features along with the noise removal.

Gated FCN is used for electricity load forecasting. Overall, the work done in this thesis

ensures automated and efﬁcient load and price forecasting at a reduced computational cost

and time.

From critical analysis point of view, it is inferred that the feature engineering performed at

L. 1

L. 3

L. 2

L. 4

Short-term load

and price

forecasting

Storage

optimization

PVC power forecasting

Energy

management

Forecasting of decomposed intrinsic mode functions

Automated short-term load forecasting

Figure 1.1: Uniﬁed system model

level 1 is quite beneﬁcial for extracting the important and less redundant features. Moreover,

at level 1, cross-validation is also performed, which proves beneﬁcial in validating the results.

The performance analysis of the model using four different performance metrics also adds to

the efﬁciency of the model. However, using different techniques for feature engineering, cross

validation and performance analysis leads to an increase in the overall computational time of

the system, which can turn out to be a major issue. The work done at level 2 does not involve

the usage of complex techniques at different steps like feature engineering and also uses less

number of performance metrics. Moreover, simultaneously achieving multiple objectives

increases the computational time and cost. Furthermore, the comparison of work done in

the thesis with multiple state-of-the-art work shows that the proposed work is efﬁcient and

outperforms the existing works. From the scalability point of view, the work done in the thesis

8

Data source Load and

price data

Data acquisition

Data pre-processing

Feature engineering

Classification/Forecasting

Figure 1.2: Steps taken during load and price forecasting

is highly scalable as multiple datasets are used having different time duration. Moreover, the

use of an automated 8-layered deep learning model at level 4 proves the efﬁcacy of the work

proposed in the thesis.

1.4 Thesis Organization

The organization of the thesis is provided in this section. Chapter 2provides the state-

of-the-art related work. Different works done on load and price forecasting are discussed

in the same chapter. Chapter 3presents the proposed model for short-term electric load

and price forecasting in the SGs using enhanced ELM optimization. Two different sets

of techniques are used for load and price forecasting using two different datasets. The

9

Table 1.1: Comparison of different levels of the proposed system model

Characteristics L.1 L.2 L.3 L.4

Main objective(s)

Short-term electric-

ity load and price

forecasting

Energy management,

storage capacity

optimization and

PVC generated

power forecasting

Forecast decom-

posed mode func-

tions

Automated short-

term load forecasting

Technique(s) used ELR and ERELM

Stackleberg game

theory, GARIMA

and CARIMA

EEMD, SVM and

EBiLSTM

EBiGRU and FCN-8

Optimizer(s) used GWO GWO and CS – ADAM

Feature engineering

technique(s) used

RFE, CART and

Relief-F

– – FCN

Cross validation tech-

nique(s) used

Monte-Carlo and K-

Fold

– K-Fold –

Dataset(s) used UCI and UMass Elia NSW and VIC ISO-NE

Performance metric(s)

used

MAE, MSE, MAPE

and RMSE

MAPE and RMSE

MAE, MAPE and

RMSE

RMSE and MAPE

Benchmark scheme(s)

CNN, LR, ELM and

RELM

SVM, SAE and BP

EMD-BiLSTM-

SVM, EMD-PSO-

GA-SVR, BiLSTM,

SVM and MLP

BiGRU, CNN-

LSTM, MLP and

LSTM

10

existing techniques, simulation results and their discussion are also presented in this chapter.

Moving ahead, Chapter 4discusses the work done on energy management using the game

theoretical approach along with the storage capacity optimization and the forecasting of the

energy generated through PVC. The simulation results and their discussion along with the

comparison between the proposed techniques using game theory and without using it are also

provided. Chapter 5highlights the contributions made and the work done to decompose the

electricity consumption time-series data into linear and non-linear mode functions, and to

forecast both time-series individually. The hybrid model is proposed to perform the required

tasks and the simulation results prove the efﬁcacy of the proposed model. Furthermore,

Chapter 6presents the work done on electricity load forecasting using a hybrid model along

with an ensemble technique. The simulation results and their discussion prove that the work

done in this chapter outperforms similar existing works. Chapter 7concludes the thesis and

presents the future directions.

