intelligent prediction system for gas metering system using particle swarm optimization in training neural network

doi: 10.1016/j.procs.2017.01.197 ScienceDirect 2016 IEEE International Symposium on Robotics and Intelligent Sensors, IRIS 2016, 17-20 December 2016, Tokyo, Japan Intelligent Prediction

Procedia Computer Science 105 ( 2017 ) 165 – 169

1877-0509 © 2017 The Authors Published by Elsevier B.V This is an open access article under the CC BY-NC-ND license

(http://creativecommons.org/licenses/by-nc-nd/4.0/)

Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors(IRIS 2016) doi: 10.1016/j.procs.2017.01.197

ScienceDirect

2016 IEEE International Symposium on Robotics and Intelligent Sensors, IRIS 2016, 17-20 December 2016,

Tokyo, Japan Intelligent Prediction System for Gas Metering System Using Particle

Swarm Optimization in Training Neural Network

N.S Roslia, R Ibrahimb, I.Ismailc *

abc Universiti Teknologi PETRONAS, Bandar Seri Iskandar, 31750 Tronoh, Perak

Abstract

In this paper, a study on development of prediction model based on an intelligent systems is discussed for gas metering system in order to validate the instrument reliability In providing reliable measurement of gas metering system, an accurate prediction model is required for model validation and parameter estimation The intelligent prediction system has been developed for gas measurement validation Then the project focused on the application of particle swarm optimization (PSO) and Genetic Algorithm (GA) in training neural network prediction model in enhancing the performance of Intelligent Prediction System (IPS) In this study, the three experiment has been conducted to improve the accuracy of the neural network prediction model The comparison of the performance of PSONN and GANN with pure ANN is presented in this paper The results shows that the proposed PSONN model give promising results in the prediction accuracy of gas measurement

© 2016 The Authors Published by Elsevier B.V

Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors (IRIS 2016)

Keywords: neural network; particle swarm optimization; genetic algorithm; prediction; gas metering system

1 Introduction

The increasing energy demand in nowadays modern innovation become industrial concern towards energy efficiency for energy saving It is a vital for the accuracy and reliable metering system in oil and gas industry to maintain the billing purposes This is because of inaccuracy of product selling to the client will bring about lost income to the organization A slight error in the bill calculation will lead to huge financial impact Between validation periods, the field device might be drifted and flow computer giving unknown readings and other scenarios that might lead to billing issues Measurement readings from billing equipment will sometimes freeze, overshoot and even zero readings Due to reliability concern, smart meters applying artificial intelligence is one

of the future technologies that can be genuine global solution1 Therefore, it is an extra effort to develop a monitoring system tool

to verify the billing data generated by measuring equipment and subsequently will enhance overall billing integrity To achieve the objective, a prevailing tool that can be used is the hybrid method which neural network are combined with PSO and GA algorithm The hybrid methods are mostly outperformed than non-hybrid methods2 It can predict the performance and accurately forecast the instrument measurement and in addition provide a reliable metering system for billing integrity

* Corresponding author Tel.: +0-000-000-0000 ; fax: +0-000-000-0000

E-mail address: author@institute.xxx

© 2017 The Authors Published by Elsevier B.V This is an open access article under the CC BY-NC-ND license

(http://creativecommons.org/licenses/by-nc-nd/4.0/)

Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent

Sensors(IRIS 2016)

Nomenclature

ANN Artificial Neural Network

PSO Particle Swarm Optimization

GA Genetic Algorithm

IPS Intelligent Prediction System

1.1 Gas Metering System

Accurate measurement of gas flow through pipelines is vital to reduce energy loss One of the greatest concerns is whether the amount of money buyer paid is justified with the amount of products sold The metering system required critical examination which result in cost reduction The measurements of gas metering station along the pipeline system are used to compute energy supplied for customer which consists of measuring equipment (e.g pressure and temperature transmitter), turbine meter, flow computer and gas chromatography These measurement will be utilized for energy consumption calculation for billing purposes Fig.1 shows the billing process from supplier to customer

Fig 1 Billing process of gas consumption The metering system do not have any reference system to validate its accurateness which also is defined as standalone system Sometimes, instruments fault may occurred that lead to billing issue Prediction and analysis on what will happen plays important role in economic operation Thus, accurate and robust prediction model can significantly improve the billing results Reliable billing verification tool also plays a significant role to increase work efficiency of billing-correction To manage this world issue, the human and computational intelligence must be developed for achieving high accuracy of prediction model of process parameter prediction

