Semester 8 Project

Report

Assessment of Forecasting Strategies on

Multivariate Time Series Data
Submitted in partial fulfillment of
the requirements for the award of the degree of

Bachelor of Technology
in
Information Technology

Submitted by

Roll No Names of Students

IIT2016042 Manavdeep Singh

IIT2016072 Badnena SVVR Upendra

Under the guidance of

Prof. Dr. O.P. Vyas

Department of Information Technology

Indian Institute of Information Technology, Allahabad
Allahabad, Uttar Pradesh, India – 211012
Apr 25, 2020
Department of Information Technology
Indian Institute of Information Technology Allahabad

Certificate

This is to certify that this is a bonafide record of the project presented by

the students whose names are given below in partial fulfilment of the require-
ments of the degree of Bachelor of Technology in Information Technology.

Roll No Names of Students

IIT2016042 Manavdeep Singh

IIT2016072 Badnena SVVR Upendra

Prof. Dr. O.P. Vyas

(Project Guide)

Date:Apr, 25 2020
 Abstract

In competitive power markets, electric utilities, power producers, and

traders are exposed to increased risks caused by electricity price volatility.
The aim of this research is to introduce a short-term electricity price forecast-
ing method that addresses the problems of price volatility, a varying number
of input parameters, varying data availability, and a large number of param-
eters and input data. Furthermore, the proposed model can be used on any
market as it targets the characteristics and specifics of each market. Here,
we combine data for the state of Ontario, Canada made available by Inde-
pendent Electricity System Operator(ieso)[1] for electric demand and prices
for ten different zones of Ontario. On the other hand, to evaluate the rela-
tionship of demand, price and weather conditions, we take the weather data
of the same time period from the website of Government of Canada[2].

In this paper, we propose to compare the effectiveness of Recurrent Neu-

ral Networks(RNN) and Long Short Term Memory(LSTM) for short term
forecasting and Hidden Markov Models on time series data. We have trained
these algorithms on our data-set and compared the results. In our experi-
ments, RNN gave better results than LSTM. We also found out that HMM is
also an applicable solution to time series prediction problems where a number
of hidden factors affect our output.
Contents

1 Introduction 1

2 Problem Statement 3

3 Literature Survey 4

4 Data Set Description 6

4.0.1 ieso Power Data Directory . . . . . . . . . . . . . . . . 6
4.0.2 Climate and Weather Data, Canada . . . . . . . . . . . 8
4.0.3 Data Preprocessing . . . . . . . . . . . . . . . . . . . . 9

5 Methodology 14
5.0.1 Recurrent Neural Network - RNN . . . . . . . . . . . . 15
5.0.2 Long Short Term Memory - LSTM . . . . . . . . . . . 16
5.0.3 Hidden Markov Models-HMM . . . . . . . . . . . . . . 18
5.0.4 Implementation Plan and Timeline . . . . . . . . . . . 19

6 Results 21
6.0.1 Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . 21

Disposal of Comments given by respected Board Members dur-

ing first evaluation 32
6.0.2 Comparison of LSTM with RNN for short term load
forecasting . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.0.3 To try the implementation of Hidden Markov Model
for time series analysis . . . . . . . . . . . . . . . . . . 32

Acknowledgements 33

References 34

i
Chapter 1

Introduction

Electricity is a very special commodity. Electricity prices differ from other

assets and commodities due to its unique features such as requirement of
having constant balance between the supply and demand sides and non-
storability[21]. Electricity demand depends on a lot of factors, including but
not limited to, weather(temperature, wind speed, precipitation, etc.), every-
day activities (time of the day, weekdays vs. weekends, holidays, etc), and
so on. The price volatility of electricity is extreme, which can be up to two
orders of magnitude higher than that of any other commodity or financial
assets[22].
If electricity price can be predicted accurately, power producers and power
generation companies could develop suitable generation plan and maximize
their profits by grasping market dynamics. Power consumers can choose the
time they want to use power and the quantity they want to buy, so that it
can reduce costs and increase the market competitiveness of enterprises.

Forecasting is prediction of some future event or events by analyzing the

historical data. It is used in many areas including business and industry,
economics, environmental science and finance. Forecasting problems can be
classified as
• Very Short term forecasting (prediction for few seconds, minutes)
• Short term forecasting (prediction for few hours, days, weeks or months).
• Medium term forecasting (prediction for 1 to 2 years).
• Long term forecasting (prediction beyond 2 years).
Many of the forecasting problems involve the analysis of time. A time
series is a series of data points indexed in time order. Most commonly, a
time series is a sequence taken at successive equally spaced points in time.

