How to Submit Proof Corrections Using Adobe ...

How to Submit Proof Corrections Using Adobe Reader Using Adobe Reader is the easiest way to submit your proposed amendments for your IGI Global proof. If you don’t have Adobe Reader, you can download it for free at http://get.adobe.com/reader/. The comment functionality makes it simple for you, the contributor, to mark up the PDF. It also makes it simple for the IGI Global staff to understand exactly what you are requesting to ensure the most flawless end result possible. Please note, however, that at this point in the process the only things you should be checking for are: Spelling of Names and Affiliations, Accuracy of Chapter Titles and Subtitles, Figure/Table Accuracy, Minor Spelling Errors/Typos, Equation Display As chapters should have been professionally copy edited and submitted in their final form, please remember that no major changes to the text can be made at this stage. Here is a quick step-by-step guide on using the comment functionality in Adobe Reader to submit your changes. 1.

Select the Comment bar at the top of page to View or Add Comments. This will open the Annotations toolbar.

2.

To note text that needs to be altered, like a subtitle or your affiliation, you may use the Highlight Text tool. Once the text is highlighted, right-click on the highlighted text and add your comment. Please be specific, and include what the text currently says and what you would like it to be changed to.

3.

If you would like text inserted, like a missing coma or punctuation mark, please use the Insert Text at Cursor tool. Please make sure to include exactly what you want inserted in the comment box.

4.

If you would like text removed, such as an erroneous duplicate word or punctuation mark, please use the Add Note to Replace Text tool and state specifically what you would like removed.

IJMSTR Editorial Board Editor-in-Chief:

Ming Yang, Southern Polytechnic State U., USA Nikolaos Bourbakis, Wright State U., USA Konstantina S. Nikita, National Technical U. of Athens, Greece

Associate Editors:

J.J. P. Tsai, Asia U. Taiwan John Volakis, The Ohio State U., USA

International Editorial Review Board: Rudiger Brause, Goethe U. Frankfurt, Germany Dimitrios Fotiadis, U. of Ioannina, Greece Tatjana Jevremovic, U. of Utah, USA Zhanpeng Jin, Binghamton U., USA

Lee Potter, The Ohio State U., USA George Tsichritzis, U. of Pireas, Greece Lefteris Tsoukalas, Purdue U., USA Maria Virvou, U. of Pireas, Greece Michalis Zervakis, Technical U. of Crete, Greece

IGI Editorial: Lindsay Johnston, Managing Director Jeff Snyder, Copy Editor Christina Henning, Production Editor Sean Eckman, Development Editor Austen DeMarco, Managing Editor Henry Ulrich, Production Assistant

International Journal of Monitoring and Surveillance Technologies Research April-June 2014, Vol. 2, No. 2

Table of Contents Research Articles 1

Creep Rupture Forecasting: A Machine Learning Approach to Useful Life Estimation Stylianos Chatzidakis, School of Nuclear Engineering, Purdue University, West Lafayette, IN, USA Miltiadis Alamaniotis, School of Nuclear Engineering, Purdue University, West Lafayette, IN, USA Lefteri H. Tsoukalas, School of Nuclear Engineering, Purdue University, West Lafayette, IN, USA

26

Fuzzy Integration of Support Vector Regression Models for Anticipatory Control of Complex Energy Systems Miltiadis Alamaniotis, School of Nuclear Engineering, Purdue University, West Lafayette, IN, USA Vivek Agarwal, Department of Human Factors, Controls and Statistics, Idaho National Laboratory, Idaho Falls, ID, USA

41

Planning and Management of Distributed Energy Resources and Loads in a Smart Microgrid Federico Delfino, Department of Naval, Electrical, Electronic & Telecommunication Engineering, University of Genoa, Genoa, Italy Mansueto Rossi, Department of Naval, Electrical, Electronic & Telecommunication Engineering, University of Genoa, Italy Luca Barillari, Department of Naval, Electrical, Electronic & Telecommunication Engineering, University of Genoa, Italy Fabio Pampararo, Department of Naval, Electrical, Electronic & Telecommunication Engineering, University of Genoa, Italy Paolo Molfino, Department of Naval, Electrical, Electronic & Telecommunication Engineering, University of Genoa, Italy Alireza Zakariazadeh, Electrical Engineering Department Iran University of Science and Technology, Tehran, Iran