1.2 Intelligent Prediction System

The implementation of the intelligent system varies in prediction, classification, clustering and pattern recognition3 A data driven technique based on historical data can be used to design reliable prediction model For example, nonlinear estimation of wind power optimization is done by ANN4 It used global optimization based on ANN Therefore, Artificial Neural Network (ANN) method is proposed to be the intelligent prediction model to learn the behavior of fault and providing reliable data for billing purposes There are several factors that affect the model development: model inputs; data pre-processing; learning algorithm and activation function5 Selection of ANN architecture also describes the model performance in view of its robustness and reliability of the system The most popular training algorithm in prediction application is backpropagation method (BP) The examples of BP algorithm are gradient descent, Lavenberg-Marquardt, Lavenberg-Marquardt with Bayesian regulation6 In this paper proposed the optimization method to be applied in training ANN to have better prediction value The approaches of hybrid intelligence is used to predict more accurate and reliable energy7 Therefore, PSO and GA is introduced in optimizing the weights

and biases from ANN training

1.3 Particle Swarm Optimization

PSO was first established by Kennedy and Eberhart as a solution to the complex non-linear optimization problem by imitating the behavior of bird flocks in the concept of function-optimization by means of a particle swarm8 PSO is also well-known as population-based search method based on the behavior of elements in nature such as fish schooling and birds flocking The population follows its leader which affected by the best-positon of each particle in the whole swarm This phenomenon is called

the global best PSO (or gbest PSO)9 Every individual will adjust its position according its own best position which follows toward

group’s objective The particle is called as personal best (or pbest PSO) A local best PSO (or lbest PSO) occurred when the pbest

corresponds to the position in neighbour’s experience Lastly, all the particles will finally move toward the desired location To achieve this condition, the velocity of the particle is updated as in equation (1) where the particles move to a new position close to

the object from initialization through time t the new position of the ith particle in a d-dimensional search space is represented by

xi = (xi1, xi2, …, xid) and determined by its velocity, vi as in equation (2)

ݒ௜ௗሺ௧ାଵሻൌ ݓݒ௜ௗሺ௧ሻ൅ ܿଵݎଵሺ݌௜ௗ൅ ݔ௜ௗሺ௧ሻሻ ൅ ܿଶݎଶሺ݌௚ௗ൅ ݔ௜ௗሺ௧ሻሻ (1)

ݔ௜ௗሺ௧ାଵሻൌ  ݔ௜ௗሺ௧ሻ൅ ݒ௜ௗሺ௧ାଵሻݐ (2) Moreover, PSO is also effected by weight, velocity constriction and clamping The inertia weight is introduced to reduce the whole position10 Based on the research on the inertia weight range, [0.5, 0.9] values reduce the amplitude of trajectories thus allowing exploration to convergence triangle11 The pbest term is known cognitive component which the particles learn from its

previous performance This element looks like an individual learning of its best particle position12 A larger size of swarm need to

be covered per iteration when larger parts of the search space In contrary, when the number of particles increase, the calculation complexity of each loop is increase hence more time-consuming6,13,14 The acceleration coefficients C1 and C2 were performed to analyse the particles in the swarm It represents the weight of the stochastic that attract the particles pbest and gbest16 PSO algorithm can optimized all parameters include the weights to perform better fine-tuning process15,17

1.4 Genetic Algorithm

John Holland introduced Genetic Algorithm (GA) by probabilistic optimization technique The first thought originated from biological development process in chromosomes The best arrangements of GA are recombined with each other to shape new structured for the survival of the fittest GA techniques usually involved of coding the problems, generating initial population, evaluating fitness, crossover, mutation and selection18 The population is a group of individual number which encoded in bit string

of fixed length Every individual interact with the other chromosome of a living thing There is a fitness function that evaluate the desired requirements of every chromosome from the population Combining of selected individuals during crossover will create more individuals which inherit the best traits while the two chromosomes are consolidated in mutation process will make a new individuals The process is repeated until the termination condition is met