1
 Through this paper, we focus on short term electricity price forecasting
using Long short-term memory(LSTM), Recurrent Neural Network(RNN)
and also comparison between these two models. Both LSTM and RNN mod-
els forecasts prices for the next 6 hours. We have also implemented Hidden
Markov Model(HMM) for our problem and discussed the results in brief. We
are predicting the price variable for 6 hours ahead, by seeing data for 5 days
behind for both LSTM and RNN, while a training size of 7000 tuples is used
to predict a test set of 3000 tuples in HM Model.

2
Chapter 2

Problem Statement

The problem of Electrical Demand Forecasting as a standalone variable was

attempted by us in the project presented in the previous semester. We used
an LSTM model to predict the load requirements for five different data sets
and compared our results with some benchmark algorithms. In this paper,
we attempt to build upon our LSTM model for Electrical Price forecasting
along with other variables on which it may depend- Demand and various
weather variables. For this purpose, we aim to construct a data-set by com-
bining data from various files available on public web interface, particularly
ieso.ca[1], climate.weather.gc.ca[2] for the region of Ontario, Canada-which
is further divided into ten different regions. We attempt to construct a text
only data-set of size greater than 1 GB. Upon this data-set, we get results
using an LSTM model and compare those results with a simpler RNN model
for Short Term Price Forecasting.
We also explore into a probabilistic model-HMM and try to identify it’s vi-
ability for a time series problem, as an additional task. We chose HMM
because all the parameters which drive the Electrical Prices are not visi-
ble to the observer and there are some hidden factors which makes prices
highly volatile in nature. Hidden Markov Models have a strong probabilistic
framework for recognizing patterns in stochastic processes. They have been
used for analyzing patterns in speech, handwriting and gestures and are still
extensively used in those areas. We include results for HMM in the end.

3
Chapter 3

Literature Survey

Paras Mandal, et al. used a Recurrent Neural Network technique based

on a Similar Day(SD) approach for day ahead electrical price forecasting[3].
They used various parameters like Mean Absolute Percentage Error, Root
Mean Square error to validate their model. They got an R2 value of 0.7758.
They get a significant accuracy to conclude that the model is good for short
term forecasting. S. Selvin, et al. quantified the performance of their sliding
window approach for futurer predictions on a short term basis for compa-
nies listed in NSE, using Percentage error[4]. They use three deep learn-
ing techniques namely RNN, LSTM, and CNN.The performance of RNN
and CNN were competitive and better than LSTM for short term forecast-
ing, with CNN giving slightly better results. Zihan Chang, et al. used an
adam-optimised LSTM network on New South Wales of Australia data-set
to illustrate the prediction accuracy of their model[5]. They achieved an
MAE value of 2.43 and MSE value of 10.84 which was better than earlier
achieved values. Israt Jahan and Sayeed Sajal used a Recurrent Neural Net-
work Model for Stock price prediction and got a percentage of error below
5%[6]. Houda Bakir, et al. used LSTM Neural Network Models to predict e-
commerce prices in their paper[7]. They compare their results with Support
Vector Regression Model. The LSTM Model was able to predict next day
prices with RMSE of 23.640 euros as opposed to 33.43 euros in case of SVR.
Jian Zheng, et al. used LSTM based Recurrent Neural Networks to predict
electrical load for short term[8]. They concluded that the results compared
favourably to many other established forecasting techniques.For very small
sections of data, their model gave results comparable to SARIMA models
when no seasonality was present. Zhitang Chen, et al., used Hidden Markov
Models for predicting future network traffic[9]. They proposed to model the
relationship between the volume of traffic and simple statistics of flow using
HMM to avoid direct calculation of traffic volume. They recorded results by