11

Chapter 2

Related Work

12

Many forecasting techniques have been used in the past for load and price forecasting. These

techniques can be categorized in three main groups: data driven [

130

]-[

134

], classical and

Artiﬁcial Intelligence (AI) [

135

]-[

139

]. Data driven techniques consider past data to predict

the desired outcomes. Classical methods comprise of the statistical and mathematical meth-

ods like ARIMA [

140

]-[

144

], Seasonal ARIMA (SARIMA) [

145

]-[

151

], Random Forest

(RF) [

152

]-[

158

], etc. Whereas, AI methods mimic the behaviour of biological neurons like

Feed Forward Neural Network (FFNN), CNN, LSTM, etc. In [

159

], behavioral analytics

are performed using Bayesian network and MLP. A number of experiments are performed

using the data obtained from the smart meters. Both short-term and long-term forecasting

is performed. In [

160

], Multiple Linear Regression (MLR) is used for forecasting purpose.

However, it has the limitation that it can not be used for long-term prediction. The authors

in [

161

] use residual network for forecasting load on the basis of weather data. The au-

thors in [

162

] use RBM to train the data and Rectiﬁed Linear Unit (ReLU) to predict the

electricity load. In [

163

], Discrete Wavelet Transform (DWT) and Inconsistency Rate (IR)

methods are proposed to select the optimal features from the feature set, which helps in

dimensionality reduction. Sperm Whale Algorithm (SWA) helps to optimize the parameters

of SVM. The authors in [

164

] propose a model for Short Term Load Forecasting (STLF).

Mutual Information (MI) is used for feature selection; whereas, better forecasting results are

achieved by modifying the ANN. In [

165

], the authors predict 24 hour ahead cooling load

of buildings using deep learning. The results show that deep learning techniques enhance

the load prediction. Similarly in [

166

], the authors use RNN, which groups the consumers

into pool of inputs. The proposed model is implemented using Tensorﬂow package and it

achieved better results as compared to conventional works.

ELM is a generalized single hidden layer FFNN that is proposed by the authors in [

167

,

168

].

It is proved to be effective in both regression and classiﬁcation methods. In [

169

]-[

174

], the

authors use the NNs for achieving better load prediction. In ELM learning processes, input

weights and biases are randomly assigned; whereas, the output weights are calculated using

the Moore–Penrose generalized inverse technique. In [

175

], the authors use Sparse Bayesian

ELM for multi-classiﬁcation purposes. The authors in [

176

] use Particle Swarm Optimiza-

tion (PSO) and Discrete Particle Swarm Optimization (DPSO) techniques for efﬁcient load

forecasting. The authors in [

177

] implement GWO with NNs to optimize weights and biases.

It is prove that optimization of weights and biases increases the efﬁciency of the network.

In [

178

], ELM is trained using Back Propagation (BP) by using context neurons as input to

hidden and input layers. Accuracy is improved by further adjusting weights using previous

13

iteration errors; whereas, biases and neurons selection affect prediction accuracy as already

discussed in [78].

In [

179

], different deep learning models are used for price forecasting. Based on the simula-

tion results, it is proved that deep learning models perform better than the statistical models.