58

Design and Early Simulations of Next Generation Intelligent Energy Systems Rafik Fainti, Department of Electrical and Computer Engineering, University of Thessaly, Volos, Greece Antonia Nasiakou, Department of Electrical and Computer Engineering, University of Thessaly, Volos, Greece Eleftherios Tsoukalas, Department of Electrical and Computer Engineering, University of Thessaly, Volos, Greece Manolis Vavalis, Department of Electrical and Computer Engineering, University of Thessaly, Volos, Greece

Copyright

The International Journal of Monitoring and Surveillance Technologies Research (IJMSTR) (ISSN 2166-7241; eISSN 2166-725X), Copyright © 2014 IGI Global. All rights, including translation into other languages reserved by the publisher. No part of this journal may be reproduced or used in any form or by any means without written permission from the publisher, except for noncommercial, educational use including classroom teaching purposes. Product or company names used in this journal are for identification purposes only. Inclusion of the names of the products or companies does not indicate a claim of ownership by IGI Global of the trademark or registered trademark. The views expressed in this journal are those of the authors but not necessarily of IGI Global. The International Journal of Monitoring and Surveillance Technologies Research is indexed or listed in the following: Cabell’s Directories; INSPEC; MediaFinder; The Standard Periodical Directory; Ulrich’s Periodicals Directory

26 International Journal of Monitoring and Surveillance Technologies Research, 2(2), 26-40, April-June 2014

Fuzzy Integration of Support Vector Regression Models for Anticipatory Control of Complex Energy Systems Miltiadis Alamaniotis, School of Nuclear Engineering, Purdue University, West Lafayette, IN, USA Vivek Agarwal, Department of Human Factors, Controls and Statistics, Idaho National Laboratory, Idaho Falls, ID, USA

ABSTRACT Anticipatory control systems are a class of systems whose decisions are based on predictions for the future state of the system under monitoring. Anticipation denotes intelligence and is an inherent property of humans that make decisions by projecting in future. Likewise, intelligent systems equipped with predictive functions may be utilized for anticipating future states of complex systems, and therefore facilitate automated control decisions. Anticipatory control of complex energy systems is paramount to their normal and safe operation. In this paper a new intelligent methodology integrating fuzzy inference with support vector regression is introduced. The proposed methodology implements an anticipatory system aiming at controlling energy systems in a robust way. Initially, a set of support vector regressors is adopted for making predictions over critical system parameters. The predicted values are used as input to a two-stage fuzzy inference system that makes decisions regarding the state of the energy system. The inference system integrates the individual predictions at its first stage, and outputs a decision together with a certainty factor computed at its second stage. The certainty factor is an index of the significance of the decision. The proposed anticipatory control system is tested on a real-world set of data obtained from a complex energy system, describing the degradation of a turbine. Results exhibit the robustness of the proposed system in controlling complex energy systems. Keywords:

Anticipatory Control, Energy Systems, Fuzzy Inference, Support Vector Regression

INTRODUCTION Energy systems are comprised of several smaller and simpler components operating in a synergistic way for energy generation, delivery, and distribution. The plethora of components, together with their interactions, significantly

increases the complexity of those systems making their control rather challenging task. More particularly, any control decision is taken only after several factors are taken into consideration. However, not all factors or operational parameters are well known at each instance;

DOI: 10.4018/ijmstr.2014040102 Copyright © 2014, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.

International Journal of Monitoring and Surveillance Technologies Research, 2(2), 26-40, April-June 2014 27