4
the application of HMM on kernel bayes rule as well as Recurrent Neural
Networks. They concluded that in high speed networks with large traffic
requirements, these models can be very successful due to non requirement
of calculating traffic volume. Tie Liu in his paper on Application of Markov
Chains to Analyze and Predict the Time Series[10], applied Markov Chains to
analyse and predict time series using first-order discrete-time Markov Chains
for some cases and higher order Markov Chains for others. He summarised
that the accuracy of markov chain models for prediciton of time series is very
well.
Hippert et al. in a paper reviewed and evaluated traditional method us-
ing neural networks for Short-Term Load Forecasting[11]. Various other ap-
proaches like using support vector machine for load forecasting for a EUNITE
competition[12], using fuzzy neural networks[13], using knowledge-based ex-
pert systems[14] have been developed or the same analysis. Another approach
develops iterative reweighed least squares algorithm for short term power
system load forecasting[15]. Another research conducted by Mohamed A.
Abu-El-Magd et al. drew a comparison between online and offline methods
for short-term electric load forecasting. The load demand was also mod-
eled using multivariate time series analysis [16]. G. T. Heinemann and team
studied temperature sensitive and non-temoerature sensitive load and did a
regression analysis for the same [17].

5
Chapter 4

Data Set Description

For the problem of price forecasting that we tried to solve, we used the
following two data sources :
1. ieso Power Data Directory. [1]
2. Climate and Weather Data, Canada. [2]
We constructed a data-set by combining data from these two data sources
for the region of Ontario, Canada.

4.0.1 ieso Power Data Directory

This data is provided by Independent Electricity System Operator(ieso). The
data source contains the market demand, zone-wise demand(10 zones in to-
tal), real-time market price of electricity in Ontario, Canada. The data is in
comma separated values(.csv) format and each file contains hourly data for
one particular year.

The market demand and zone-wise demand is hourly data whereas the
real-time market data is available for every 5 minute interval. The availability
of data is as follows:-
1. Ontario and Market Demand - 2002-present, avg. file size = 153 kB
2. Zonal Demand - 2003-present, avg. file size = 429 kB
3. Real time market price report - 2010-present, avg. file size = 24 MB
As can be seen, for data to be available for all attributes, i.e. the years for
which data is available for all the three above mentioned reports, we have
taken the data for the years 2010-2019.

6
 Column Name Datatype
Date datetime
Hour int
Market Demand int
Ontario Demand int

Table 4.1: Data Description - Ontario and Market Demand

Column Name Datatype

Date datetime
Hour int
Zonal Demand(count=10) int
Zone Total int
Diff int

Table 4.2: Data Description - Zonal Demand, Total columns = 10 + 4

Column Name Datatype

DELIVERY DATE datetime
DELIVERY HOUR int
INTERVAL int
Regions(count=15)
10N float
30R float
ENGY float

Table 4.3: Data Description - Real time Market Price Report,

Total columns = 3*15 + 3

7
4.0.2 Climate and Weather Data, Canada
This data-set is provided by Government of Canada. The data-set contains
the weather data of Ontario, Canada. There are 20+ attributes in each file
such as temperature, dew point, humidity, etc. The data is in comma sepa-
rated values(.csv) format and each file contains hourly data for one particular
month in a year. The average size of each file is 119 kB.

Column Name Datatype

Longitude float
Latitude float
Station Name string
Climate ID int
Date/Time datetime
Year int
Month int
Day int
Time time
Temp float
Dew Point Temp float
Dew Point Temp Flag -
Rel Hum int
Rel Hum Flag -
Wind Dir (10s deg) int
Wind Dir Flag -
Wind Spd (km/h) int
Wind Spd Flag -
Visibility (km) float
Visibility Flag -
Stn Press (kPa) float
Stn Press Flag -
Hmdx -
Hmdx Flag -
Wind Chill float
Wind Chill Flag -
Weather string

Table 4.4: Climate and Weather Data, Canada.

8
4.0.3 Data Preprocessing
In ieso Power Data Directory data-set, not all three i.e Ontario and Market
Demand, Zonal Demand and Real time market price report were not avail-
able from 2002. As we have to combine all three we have used the data from
2010 to 2019. Climate and Weather Data is also available for these years.

Now we had three issues to address from the collected data:

• Missing values in the data.

• While Ontario and Market Demand, Zonal Demand and Climate and
Weather Data are in the form of hourly data, Real time market price
report is available for every 5 minute interval. Thus, one has to be
adjusted in the form of the other.

• In Climate and Weather Data as there were 20+ attributes. Selection

of only relevant attributes is necessary.

