# Neural Network and Time Series Analysis Approaches in ... Electricity Consumption of Public...

date post

04-Jun-2018Category

## Documents

view

212download

0

Embed Size (px)

### Transcript of Neural Network and Time Series Analysis Approaches in ... Electricity Consumption of Public...

Neural Network and Time Series Analysis Approaches in Predicting

Electricity Consumption of Public Transportation Vehicles

CLAUDIO GUARNACCIA, JOSEPH QUARTIERI, CARMINE TEPEDINO

Department of Industrial Engineering

University of Salerno

Via Giovanni Paolo II 132, Fisciano (SA)

ITALY

cguarnaccia@unisa.it , quartieri@unisa.it , ctepedino@unisa.it

SVETOSLAV ILIEV, SILVIYA POPOVA

Department of Bioengineering and Unique Instruments, Components and Structures

Institute of System Engineering and Robotics

Bulgarian Academy of Sciences

Sofia 1113, Akad. G. Bonchev str. bl 2

BULGARIA

sd_iliev@abv.bg , popova_silvia2000@yahoo.com

Abstract: - Public transportation is a relevant issue to be considered in urban planning and in network design,

thus efficient management of modern electrical transport systems is a very important but difficult task. Tram

and trolley-bus transport in Sofia, Bulgaria, is largely developed. It is one of the largest consumers of electricity

in the city, which makes the question of electricity prediction very important for its operation. In fact, they are

required to notify the energy provider about the expected energy consumption for a given time range.

In this paper, two models are presented and compared in terms of predictive performances and error

distributions: one is based on Artificial Neural Networks (ANN) and the other on Time Series Analysis (TSA)

methods. They will be applied to the energy consumption related to public transportation, observed in Sofia,

during 2011, 2012 and 2013.

The main conclusion will be that the ANN model is much more precise but requires more preliminary

information and computational efforts, while the TSA model, against some errors, shows a low demanding

input entries and a lower power of calculation. In addition, the ANN model has a lower time range of

prediction, since it needs many recent inputs in order to produce the output. On the contrary, the TSA model

prediction, once the model has been calibrated on a certain time range, can be extended at any time period.

Key-Words: - Neural Network, Time Series Analysis, Electricity consumption prediction, Public transportation

1 Introduction In many big European cities, a large amount of

resources is adopted to develop an efficient network

of public transportation. The growing number of

inhabitants in urban areas leads to the necessity to

control the vehicular traffic due to private

transportation. For this reason, electrical tram and

trolley bus are preferred. The reduction of

combustion engines usage allows to reduce physical

and chemical polluting agents in highly populated

areas. Electrical engines are also very quiet, from

the acoustical point of view, and contribute to a

reduction of noise due to vehicular road traffic [1-

13].

Sofia, the capital of Bulgaria, has a very

developed network of electrical public

transportation vehicles. Anyway, the high electricity

absorption must be carefully monitored, both for

cost and electrical network stability reasons.

Different predictive models can be found in

literature, based on various approaches, such as

Neural Networks, Support Vector Machines, Fuzzy

logic, statistical tools, etc. [14-20].

In this paper, the predictive performances of two

different modelling techniques are compared. The

first method is based on an Artificial Neural

Network (ANN) of the multilayer perceptron

typology, thus able to extract the non-linear

relations in a data matrix. The second technique

makes statistical inference using the time periodicity

of the electrical absorption, by means of a model

based on Time Series Analysis (TSA).

WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENTClaudio Guarnaccia, Joseph Quartieri,

Carmine Tepedino, Svetoslav Iliev, Silviya Popova

E-ISSN: 2224-3496 312 Volume 11, 2015

mailto:cguarnaccia@unisa.itmailto:quartieri@unisa.itmailto:ctepedino@unisa.itmailto:sd_iliev@abv.bgmailto:popova_silvia2000@yahoo.com

After having presented the models, they will be

tested on 4 different datasets, that are four months of

2013. The differences between results obtained with

the two models will be highlighted in terms of error

evaluation and analysis.