hence, control decisions are made under some uncertainty. Anticipatory decision systems are a class of systems that make decisions taking into account both the current and future states (Tsoukalas, 1998). While the current state contains the most recent measured values of the system under monitoring, anticipation of future states is performed by analytical prediction models (Agarwal et al., 2013). Analytical models are utilized to make predictions regarding the system’s own state variables and/or its environment. In other words, anticipatory systems mimic the way humans make their decisions, observe the current situation, and foresee possible future situations that may be attained by specific actions (Rosen, 2012). Intelligent control focuses on realizing complex automatic control systems by utilizing functions, such as solution comparison, solution search using a heuristic criterion, prediction making, and generally by employing functions and tools from the artificial intelligence realm (Tsoukalas & Uhrig, 1996; Harris et al., 1993). One of the crucial factors in human thinking is the ability to learn how to make predictions over the output of actions. However, learning how to predict requires prior observation of the same or similar phenomenon. By observing, humans develop inherent ways of how to think and anticipate the possible output following specific actions. Thus, anticipation is a crucial factor for implementing intelligent control systems (Uhrig et al., 1994). The majority of proposed anticipatory systems to implement intelligent control employ tools from artificial intelligence and statistics. Ikonomopoulos et al. (1993) proposed an anticipatory system based on the integration of neural networks and fuzzy logic for measuring operational parameters in complex systems, while Xinging et al. (1996) implemented a neurofuzzy system for complex system control. Uhrig and Tsoukalas (2003) applied anticipatory control for complex nuclear systems by using intelligent multi-agent systems. A hybrid expert system-neural network methodology was presented by Tsoukalas and Reyes-Jimenez

(1990), and a possibilistic-probabilistic formalism by Tsoukalas (1989). Furthermore, the use of dynamic control equations together with data mining tools was presented for controlling complex power systems like wind turbines in (Kusiak et al. 2009). Further, fuzzy logic rules were used for anticipation in Chiu et al.’s article (1991), while a distributed intelligence approach for the anticipation of transportation infrastructures is proposed by Van Dam et al. (2004). The above examples strictly adopt a single model for making predictions; therefore, the examples exhibit significant limitations in prediction accuracy. This paper presents a new methodology for implementing anticipatory control systems. The proposed methodology combines support vector regression, SVR, (Bishop, 2006) and fuzzy logic (Ross, 2013). More particularly a set of SVR models is adopted for prediction making (Alamaniotis et al., 2012), while fuzzy logic is employed for fusing the individual SVR predictions and decision making. The proposed system is applied for making control decisions in complex energy systems (Alamaniotis et al., July 2014). In the current manuscript, the anticipatory control system is adopted for controlling the maintenance actions of a turbine, which is the basic component in all energy generation systems. The remainder of the paper is organized as follows: in the next two sections, support vector regression and fuzzy logic are briefly introduced, and the proposed anticipatory system is described. The results of the proposed system on making control decision regarding degradation of a turbine are given in the results section. Finally conclusions are drawn based on main contribution of the paper.

BACKGROUND Support Vector Regression Kernel machines are a class of computational tools that belongs to the machine learning realm. One popular paradigm among kernel machines is the support vector machine (SVM), which is

Copyright © 2014, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.


used for classification and regression problems (Bishop, 2006). In the latter case, the SVM is called support vector regression (Basak, Pal, & Patranabis, 2007). SVM has been widely used in various applications because of its efficiency and simplicity in implementation. In the current manuscript, SVM is used for prediction making in continuous space (i.e., SVR). Derivation of the framework for SVR starts from the simple linear regression (Bishop, 2006). Simple linear regression aims at minimizing a regularized error function, by utilizing a set of historical datasets called the training datasets. Though in simple regression all the available points participate in obtaining the regression line, that is not the case for SVR. More particularly, SVR provides a regression curve by taking into account the data points that lie outside the ε-sensitive band (Vapnik et al., 1997). These points are used for minimizing the regularized error function and are known as support vectors (Bishop, 2006), while data points inside the ε-sensitive band are totally ignored. To make it clearer, SVR curve is obtained by utilizing only a subset of the available training data points (i.e., the support vectors). Analytically, the ε-sensitive band is expressed as:

Eε ( y ( x) − t ) = 0, if | y ( x) − t |< ε      E ( y ( x), t ), if | y ( x) − t |≥ ε 

(1)

where E(y(x),t) is the regularized error function, y(x) is the SVR curve, and t is the known target data point (vector in multidimensional problems). In Equation 1, observe that the ε-sensitive error function assigns a zero error when the absolute difference between the regression curve y(x) and actual value t is between the band determined by [y(x)-ε, y(x)+ε]. In general, the regularized error function for which SVR curve is obtained using a N long training dataset is given below (Bishop, 2006): N