4.0.3.1 Filling missing values

In both Electricity data and Weather data there were missing values due to
which the output variable will have an inherent bias. There can be many
ways to handle missing values, for example- filling with mean, median or
some statistical quantity, or filling a random value. Another way can be to
ignore missing values at all or ignoring entire tuples with missing values. But
to make our data not decrease in size and still be robust, we chose to fill the
missing values using a technique called linear interpolation.
In order to remove this bias we have filled the missing values using 1-D Linear
Interpolation.

Linear interpolation is a linear function. Suppose two points (x0, y0) and
(x1, y1) exist, and an unknown function f, then interpolating a point between
two points, thus, the interpolated point f(x, y) can be denoted as:

f ( x, y) = f ( x0 ) + w · ( f ( x1 ) − f ( x0 ) )

= ( 1 − w) f ( x0 ) + w · f ( x1 )
Where w is denoted by:
y − y0 x − x0
w= =
y1 − y0 x1 − x0

9
 In python, there is a in-built function interpolate() which is used to fill NA
values in the dataframe or series. interpolate() function takes arguments like
method of interpolation to be used and direction in which it has to perform
interpolation. Linear Interpolation can be done two directions in python:
forward and backward. We have used linear interpolation in backward direc-
tion as there were more missing values in the beginning of our dataset i.e in
the year 2010.

4.0.3.2 Upsampling Data

In order to combine both data-sets i.e Electricity Data and Weather Data we
had to upsample Ontario and Market Demand, Zonal Demand and Weather
Data as they were hourly data and Real-time Market price was available for
every 5 minute interval.

As we had hourly values and we needed values for every 5 minute in-
terval again it is similar to missing values problem and we have used linear
interpolation again to upsample the data. After upsampling the data and
combining both weather and electricity data the resulting dataset(Table 4.5)
had 170 columns and about 1.2 gigabytes in size.

4.0.3.3 Feature Selection

With respect to a data-set, a feature simply means a column. Each column
has some impact on the output variable. While some features are necessary
because the output depends on them to some considerable extent, including
irrelevant features will only make our model worse.
Since the Climate and Weather data that we had for the province of Ontario,
Canada, had more than ten different weather attributes, from Temperature
to visibility, there was a requirement for feature selection for selecting only
the relevant variables. Since our data is numeric and continuous in nature,
we only explored numerical feature selection techniques. We discuss them
one by one hereforth.

1. Pearson Coefficient :- We can make a correlation matrix using Pear-

son Coefficient. We noted the three attributes with highest correlation
values given that they have correlation value greater than 0.05. We
used the corr()[18] function available in pandas library for python 3.
The formula for pearson coefficient is as follows:
cov(X, Y )
ρX,Y =
σX σY

10
 Column Name Datatype
Date datetime
Hour int
Minute int
Market Demand float
Ontario Demand float
Regions(count=15)
10N float
30R float
ENGY float
Zones(count=10)
Zone Demand float
Longitude float
Latitude float
Station Name string
Temp float
Dew Point Temp float
Rel Hum float
Wind Dir (10s deg) float
Wind Spd (km/h) float
Visibility (km) float
Stn Press (kPa) float
Wind Chill float

Table 4.5: Data Description - Prepared Dataset,

Total columns = 15*3 + 10*12 + 5

11
 • cov is the covariance
• σA is the standard deviation of A

2. Backward Elimination :- This is a computationally expensive algo-

rithm than Pearson Coefficient, but it is an iterative and more accurate
model than it. We use an Ordinary Least Squares model for performing
Linear Regression and selecting the best attributes. Attributes are re-
moved one by one according to pvalue performance matrix. Finally we
get the attributes namely, Temp, Dew Point Temp and Rel Hum, which
incidentally are the same as those obtained by Pearson Coefficient.

3. Recursive Feature Elimination :- As the name suggests, this method

recursively removes irrelevant features. It uses an accuracy metric to
rank attributes by importance and gives them a number starting from
1 as most important. Here again, Temp, Dew Point Temp and Rel
Hum were among the most important top three features.

Therefore, from all available attributes for weather data, we used Temp,
Dew Point Temp and Rel Hum as the relevant attributes, because these were
common amongst all above techniques used for feature selection. After re-
moving all the irrelevant attributes from our data-set, now we had 46 columns
and about 300 megabytes in size(Table 4.6).