The ANN model ensures a more accurate mean

prediction, but it needs more input information,

higher elaboration and computation abilities and

input data measured closer to the periods that are

under prediction. The TSA model is slightly less

precise in the prediction but needs as input only the

energy consumption registered in a sufficient

number of previous time periods. In addition, the

TSA model requires a low computing power and it

is able to provide reliable predictions even in time

periods far from the data used in the calibration and

in the parameters evaluation.

2 Models Presentation In this section, the ANN and TSA models will be

shortly presented and discussed.

The dataset is related to energy consumption in

2011, 2012 and 2013. The first two years (2011 and

2012) are used for training and calibration of the

models, while some intervals of 2013 (January,

May, July and November) are used for testing, i.e.

comparison between real and predicted data.

2.1 Artificial Neural Network model Artificial neural networks (ANN) have been applied

successfully to a large number of engineering

problems. The great advantage of ANN is that they

impose less restrictive requirements with respect to

the available information about the character of the

relationships between the processed data, the

functional models, the type of distribution, etc. They

provide a rich, powerful and robust non-parametric

modelling framework with proven efficiency and

potential for applications in many fields of science.

The advantages of ANN encouraged many

researchers to use these models in a broad spectrum

of real-world applications. In some cases, the ANNs

are a better alternative, either substitutive or

complementary, to the traditional computational

schemes for solving many engineering problems.

The approach based on ANN has some significant

advantages over conventional methods, such as

adaptive learning and nonlinear mapping.

In many engineering and scientific applications a

system having an unknown structure has measurable

or observable input or output signals. Neural

networks have been the most widely applied for

modelling of systems [14, 21-27]. Artificial neural

networks, coupled with an appropriate learning

algorithm, have been used to learn complex

relationships from a set of associated input-output

vectors.

There are four reasons for using neural network

for electricity consumption prediction in tram and

trolleybus transport:

1. The dependence between input and output data is nonlinear and the neural networks have ability to

model non-linear patterns.

2. The neural network learns the main characteristics of a system through an iterative

training process. It can also automatically update

its learned knowledge on-line over time. This

automatic learning facility makes a neural

network based system inherently adaptive.

3. ANN can be more reliable at predicting. It is well-known that forecasting techniques based on

artificial neural networks are appropriate means

for prediction from previously gathered data. The

neural networks make possible to define the

relation (linear or nonlinear) among a number of

variables without their appropriate knowledge.

4. There is a big number of data available. The neural network, trained with these data, adjusts

the weights and predicts output with small error

when working on new data with the same or

similar characteristics of the input data.

2.1.1 ANN model details

Two-layer network with error back propagation

training algorithm is used to predict electricity

consumption. The network has one hidden layer

with forty-three neurons and an output layer with

one neuron. The sigmoid tansig transfer function is

used for the hidden layer and for the output layer the

activation function is the linear function purelin. Six

input factors: mileage, air temperature, time of day,

weekday or holiday, month, schedule (summer /

winter).

Training data for 2011 and 2012 years with a

total of 17496 items were used. The best result in

the training of the network is achieved after 158

iterations, as mean square error (performance) is

0,0776 .

In Fig. 1 the multiple-correlation coefficients and

comparison between linear regression and ANN for

training, validation and testing are shown, while in

Fig. 2, the error histogram in the complete training

process is reported.

WSEAS TRANSACTIONS on ENVIRONMENT and DEVELOPMENTClaudio Guarnaccia, Joseph Quartieri,

Carmine Tepedino, Svetoslav Iliev, Silviya Popova

E-ISSN: 2224-3496 313 Volume 11, 2015

Fig. 1: Comparison between linear regression and ANN model

results plotted versus the observed values for training,

validation and testin

Recommended

*View more*