In this paper, GRU is used, which is a variant of RNN. GRU outperform LSTM and many

other statistical models in terms of accuracy. In [

180

], price forecasting is done using a variant

of Auto Regressive Moving Average Model (ARMAX), known as Hilbertian ARMAX, which

uses the exogenous variables. The functional parameters used in the proposed model are

designed as the linear combinations of the sigmoid functions. These parameters are then

optimized using a Quasi Newton (QN) algorithm. In [

181

], CNN and LSTM are used for

price forecasting in PJM electricity market. In [

182

], Deep Neural Network (DNN) is used to

extract the complex patterns from the price dataset of Belgium. In [

183

], Grey Correlation

Analysis (GCA) is used along with Kernel Principal Component Analysis (KPCA) to deal

with the dimensionality reduction issue. For prediction, SVM is used in combination with

Differential Evolution (DE), where DE is used to tune the parameters of SVM. In [

184

],

a variant of autoencoder is used, which comprises of encoder and decoder. First, the data

is encoded to deal with space complexity. Once the output is obtained, it is decoded into

original form. The authors in [

185

] implement an enhanced form of Artiﬁcial Bee Colony

(ABC) algorithm, known as Time Varying Coefﬁcients Artiﬁcal Bee Colony (TVC-ABC),

for parameter tuning of Nonlinear Least Square SVM (NLS-SVM). The inputs are ﬁrst fed to

ARIMA and then the output of ARIMA is given to NLS-SVM. This ARIMA plus TVC-ABC

NLS-SVM is a Multi Input Multi Output (MIMO) forecast engine.

The limitations of gradient descent methods led researchers to evolve ELM based upon local

minima, learning rate, stopping condition and iterations of learning [

186

]. ELM performs

differently as compared to the traditional learning algorithms because it gives comparatively

less forecasting error and better generalization performance in [

187

]. Different versions of

ELM have also been proposed by researchers like Kernel Based Extreme Learning Machine

(KELM). Robust classiﬁcation is also performed in literature inspired by Mercer condi-

tion [

188

]. Moving ahead, solving the DEM problem at MG level in the Energy Internet

using both game theory and data analytics techniques is the modern trend. There is a recent

surge in data analytics that introduces various mathematical tools to address the uncertainties

[

189

]. While managing energy, two main methodologies are used: stochastic optimization

and robust optimization. These are widely used to handle data uncertainties [

190

]. Stochastic

optimization in energy management solutions is considered as one of the effective techniques

14

for the optimization of statistical objective function. The undeﬁned numerical data has

to be assumed in order to follow a well-known probability distribution. Real time energy

management techniques that utilize stochastic optimization are proposed to reduce the op-

erational cost of MG [

191

]-[

195

]. In [

196

], a multistage framework is proposed in order

to reduce the cost of the energy system based on stochastic optimization. In order to cater

multi dimensional energy management, stochastic dynamic programming methods are used

[

197

]-[

200

]. Nonetheless, the precise calculation of probability distribution of uncertain

data causes tremendous challenges in practical applications, which include the complex

operations and constraints. Optimality performance is also affected by the impact of data

uncertainties. Such issues are not dealt in the existing energy management approaches that

contain stochastic optimization.

Robust optimization based energy management approaches rely on limited information and

enable distribution free models for handling the uncertainties [

201

]-[

206

]. In the optimized

energy management systems, the worst case operational scenarios are considered. Besides,

optimized energy management approaches help in removing the negative impact of the

uncertainties faced while achieving optimal performance. Thus, a novel pricing strategy is

proposed, which promotes robustness against uncertainties of power inputs [

207

]. Nonethe-

less, the robust version of controllable electric load management is ensured to be intractable

that relies on appropriate designs for objective function modeling and building of uncertainty

set.

Ever since the growing advancement in the ﬁeld of ICT, a large amount of data is collected

regarding consumers’ behavior, states of battery, substations, customer devices, distributed

energy resources, renewable output, weather conditions, video surveillance, etc., [

208

]. The

energy generation using PVC is considered to be one of the important components of electric-

ity sector. It is also gaining attention of government because of decreasing environmental

issues and being cost effective. In [

209

], it is stated that PVC energy will be responsible for

16% of the total energy generation. PVC energy relies on solar radiations, which exhibits the

intermittent nature due to the position of sun and the movement of clouds that causes varia-

tions throughout the day. In addition, information of solar energy is the greatest concern both

for operators and planners of electrical systems. Hence, there is a sheer need of forecasting