12
 Column Name Datatype
Date datetime
Hour int
Minute int
Market Demand float
Ontario Demand float
Ontario Price float
Zones(count=10)
Zone Demand float
Temp float
Dew Point Temp float
Rel Hum float

Table 4.6: Data Description - Final dataset after Feature Selection,

Total columns = 10*4 + 6

13
Chapter 5

Methodology

In this section we are going to explain the steps that we followed. The
methodology can be represented as the following pipeline:-

Figure 5.1: Block Diagram of the pipeline

We started the process with the preparation of data, the steps to which
we have discussed in the previous section. Before explaining about each
algorithm, we shall explain the steps that were common to all of them. Our
data-set now had three columns namely-Date, Hour, Minute, upon combining
which, we got our time stamp, named ”period”, which we used as the index
of our data moving forward. To fill some missing values that might have been
left, we backward interpolated our data because the largest chunk of missing
data was present at the start of the year 2010.
In addition to these steps which were common for all three algorithms, to
make a comparative study between RNN and LSTM, we used similar model
definitions for both, about which we will be talking in the coming sections.
We used tensorflow’s open source library-HParam for hyperparameter tuning,

14
including number of cells in each layer, the optimiser function and the drop
rate.
We shall now look at each of the algorithms one by one.

5.0.1 Recurrent Neural Network - RNN

Neural networks are aimed at resembling the human brain cells and the way
they think. But traditional neural networks lack the principle of persis-
tence, that is human brain stores some information and doesn’t always start
thinking from scratch. This seems like a major shortcoming. This issue is
addressed by Recurrent Neural Networks, which have a loop in their struc-
ture that allows the persistence of information.

Figure 5.2: loop in RNN structure

Upon unrolling the loop, we might be able to see the working more clearly.
At each stage, the chunk of Neural Network - A, takes some input xt and
gives some output ht . The loop allows it to take some part of the output and
pass it to the next time step t+1.
RNNs have been applied to a variety of problem incuding but not limited
to speech recognition, image captioning, transation and others. But on the
question of prediction with time series data, a special type of RNN- Long
Short Term Memory of LSTM Networks have been used majorly. While
Long Short Term Memory networks have a clear advantage over observing
long term dependencies due to their innate feature of memory, for the prob-
lem of Short term forecasting we wanted to study if RNNs might just do the
trick as we won’t need to study dependencies over a long term.

15
 Figure 5.3: an unrolled Recurrent Neural Network

Upon using tensorflow’s HParam API, we zeroed in on the following pipeline

for our model(s).

Figure 5.4: our LSTM pipeline

5.0.2 Long Short Term Memory - LSTM

Long Short Term Memory(LSTM) Networks are a special type of recurrent
neural networks, which have been used for various tasks like music comple-
tion, handwriting generation among others. They are much more effective to
those specific problem areas, than the standard version.
RNNs appeal to computer scientists due to their ability to use recent past
data to predict the near future characteristics of the system. But, the prob-
lem area is when data older than just the immediately previous plays some
part in determining the future output. Here the gap between relevant in-
formation and the point where it is needed becomes sufficiently large to be
out of the scope of standard RNNs. Theoretically, RNNs can handle such
dependencies, but, in practice, they fail to do so.
LSTMs, on the other hand, are capable of learning long term dependencies.
They were introduced by Hochreiter Schmidhuber (1997)[21]. LSTMs were
designed to solve the long term dependency problem. By their default be-
havior, they remember long term information. LSTMs have a similar chain
like structure like RNNs but they differ in the repeating module such that
instead of a single neural network layer, they have four, as shown in the figure

16
below.

Figure 5.5: four interacting layers in LSTM repeating module

The key to LSTMs is the cell state which is the straight line running
through the repeating module. It is like a conveyor belt which runs through
the module with minor interactions which are controlled by gates. Gates are
composed of sigmoid neural net layers and multiplication pointwise functions.
They output numbers between 0 and 1 where 0 means do not let anything
through and 1 means let everything through.
The first step is the forget gate layer, which looks at the current cell state,
and from h(t-1) and x(t), decides how much part to keep and what to forget.

The next step is the input gate layer in which new information is added
to the cell state through a sigmoid neural net layer and a tanh function. In
the next step, these two layers are combined to give an update to the state.
The old cell state is now updated to the new cell state by first multiplying

with ft and then adding it *Ct .

Finally, we decide what to output by filtering the previously generated

new cell state. This is again done by using a sigmoid layer and a tanh layer.