C ∑ E ε (y(x i ) − ti ) + (1 / 2) w i =1

2

(2)

where C is a regularization parameter, Eε() is obtained by Equation 1, y(x) is the predicted value for an input xi, and ti stands for the respective actual value. A general illustration of the SVR is visualized in Figure 1 where the support vectors and the ε-sensitive band are marked. One approach in SVR implementation is the vSVR, where ν denotes a fixed minimum portion of data points that remains outside the ε-insensitive band. Additionally, by introducing Lagrange multipliers for each support vector, SVR reduces to an optimization problem. Solution of the optimization problem provides the following equations: N

y ( x ) = ∑ ( a j − b j ) k ( x, x j ) + β

(3)

j =1

N

β = ti − ε − ∑ (al − bl )k (x i , xl )

(4)

l =1

where α and b are the Lagrange multipliers while k(x,x) is a kernel function (Cristianini & Shawe-Taylor, 2000). Kernel function is any valid analytical function f(x) that may be written in the so called dual form: T

k ( x1 , x2 ) = f ( x1 ) f ( x2 )

(5)

where T denotes the transpose of the function f(). Equations 3, 4, and 5 exhibit the dependence of the SVR formulation on selection of the kernel function. More details on SVR implementation can be found in Bishop (2006), Vapnik (1995), and Shawe-Taylor, & Christanini (2000).

Elements of Fuzzy Logic Fuzzy logic belongs to the wider area of artificial intelligence and mimics the way humans think and make inferences. It facilitates modeling of complex processes without explicitly taking into account all the physical phenomena involved. As stated in (Tsoukalas & Uhrig, 1997) “fuzzy



Figure 1. Visualization of support vector regression framework

logic systems address the imprecision of the input and output variables directly by defining them with fuzzy numbers that can be expressed in linguistic terms.” The main principle of fuzzy logic is the use of fuzzy sets. Unlike classical set theory in which an object either belongs to a set or not, in fuzzy set theory an object may belong to various sets with a different degree of membership. The degree of membership takes values in the interval [0, 1] where a value of zero denotes absolute certainty that the object does not belong to the set while a value of one denotes absolute certainty that the object does belong to the set. The set of all objects is called the universe of discourse (Tsoukalas & Uhrig, 1997). A fuzzy set A is modeled with a membership function μΑ(x):

A = {(x,µA(x ))}, x ∈ X

(6)

where A is the linguistic term of the set, x is an object in the universe of discourse and μΑ(x) is the degree of membership of x in set A as computed by the membership function. An example of fuzzy representation of a variable Temperature is given in Figure 2. The universe of discourse includes the range of temperatures from 0 to 100°C, while the variable temperature is modeled with three Gaussian shaped fuzzy sets: low, medium, and high. One of the advantages of fuzzy logic is the use of “IF…THEN” rules that express the association between fuzzy sets between two or more variables. Such rules allow the modeling of correlated knowledge expressed in the form linguistic terms as used in everyday life. The general form of fuzzy rules is given below: a. IF variable(1) is SET(1), …AND/OR variable(n) is SET(n), THEN variable(m) is SET(m).



Figure 2. Example of fuzzy set representation of the variable temperature

where SET(x) denotes a fuzzy set. In addition, the use of the set operators AND/OR aims at combining various conditions.

ANTICIPATORY CONTROL SYSTEM In this section, the anticipatory control system for complex energy systems is presented. The proposed system includes anticipation and decision making through learning by combining historical datasets and intelligent tools. More particularly, a set of SVR models equipped with a kernel function is used synergistically with fuzzy logic. In general, the proposed anticipatory system can be characterized as a two part system: the individual SVR models prediction values (first part – SVR part), and the fusion of predictions and decision making part by means of fuzzy logic (second part – fuzzy part). The block diagram of the anticipatory control system is presented in Figure 3, where its two parts as well as their individual blocks are explicitly shown.