PVC power generation [

210

] using machine learning techniques. The forecasting models that

are mainly used for prediction of electrical load and the RES are grouped in three categories:

statistical models like Auto Regressive (AR), Exponential Smoothing (ES) models [

211

]; AI

models like NN, CNN, and hybrid models like neuro fuzzy models [

212

,

213

]. As a case

15

study, electricity demand of 10 countries are taken into account in order to analyze these

methods. In [

214

], different forecasting techniques are used like multi-model, iterative, single

model multi-variate forecasting are analyzed in detail. It also covers different issues like

NN designing, implementation and validation. In [

215

], a combination of NN and enhanced

PSO are used in order to perform power forecasting, which focuses on feature selection. The

time-series of power generated by PVC contains many well deﬁned patterns, such as peaks

at afternoon and off peaks in morning and evening. Subsequently, there is no generation

during the night. Therefore, time-series patterns can not be ignored [

216

]. ARIMA is used for

uni-variate time-series forecasting because solar generation tends to follow speciﬁc patterns

and it is proven to be efﬁcient on the account of ﬂexibility. It also performs its orderly

searching at every level. Thus, it determines the best ﬁt model for particular time-series. The

optimization of parameters has signiﬁcant impact on the performance of forecasting algo-

rithms. Different meta heuristic techniques are applied for forecasting purposes like ARIMA

[

217

], NN [

218

], SVR [

219

] in literature for improving performance. Game theory concept

is widely applied on MG energy management studies. It provides distributed self organizing

and self optimizing solutions for the problems having conﬂicting objective functions. In a

broader aspect, game theory is characterized in two categories in context of players. The

ﬁrst category includes the players that have a binding agreement between them. While the

second category includes those players who are not having binding agreement between them

[

95

]. In non-cooperative game theory, the main focus is on predicting individual strategies

and it also assesses the players that make decisions to ﬁnd NE. It provides a framework for

performing analysis, i.e., DEM, which is devised for characterizing the interaction between

players and decision making process to achieve NE. The strategic outcome between players

can be improved under mutual commitment.

In order to handle non-cooperative game theory based energy management, a multi user

based Stackelberg game is used to optimize the payoff of each player [

220

]. In [

221

], a multi

stage market model is proposed, which is based on cooperative game to reduce the cost of

utility. However, it maximizes the total proﬁt of the market. In order to solve the dispatch

problem in integration of renewable resources generation and energy storage, a cooperative

distributed energy scheduling algorithm is proposed in [222].

Future events are determined by time-series forecasting. It is based on a series of given

historical events. In [

223

], building’s energy consumption is forecasted using linear regression

in the year 1980. The authors presented regression based forecasting technique that forecasts

16

residential energy consumption in Canada [

224

]. In 2012, a regression model was developed

to forecast load in banking sector in Spain [225].

There are two types of learning techniques in the literature: machine learning techniques

and deep learning techniques. The machine learning techniques are used in various ﬁelds

for forecasting purposes. They perform the best in pattern recognition as well. In [

61

], a

time-series of wind speed is used for its prediction along with the parameter optimization. It

includes two evolutionary algorithms: ANN and SVM. In a couple of decades, several hybrid

machine learning and deep learning techniques are introduced to improve the prediction

accuracy [

62

]. Gunay et al. proposed a hybrid of regression model and ANN in order to

forecast the electricity demand of Turkey [

63

]. It mainly works using a multi-variate dataset

where a regression model is used for feature selection. Afterwards, the selected features

are forwarded to the ANN for annual electricity demand forecasting. The approximation

capability of ANN is proven to be very effective. However, there are certain limitations,

such as overﬁtting and high computational time. Therefore, in literature, ANN is used in

hybrid models with other regression models, such as SVM [

226

]. Each model possesses

certain pros and cons, which cannot be overlooked. Neural networks such as MLP give

accurate forecasting results in case of the complex and non-linear scenarios. However, the

computational time is high, which consequently increases the overall cost of the system.