17
 In our model, we have used the following architecture. The results achieved
are discussed further in the paper.

Figure 5.6: our LSTM pipeline

5.0.3 Hidden Markov Models-HMM

The Hidden Markov Model was introduced in 1966 by Rabiner and Juang[19].
Hidden Markov Models are a great tool to observe non stationary systems,
electrical price is one such system. The observations of electrical price system
are continuous in nature. There are some visible states in a markov process
and there can be many hidden states. A typical Hidden Markov Process is
shown in the diagram below:

The three main problems in solving of an HMM are:

1. to compute probabilities of observations based on observation data and

model parameters.

2. finding the best corresponding sequence based on the observation data

and model parameters.

3. calibrating the HMM parameters.

Due to dearth of available resources and the complexity of solving the above
problems, we used a model similar to the one proposed by Nguyet Nguyen[20],
by using the the Bayesian information criterion, BIC to test the performances
of HMM with different numbers of states, variable between 2 and 15. Due to
the property of the above said model of prefering higher number of states to

18
 Figure 5.7: hidden markov process

form complex models leads to overfitting and that can be seen in our results
which have high r2 values, above 0.99.
The excessive run times caused due to the lack of computational resources in
our personal systems lead us to using only 1/100th of our data set for pre-
diction. Thus we present HMM only for completion and not for comparison
to the above two stated algorithms.

5.0.4 Implementation Plan and Timeline

As can be seen in the image, we went on with the implementation of the
solution to the problem statement by starting with literature review of similar
papers written in the past which talk about some of the ways in which the
problem has been tackled. Then we went on to implement the algorithm
of LSTM and then followed it by other algorithms like RNN and HMM.
Further, after successful implementation of each, we compared and compiled
the results obtained through two parameters on our data-set, namely r2 error
and root-mean-square error. We also measure mean absolute percentage error
in case of HMM. We present a tabulated presentation of the same further.

19
Figure 5.8: our implementation plan and timeline

20
Chapter 6

Results

6.0.1 Outcomes
We will start the results with the tables of both Long Short Term Mem-
ory(LSTM) and Recurrent Neural Network(RNN) parameters. Then we pro-
ceed on to the comparative analysis of both via images. Then we present
the Table of Parameters of Hidden Markov Model(HMM) and in the end are
some of the images for sample outputs.
From the tables, it is easy to see that for short term load forecasting, a sim-
ple RNN model performs better than an LSTM Model. These results are
also illustrated in the graphs that follow. This can be attributed to the fact
that long term dependencies need not be considered in the case of short term
forecasting, therefore a simpler model outperforms the complex one.

21
Figure 6.1: LSTM results

22
Figure 6.2: RNN results

23
Figure 6.3: RMS comparison of RNN and LSTM

Figure 6.4: r2 comparison of RNN and LSTM

24
Figure 6.5: Training time comparison of RNN and LSTM

Figure 6.6: Testing time comparison of RNN and LSTM

25
Figure 6.7: Runtime comparison of RNN and LSTM

26
 27
Figure 6.8: HMM results
Figure 6.9: RMS of HMM

Figure 6.10: r2 of HMM

28
Figure 6.11: MAPE of HMM

Figure 6.12: Runtime of HMM

29
Figure 6.13: Sample RNN Output

Figure 6.14: Sample LSTM Output

30
Figure 6.15: Sample HMM Output

31
Disposal of Comments given by
respected Board Members
during first evaluation

6.0.2 Comparison of LSTM with RNN for short term

load forecasting
Comparison using two parameters namely Root Mean Square Error and R2
Score have been used. For our data-set, RNN gives a better result as com-
pared to LSTM for short term price forecasting.

6.0.3 To try the implementation of Hidden Markov

Model for time series analysis
We were successful in implementing Hidden Markov Model for our data-set
and the results that were obtained were noted and plotted.

32
Acknowledgments

We are extremely grateful to our project supervisor Prof. Dr. O.P. Vyas for
his guidance and support at every stage of the project. We are also thank-
ful to all the friends who have supported and helped in every phase of the
project till now.

Thank you.

April 2020
Indian Institute of Information Technology, Allahabad

33
References

[1] Independent Electricity System Operator, accessed February, 2020,

<ieso.ca>.

[2] Canada, Government of, Historical Climate Data, accessed February,

2020, <https://climate.weather.gc.ca/>.

[3] Paras Mandal, et al., A new recursive neural network algorithm to fore-
cast electricity price for PJM day-ahead market, International Journal
of Energy Resource, Vol. 34 Issue 6, May 2010, pages 507-522.