Before any operation takes place, a set of system decision variables should be identified; the number of variables is equal to M. The decision variables consist of the entries of the decision vector (i.e., X1, X 2 ..., X M  ) of length M that adequately supports decision making about the complex system operation at each instance. Once decision variables have been identified, a group of #M SVR models is employed (i.e., one SVR model for each decision variable). The SVR models are equipped with the kernel function whose analytical form is given below:

(

k ( x1 , x2 ) = exp − x1 − x2

2

)

/ 2σ 2 (7)

where σ 2 is a parameter evaluated in the training phase. The kernel function in Equation 7 is known as the Gaussian kernel function (Bishop, 2006; Rasmussen, 2006).



Figure 3. Block diagram of the anticipatory control decision system

In machine learning parlance, training is the process during which the model is exposed to known inputs and outputs and its parameters are tuned with respect to prediction error. Therefore, historical or simulated datasets for the decision variables are required for training the SVR models. Training is two-fold: evaluation of the Gaussian kernel parameters, and selecting the support vectors for obtaining the SVR curve. Hence, each SVR model is trained using the respective training datasets. Once the training phase is over, the models are utilized to predict the next-time-step values for each decision variable. It should be noted that the prediction horizon (i.e., the number of time steps) is deter-

mined by the modeler and depends on the details of the problem at hand. However, for energy systems an one-step-ahead prediction horizon may be employed (Alamaniotis et al., 2011). Once the individual predictions are collected, they are put together to consist the predicted decision vector, which is also shown in Figure 3. Regarding the entries of the predicted decision vector, the proposed anticipatory system implements a fusion of predicted values by utilizing recently arrived measurements with the respective previously predicted values. More particularly, a new measurement is obtained (current decision vector in Figure 3), and then compared to the respective SVR predicted



Figure 4. Fuzzy sets representation of the absolute error variable

value. Comparison is expressed as the absolute difference between the measurement and the prediction: Absolute _ Error = Rt − Pt

(8)

with R being the measured and P the predicted value respectively at time step t. Equation 8 is applied to all decision variables (i.e., entries of decision vector in Figure 3); which provides M absolute error values. Next, the M absolute errors are forwarded to a fuzzy inference system to indicate the quality of predictions obtained by each SVR model. More particularly the inference system gets as an input the absolute error and provides as an output a quality index in the interval [0, 1], where zero value implies bad prediction while the value one denotes that prediction coincides with measurement (i.e., perfect prediction). The absolute error is fuzzified using the fuzzy sets as shown in Figure 4; the parameters of the fuzzy sets are determined by the modeler according to the complexity and the type of the energy system under monitoring. In

Figure 4, the general frame of the fuzzy sets is presented. In addition the fuzzy sets for modeling the prediction quality are depicted in Figure 5. Association of the fuzzy sets that model the absolute error with the prediction quality fuzzy sets is performed by defining a set of fuzzy rules. In particular, the set of fuzzy rules employed for implementing the fuzzy inference mechanism are given below: 1. IF Absolute_Error is VERY SMALL, THEN Prediction Quality is VERY HIGH. 2. IF Absolute_Error is SMALL, THEN Prediction Quality is HIGH. 3. IF Absolute_Error is MEDIUM, THEN Prediction Quality is MEDIUM. 4. IF Absolute_Error is LARGE, THEN Prediction Quality is LOW. 5. IF Absolute_Error is VERY LARGE, THEN Prediction Quality is VERY LOW. The output of the inference system is a set of values in the interval [0, 1] that includes the indexes of prediction quality for each SVR model: M quality indexes for M models. There-



Figure 5. Fuzzy sets for modeling prediction quality

fore, quality indexes consist of indication of the prediction success of each predictor in the previous prediction step. Indexes are a piece of information utilized in the next step prediction. In the next step, the SVR models are used for prediction making. Initially, the predictions are obtained, then multiplied with the respective quality index and divided by their sum giving a weighted prediction, called the indicator value: M