Therefore, it is the need of the hour to propose the hybrid techniques, which use ANN

for forecasting the complex and non-linear characteristics of the time-series. Whereas,

linear and stationary characteristics need to be forecasted using the techniques that consume

less computational resources, such as SVM. The characteristics of time-series data can be

decomposed and separated with the help of signal decomposition techniques [

227

]. ANN

neither performs feature engineering nor requires hand crafted feature selection. It rather

requires intensive domain knowledge, which makes it a difﬁcult choice. Therefore, it is

required to incorporate deep learning techniques such as CNN and LSTM. These techniques

are considered as the building blocks of deep learning models [

67

], [

228

]. Various variants

of these building blocks have been proposed in order to enhance the network ﬂexibility like

BiLSTM [

68

]. However, BiLSTM has an inherited issue of poor generalization. Therefore,

an ensemble strategy, which takes the average weights of a multi-model BiLSTM, is used in

the proposed work.

The aforementioned literature contains certain drawbacks. The researchers are using the same

learning model for time-series forecasting. However, it is a known fact that different models

17

handle volatility and variability of the data in different ways. Therefore, a large number

of researchers are focusing on the concept of decomposing the time-series into multiple

sub-series. So that the sub-series forecasting can be handled separately by adjusting the

models. These decomposition techniques are proven to be efﬁcient in forecasting as compared

to the sole forecasting techniques [69].

The electricity load time-series shows ﬂuctuations in intensities. Different techniques have

been introduced in order to analyze different levels of data intensities like wavelet method

[

229

]-[

233

], Fourier transform [

234

]-[

238

], EMD and EEMD [

70

]. Among the ensemble

methods, divide and conquer model is considered to be the best. Where, the actual time-series

is decomposed into multiple orthonormal sub-series on the basis of time domain. It helps

in removing the noise as well as in extracting high and low intensity based sub-series. All

of these techniques are applied on different time-series data and the ﬁndings reveal that

EEMD, which is a modiﬁed form of EMD, performs well in terms of acquiring forecasting

accuracy. In EMD, the time-series signal is decomposed in a number of Intrinsic Mode

Functions (IMFs) and residue. These obtained IMFs and residue by EMD reveal different

characteristics. EMD faces the mode-mixing problem while decomposition of the signal

takes place. Mode-mixing problem is either deﬁned as a single IMF that contains multiple

characteristics or representation of same kind of characteristics by different IMFs. It makes

the interpretation of sub-series difﬁcult. Therefore, EEMD is brought to the light by Wu

[

239

], which is a noise assisted data preprocessing technique that adds white noise to the

time-series. Furthermore, it also helps in the alleviation of mode-mixing problem. In [

240

],

the EMD signal decomposition technique is used to decompose the signal into multiple

sub-series and residue. The extracted sub-series are forecasted using LSTM forecasting

algorithm. There is a drawback that sub-series are not classiﬁed separately according to

linearity and non-linearity. Consequently, forecasting becomes extensively time consuming.

In the proposed work, EEMD is integrated with EBiLSTM and SVM. EBiLSTM is a deep

learning model that handles the complex and non-linear data efﬁciently. However, it has

high computational overhead. In order to reduce the computational time, linear IMFs are

forecasted using SVM as it shows good accuracy in order to learn linear data.

Table 2.1: Summary of related work

Objectives Forecast

Horizons

Data Resolution Datasets Techniques

18

Load forecasting

[36]

Short-term

Hourly building

cooling loads and

the meteorologi-

cal data

TRNSYS simula-

tion tool for elec-

tric load demand

data of Northeast

China

ANN and ensem-

ble method

Load Forecast-

ing [40]

Short-term

Monthly average

load demand

Assam north-

eastern state of

India data

GOA and SVM

Load forecasting

[45]

Long-term Yearly data

China electricity

data

SVM

Load forecasting

[60]