2 Methodology

2.1 Development of IPS Framework

Based on the research application in flow measurement19,20,21, it depicts that neural networks is one of reliable way to improve energy efficiency Fig 2 shows the framework of proposed prediction model in order to get the best performance of IPS

Determine NN and PSO parameter Start training Determine NN structure

Adjust weight according gbest particle

Initialize NN and GA parameter Start GA model Determine NN structure

Training new parameters by NN

output

Implementation of IPS Performance analysis and comparison

Fig 2 Intelligent Prediction System (PPS) Development 2.2 Development of ANN Architecture

In order to obtain the optimum parameter for neural network model, the architecture of ANN prediction model is required to be investigated Neural network needs to be trained based on good historical data from the well-functioned instrument The total data available is 4560 used for training the ANN model In this investigation, the best combination of these parameters are selected based on the least root means square error (RMSE) From the investigation, ANN model was trained by the Lavenberg-Marquardt algorithm as the best parameter of the neural network model that gives the more promising result as compared to other learning

algorithms The optimal number of neurons is 10 with 2 layers of hidden layers These parameters are well investigated before

proceeding with the PSO and GA program

2.3 Parameter of particle swarm optimization

Proposed prediction model based on PSO algorithm is implemented to have better prediction result To define the best model, PSO parameters such as swarm size and the coefficient of velocity equation are investigated Then the value of particle’s position and velocity are updated based on the best fitness values The fitness evaluation is calculated based on mean square error (MSE)

in equation (3)

ܯܵܧ ൌ ଵ௡σ௡ ሺݕ௜െ ݕ෤௜ሻଶ

The searching process is continued until the stopping criteria is met in order to gain the best position of pbest and gbest The accuracy and robustness of IPS model are analyzed and tested based on predicted value provide by the network In summarize, the parameters that have been finalized for PSONN are shown in Table 1

Table 1 Parameters of PSONN

Acceleration Coefficients (C 1 and C 2 ) 2.0

Initial weight, position and velocity Random

2.4 Parameter of genetic algorithm

GA is typically connected in ANN to advance the system due to its effectiveness in providing the best parameters, for example, learning process do maintain a strategic distance from being caught in a local minima and speed up the convergence rate Besides that, GA has been utilized to create the best NN weight optimization and design The generation is prepared for evaluation and the procedure proceeds until the best performance is achieved GA is ended up being successful to direct the ANN adapting, thus, it is generally utilized as a part of numerous applications Subsequently, PSO performance and outcome are defined to be compared with GANN Table 2 demonstrates the optimal value of GA parameters that have been investigated in order to provide the best result in NN learning

Table 2 Parameters of GANN

3 Results and Discussion

The prediction performance of ANN, PSO-based ANN and GA-based ANN models were examined for comparison purposes The result for training and testing the prediction model shows in the Table 3 The performance of PSO and GA algorithm has been compared with ANN model only The accuracy of the model performance are analysed by the root mean square error (RMSE) and percentage error between predicted and actual value

Table 3: Comparison between conventional ANN with PSO-based ANN model

Parameter ANN

(RMSE)

PSONN (RMSE)

GANN (RMSE)

ANN (Percentage Error)

PSONN (Percentage Error)

GANN (Percentage Error)

Results in Table 3 shows that PSONN mostly gives better performance compared to ANN and GANN with the least RMSE produced compared to GANN and ANN In PSONN, selection of network parameters for the prediction model give impact to the performance of the system Adjustment of parameter can be done as well as to achieve better enhancement While the weight of

GA gives the basic of standard LM learning during the learning rate of GANN This is to ensure faster convergence time and better

performance results In any case, the selection process of GA parameter takes longer time compared with PSONN Therefore, the best parameters of PSONN will be used to develop the IPS

4 Conclusion

In conclusion, PSO and GA are effectively optimized the neural network prediction model They have been tested using data from gas metering system The investigation is done by comparing accuracy result produced by hybrid learning method of PSO and GA with ANN model The most critical knowledge in this study is the proposed PSO is a straightforward optimization technique with less computation that can be implemented in neural network with high accuracy compared to GANN While the implementation of PSO-based ANN model in the metering instrument is applicable for the user to perform the process parameter prediction and calculation using IPS For the future works, IPS can be developed for other application in the industry with the expert integration system