Indicator =

∑ i ⋅V i =1 M

∑i

i

(9)

i =1

where i is the quality index and Vi is the predicted value. The indicator value expresses the weighted average prediction for the monitored energy system variable. Once the indicator is evaluated, then the final decision-making process takes place. More particularly, a fuzzy inference engine is employed for selecting a decision among a set of anticipated decisions/actions and for comput-

ing a certainty factor about that decision. The decision making inference system adopts fuzzy rules of the form: IF indicator is A(x), THEN decision should be D(x) (10) where A(x) represents a fuzzy set from those depicted in Figure 5, and decision is a fuzzy variable whose arguments are also fuzzy sets expressed as D(x). Details of the form of sets A(x) and D(x) as well as the number of fuzzy rules depend on the specifics of the application and the system performance (Alamaniotis et al., 2009). Evaluation of the decisions as expressed by the rules in Equation 10 allows implementation of anticipatory strategies. In general, rules and sets in Equation 10 are determined according to the modeler’s experience; the latter is the big advantage of the methodology since it can incorporate system operator’s experience in the model.



ANTICIPATORY CONTROL OF ENERGY SYSTEMS: APPLICATION ON TURBINE DEGRADATION RELATED DECISIONS Problem Statement The presented anticipatory system is applied on a case of monitoring turbine degradation (Liu et al., 2002). Datasets contains five different real measured histories of turbine blade degradation and may be found on ReliaSoft website (2014). More particularly, the five histories include experimental obtained time series expressing the crack length evolution of the respective turbine blade. Time is measured in terms of operation cycles, while measurements are taken every 1x106 cycles (i.e., the five measurements are taken at 1 × 106, 2 × 106, 3 × 106, 4 × 106 and 5 × 106 cycles). A blade fails to perform its normal operation when the crack length is equal to 30mm. The essential role of a turbine in energy generation power plants (i.e., conventional or nuclear) mandates its continuous monitoring and on-time maintenance or replacement (Alamaniotis et al., October 2014). Unexpected turbine failure may lead to long-term delay in energy production with severe and costly consequences for both suppliers and consumers. Anticipatory control is paramount for implementing predictive maintenance, and subsequently optimizing operational expenses. Thus, by predicting the crack length of the next measurement, the system operator can make control decisions about the turbine maintenance actions. In the current work, five datasets are used to validate the proposed methodology via a cross validation approach (Bishop, 2006). More particularly, at each instance the four datasets are used to train the four SVR models individually, while the fifth dataset is used for testing. In the test case, it is assumed that the decision variables express the crack length of four blades. Therefore, the decision vector is comprised of four entries; each entry is the prediction of an SVR model that has been trained using one of the available blade crack histories.

Thus, the values for the fuzzy parameters presented in Figure 4 are taken to be respectively: A = 5, B = 10, C = 15 and D = 20mm . In the current case, the indicator value computed by Equation 9, expresses the expected predicted crack length in the fifth blade. Furthermore, the SVR parameters (Bishop, 2006) were empirically selected as C=4 and v=0.5 (Alamaniotis et al., 2012; 2011). The performance indicator value is compared to the failure threshold. It should be noted that in the current study case the performance indicator value expresses the average prediction of the four SVR models. The absolute difference between the indicator value and the failure threshold (i.e., 30mm) is computed and fuzzified using the fuzzy sets presented in Figure 6. Furthermore, the maintenance control decisions (i.e., D(x) in Equation 10), which express the need for maintenance (decision to perform maintenance), are shown in Figure 7, while the fuzzy sets representing the need for replacement (decision to replace the component) of the component are depicted in Figure 8. The respective fuzzy logic rules (i.e., rules following form expressed in Equation 10, are the following: 1. IF Absolute_Difference is SMALL, THEN Need for Maintenance is HIGH. 2. IF Absolute_Difference is LARGE, THEN Need for Maintenance is LOW. 3. IF Absolute_Difference is VERY SMALL, THEN is Need for Replacement is HIGH. 4. IF Absolute_Difference is VERY LARGE, THEN Need for Replacement is LOW. 5. IF Absolute_Difference is MEDIUM, THEN Need for Maintenance is LOW and Need for Replacement is LOW.

Results In this section results on anticipatory control of the turbine degradation and maintenance are presented. Table 1 presents the leave-one-out cross-validation schema and provides five test cases as a result.