Short-term

and mid term

Half hourly data AEMO data EEMD-DBN

Wind speed fore-

casting [61]

Short-term

Hourly wind

speed time-series

data

Two wind speed

time-series data

EEMD with

adaptive noise

ANN

Load forecasting

[99]

Short-term

Hourly historical

data

Provincial capital

of Soria data

ANN

Load Forecast-

ing [100]

Short-term

Daily and hourly

data

PJM data EEMD-LSTM

Price and load

forecasting

[101]

Short-term

and mid term

Hourly data

NYISO and PJM

data

ESAE, NARX

and DE-ELM

Load forecasting

[102]

Very short-

term and mid

term

Hourly data EUNITE data

EEMD and

LSTM

Price forecasting

[103]

Long-term Hourly data

China’s power

grid investment

data

EMD, GA-SVM

and Radial basis

function ANN

Load forecasting

[119]

Short-term

Half hourly resi-

dential data

Data of 10,000

different cus-

tomers in NSW

LSTM

19

Load forecasting

[120]

Short-term

and mid term

Electricity hourly

load datasets that

contain daily av-

erage load

PJM data EEMD-SVR

Bayesian, MLP

[159]

UKDale UK

Forecasting

done using

behavioural

analytic

Requires inten-

sive training

MLR, BaggedT,

NN [160]

Beijing Beijing, China

Comparison

between tech-

niques done to

overcome the

limitations

Not suitable for

long term fore-

casting

DRN [161] ISO-NE New England

Load forecast-

ing done using

weather data

Overﬁtting

RBM,

ReLU [162]

Korea

Electric

Company

Korea

Two stage

forecasting

performed

Long-term fore-

casting not sup-

ported

DWT-IR, SVM,

SWA [163]

NYISO,

AEMO

Australia, US

Dimensionality

reduction and

paramater opti-

mization done

Time complexity

Modiﬁed MI,

ANN [164]

PJM US

Two stage fore-

casting is done

Time complexity

DAE [165] Hong Kong Hong Kong

Cooling load pre-

diction done

Time and space

consuming

Pooling Deep

RNN [166]

IRISH Ireland

Pooling of con-

sumers done

for aggregated

prediction

Difﬁcult to train

20

ELM [167,168] USF US

Long, medium

and short-term

forecasting done

Overﬁtting

SLFN [

169

,

174]

Marine

Resources

Division

Australia

Optimization of

weights

Overﬁtting by

using moore-

penrose inverse

Sparse Bayesian

ELM [175]

Harvard

medical

college

USA

Optimization of

weights and bi-

ases using BP

Require intensive

training

PSO,

DPSO [176]

US USA

Compact ANNs

are produced

Large computa-

tional time

GWO with

NN [177]

Load USA

Weights and bi-

ases optimization

Time complexity

RELM [178]

Bench mark

UCI machine

Portugal

Use of context

neurons

Computationally

expensive

LSTM, DNN,

GRU [179]

EPEX Belgium

Comparison

between different

models

Overﬁtting

Hilbertian, AR-

MAX [180]

EPEX Spain, Germany

Optimization of

functional param-

eters for price

forecasting

Non-linearity

CNN,

LSTM [181]

PJM US

Two NNs are

used for price

forecasting

Computationally

expensive

DNN [182] EPEX Belgium

Complex patterns

are extracted for

prediction

Space complex-

ity

GCA, KPCA,

DE-SVM [183]

ISO NE-CA New England

Dimensionality

reduction is

removed using

hybrid of KPCA

and GCA

Overﬁtting

21

SDA [184] MISO

Arkansas, Texas

and Indiana

Variant of autoen-

coder used

Computationally

expensive

ARIMA, TVC-

ABC, NLS-

SVM [185]

PJM, NY-

ISO, AEMO

Australia, US

Parameter tuning

of SVM done us-

ing TV-ABC

Computationally

expensive

ANN with

meta heuristics

optimization

methods [186]