Acknowledgements

This research is funded and supported by Universiti Teknologi PETRONAS

References

1 Lodder AR and Wisman T Artificial Intelligence Techniques and the Smart Grid: Towards Smart Meter Convenience While Maintaining Privacy Journal of

Internet Law (Dec 2015), 2016 19(6): p 20-27

2 Okumus I and Dinler A Current status of wind energy forecasting and a hybrid method for hourly predictions Energy Conversion and Management, 2016

123: p 362-371

3 Sridevi S, Rajaram S, and Swadhikar C An intelligent prediction system for time series data using periodic pattern mining in temporal databases, in Proceedings

of the First International Conference on Intelligent Interactive Technologies and Multimedia 2010, ACM: Allahabad, India p 163-171

4 Saravanakumar R and Jena D Nonlinear control of a wind turbine based on nonlinear estimation techniques for maximum power extraction International

Journal of Green Energy, 2016 13(3): p 309-319

5 Ismail MJ and Ibrahim R Selection of network architecture and input sensitivity analysis for a Neural Network Energy Prediction Model in Intelligent and

Advanced Systems (ICIAS), 2010 International Conference on 2010

6 Vilovic I, Burum N, and Milic D Using particle swarm optimization in training neural network for indoor field strength prediction in ELMAR, 2009

ELMAR'09 International Symposium 2009 IEEE

7 Zahraee SM, Khalaji Assadi M, and Saidur R Application of Artificial Intelligence Methods for Hybrid Energy System Optimization Renewable and

Sustainable Energy Reviews, 2016 66: p 617-630

8 Eberhart RC and Yuhui S Particle swarm optimization: developments, applications and resources in Evolutionary Computation, 2001 Proceedings of the

2001 Congress on 2001

9 Zaji AH and Bonakdari H Performance evaluation of two different neural network and particle swarm optimization methods for prediction of discharge capacity

of modified triangular side weirs Flow Measurement and Instrumentation, 2014 40(0): p 149-156

10.Bonyadi M and Michalewicz Z A locally convergent rotationally invariant particle swarm optimization algorithm Swarm Intelligence, 2014 8(3): p 159-198

11.Fernández-Martínez J, García-Gonzalo E, and Fernández-Alvarez J Theoretical analysis of particle swarm trajectories through a mechanical analogy

International Journal of Computational Intelligence Research, 2008 4(2)

12.Emery NJ Cognitive ornithology: The evolution of avian intelligence Philosophical Transactions of the Royal Society of London Series B: Biological Sciences,

2006 361(1465): p 23-43

13.Pulido M, Melin P, and Castillo O Particle swarm optimization of ensemble neural networks with fuzzy aggregation for time series prediction of the Mexican

Stock Exchange Information Sciences, 2014 280: p 188-204

14.Momeni E, et al Prediction of uniaxial compressive strength of rock samples using hybrid particle swarm optimization-based artificial neural networks

Measurement: Journal of the International Measurement Confederation, 2015 60: p 50-63

15.Pengajian, S., Judul: Particle Swarm Optimization For Neural Network Learning Enhancement 2006, Universiti Teknologi Malaysia

16.Eberhart RC, and Shi Y Particle swarm optimization: developments, applications and resources in Evolutionary Computation, 2001 Proceedings of the 2001

Congress on 2001 IEEE

17.Das G, Pattnaik PK, and Padhy SK Artificial Neural Network trained by Particle Swarm Optimization for non-linear channel equalization Expert Systems with

Applications, 2014 41(7): p 3491-3496

18.Samira G, Jarek K, Yousef M A hybrid Genetic Algorithm and Monte Carlo simulation approach to predict hourly energy consumption and generation by a

cluster of Net Zero Energy Buildings Applied Energy 2016 179:p 626-637

19.Borg D, Suetake M, and Brandão D A neural network developed in a Foundation Fieldbus environment to calculate flow rates for compressible fluid Flow

Measurement and Instrumentation, 2014 40(0): p 142-148

20.Dragoi EN, Curteanu S, and Fissore D On the Use of Artificial Neural Networks to Monitor a Pharmaceutical Freeze-Drying Process Drying Technology,

2013 31(1): p 72-81

21.Liu RP, et al A neural network to correct mass flow errors caused by two-phase flow in a digital coriolis mass