[4] S. Selvin, R. Vinayakumar, E. A. Gopalakrishnan, V. K. Menon and

K. P. Soman, ”Stock price prediction using LSTM, RNN and CNN-
sliding window model,” 2017 International Conference on Advances in
Computing, Communications and Informatics (ICACCI), Udupi, 2017,
pp. 1643-1647.

[5] Z. Chang, Y. Zhang and W. Chen, ”Effective Adam-Optimized LSTM

Neural Network for Electricity Price Forecasting,” 2018 IEEE 9th Inter-
national Conference on Software Engineering and Service Science (IC-
SESS), Beijing, China, 2018, pp. 245-248.

[6] Israt Jahan, Sayeed Sajal, Stock Price Prediction using Recurrent Neural
Network (RNN) Algorithm on Time-Series Data, MIC Symposium-2018,
paper 55.

[7] Houda Bakir, Ghassen Chniti, and Hédi Zaher, E-Commerce Price Fore-
casting Using LSTM Neural Networks, International Journal of Machine
Learning and Computing, Vol. 8, No. 2, April 2018.

[8] Jian Zheng, Cencen Xu, Ziang Zhang and Xiaohua Li, ”Electric load
forecasting in smart grids using Long-Short-Term-Memory based Re-
current Neural Network,” 2017 51st Annual Conference on Information
Sciences and Systems (CISS), Baltimore, MD, 2017, pp. 1-6.

34
 [9] Zhitang Chen, Jiayao Wen and Yanhui Geng, Predicting Future Traffic
using Hidden Markov Models, 2016 IEEE 24th International Confer-
ence on Network Protocols (ICNP) Workshop on Machine Learning in
Computer Networks (NetworkML 2016).

[10] Tie Liu, Application of Markov Chains to Analyze and Predict the Time
Series, Modern Applied Science, Specialized Scientific Research Program
on Scientific Research of high-level talents in Ankang University (Pro-
gram No. AYQDZR200705).

[11] Hippert, H.s Pedreira, Carlos Souza, Reinaldo. (2001). Neural Networks
for Short-Term Load Forecasting: A Review and Evaluation. Power Sys-
tems, IEEE Transactions on. 16. 44 - 55. 10.1109/59.910780.

[12] “Load forecasting using support vector machines: A study on EUNITE

competition 2001”, IEEE Transactions on Power Systems, Vol.19, No.4,
November 2004.

[13] “Short term load forecasting using fuzzy neural networks”, IEEE Trans-
actions on Power Systems, Vol.10, No.3 August 1995.

[14] “Short term load forecasting for fast developing utility using knowledge-
based expert systems”, IEEE Transactions on Power Systems, Vol.17,
No.4, May 2002.

[15] “Short term power system load forecasting using the iteratively
reweighed least squares algorithm”, Electric power system research,
19(1990) pp.11-12

[16] M. A. Abu-El-Magd and N. K. Sinha, ”Short-Term Load Demand Mod-

eling and Forecasting: A Review,” in IEEE Transactions on Systems,
Man, and Cybernetics, vol. 12, no. 3, pp. 370-382, May 1982. doi:
10.1109/TSMC.1982.4308827

[17] G. T. Heinemann, D. A. Nordmian and E. C. Plant, ”The Relationship

Between Summer Weather and Summer Loads - A Regression Analysis,”
in IEEE Transactions on Power Apparatus and Systems, vol. PAS-85,
no. 11, pp. 1144-1154, Nov. 1966.

[18] the pandas development team, ”pandas.DataFrame.corr”, ac-

cessed April 11, 2020, <https://pandas.pydata.org/pandas-
docs/stable/reference/api/pandas.DataFrame.corr.html>

35
[19] L. R. .Rabiner, B. H. Juang, An Introduction to Hidden Markov Models,
1966.

[20] Nguyet Nguyen, Stock Price Prediction using Hidden Markov Model,
March 21, 2016.

[21] Shahidehpour, Hatim Yamin, Zuyi Li, Market Operations in Electric

Power Systems: Forecasting, Scheduling, and Risk Management, Wiley.

[22] Weron, Rafal. (2014). Electricity price forecasting: A review of the state-
of-the-art with a look into the future. International Journal of Forecast-
ing. 30. 10.1016/j.ijforecast.2014.08.008.

36