Load and

price/Com-

mercial load

of building

in China, Tai-

wan regional

electricity

load

Taiwan

Various para-

mater calcula-

tions done for

accuracy

Accuracy of mod-

els depend on na-

ture of dataset

OS-ELM with

kernel [187]

Load and

price/Sylva

bench mark

US

Comparison

of different

algorithms done

Restrict to the

computation of

streamed data

ELM in multi

class sce-

nario [188]

Load and

price/Uni-

versity of

California

Irvine

Canada

Robust classiﬁca-

tion

High computa-

tional cost

Many studies are conducted in literature for EC forecasting, which are based on statistical,

machine learning and deep learning models. There are several techniques based on these

models that are proposed by researchers in the course of time. These techniques include

Auto-Regressive Integrated Moving Average (ARIMA) [

226

], LR [

241

], SVM [

242

], ANN

[

243

], sequence to sequence learning [

244

], DANN [

245

], etc. Table 2.1 presents the main

ﬁndings of literature review in terms of proposed schemes, contributions, performance

metrics, results and limitations. According to the statistics provided in [

246

] regarding the

application of these forecasting techniques in research, ANN is being used up to 47%, SVM

25%, Decision Tree (DT) 4% and the rest of the techniques are employed up to 25% in the

literature. The forecasting of residential electricity load is performed using ANN in [

127

].

However, the effectiveness of ANN increases as per the availability of different exogenous

22

variables like weather data, temperature, wind speed and humidity. The nonavailability of

dependent variables adversely affects the performance of ANN. RNN resolves this issue with

the help of BP in the network. It yields promising results while forecasting sequence-based

time-series. However, long-term dependencies are hard to learn by traditional RNN as it faces

the vanishing gradient issues due to time depth and single hidden layer. In order to overcome

these issues, the LSTM forecasting technique was proposed in 1997 [

129

]. It combines

short-term memory with long-term memory by incorporating three memory gates, which

save the long-term dependencies in them. In [

247

], LSTM is combined with GA in order to

optimize the model’s hyper-parameters. The comparison between different deep learning and

machine learning models is performed in [

248

], which proves that the performance of deep

learning models is reasonably better as compared to the machine learning models. DANN is

used as the forecasting technique in [

128

]. It is a multichannel deep convolution network,

which uses the variables’ data to detect latent features. The extracted latent features are fed to

MLP for forecasting the time-series data. However, MLP performs poorly when detecting

temporal dependencies. In [249]-[253], deep LSTM is used in order to perform forecasting.

Various variants of LSTM are proposed in the literature that include DLSTM. It uses skip

connections that are extracted from the concept of ResNet [

161

]. Moreover, DLSTM enhances

the efﬁciency of the network by reducing the vanishing gradient issue [

254

]. BiLSTM is used

in literature to learn two-way dependencies of the data: past and future [

255

]. GRU is an

advanced version of LSTM that combines the input gate and forget gate of the network into

one gate, termed as an update gate [

256

]. Multi GRU is also a variant of GRU that is used to

optimize the electricity dispatch plan [

257

]. In [

258

], GRU is used in order to forecast PV

power generation. Apart from time series sequence data, there are other high dimensional

information as well like spatiotemporal matrix that exists in the time-series, which can not

be learned by GRU. Therefore, it is required to add such features in GRU that can optimally

learn the high dimensional temporal data. In deep learning methods, feature extraction is

ideally performed by CNN. It is widely used in the ﬁeld of image and pattern recognition. It

also learns local features that are based on a strong relationship between nearby points [

259

].

Researchers in the ﬁelds of image recognition, speech recognition and emotion recognition

frequently use CNN. It is an end-to-end learning model that simultaneously covers spatial and

temporal trends [

260

]. Moreover, CNN-LSTM is used in speech recognition to learn global

and local emotion-based features. Multiple blocks are used in the network, comprising of

convolution block and pooling block that learn the local correlation in the data. LSTM learns