Figure 6. Fuzzy sets for modeling absolute difference between expected crack length and failure threshold

Figure 7. Fuzzy representation of variable “need for maintenance” used to make maintenance decisions

The prediction results obtained by using the presented anticipatory system for each of the five cases are shown in Table 2. More particularly,

the crack length prediction at each step (i.e., the index value computed by the anticipatory system) and its absolute difference from the



Figure 8. Fuzzy representation of variable “need for replacement” used to make component replacement decisions

real crack length are provided. In addition, the respective decision, which is either “maintenance” or “replacement,” is also provided together with its respective certainty. Observe that in all the five cases, the system prefers (in particular, it decides with higher certainty) in the first steps, the decision of “maintenance” of the turbine over replacement. While as the turbine crack evolves with time, and its length comes closer to the failure threshold, it is observed that the decision of “replacement” is preferred by the system. This trend is observed in all five cases, whose last step prediction crack is very close to the turbine failure threshold of 30 mm,

giving very high certainty in deciding turbine replacement. In Figure 9 we compare the prediction performance of the models involved in the system. More specifically, predictions obtained by the anticipatory system and each individual SVR model are compared by computing the average value of the absolute error over all five prediction steps. The absolute error expresses the non-negative difference between the respective model prediction and the measured (real) value. Furthermore, the model that provides the lowest error at each case is marked and indicated in Figure 9. Overall, observe that there is no

Table 1. Cross validation schema for the turbine degradation datasets under study Case No.

Training Datasets

Test Dataset

Case 1

Dataset 1

Dataset 2

Dataset 3

Dataset 4

Dataset 5

Case 2

Dataset 1

Dataset 2

Dataset 3

Dataset 5

Dataset 4

Case 3

Dataset 1

Dataset 2

Dataset 4

Dataset 5

Dataset 3

Case 4

Dataset 1

Dataset 3

Dataset 4

Dataset 5

Dataset 2

Case 5

Dataset 2

Dataset 3

Dataset 4

Dataset 5

Dataset 1



Table 2. Prediction results and respective decisions obtained by the anticipatory system at each step of each test case Step

Decision Process Anticipatory Index Value

Decision

Absolute Difference |Index Value-Real|

Action

Certainty

Case 1: Test Dataset 5 Step 1

11.9639

1.9639

Maintenance

64.7%

Step 2

15.4697

0.4697

Maintenance

33.9%

Step 3

18.2225

1.7775

Replacement

56.8%

Step 4

24.8454

1.1546

Replacement

72.5%

Step 5

29.1029

3.8971

Replacement

99%


12.0022

0.0022

Maintenance

26.1%

Step 2

15.6277

0.3723

Maintenance

26.4%

Step 3

19.6862

2.6862

Replacement

50%

Step 4

24.5742

4.5742

Replacement

50%

Step 5

29.9992

3.9992

Replacement

99%

5.6263

Maintenance

26.2%


11.3737

Step 2

16.7450

8.255

Maintenance

28.1%

Step 3

26.5385

0.5385

Replacement

46.4%

Step 4

25.1760

1.824

Replacement

48.9%

Step 5

29.2140

3.786

Replacement

98.1%

2.7299

Maintenance

28.1%


12.7299

Step 2

17.6221

2.6221

Maintenance

29.5%

Step 3

22.0691

2.0691

Maintenance

68.0%

Step 4

24.8500

0.1500

Replacement

74.0%

Step 5

29.7800

0.2200

Replacement

98.3%


11.4696

3.5304

Maintenance

26.1%

Step 2

19.3177

0.6823

Maintenance

30.8%

Step 3

22.1192

0.1192

Maintenance

68.0%

Step 4

25.4014

0.5986

Replacement

48.7%

Step 5

29.6835

0.6835

Replacement

99%

unique model that provides the lowest error in all cases, but instead a different model in each case. However, the anticipatory system has the best prediction performance in two out of five

cases while it consistently provides the lowest error. In other words, the average predictions obtained by the anticipatory systems, if not the best then it is one of the best among the models.



Figure 9. Average absolute errors for individual SVR models and the presented anticipatory system for Cases 1–5 of Table 2

The latter observation validates the robustness of the presented system in prediction and decision making. It should be noted that the results presented in Table 2 and Figure 9, respectively, confirm the advantage of the presented anticipatory system in robust control of complex energy systems. The proposed approach can be generalized to other energy systems.

CONCLUSION The significance of energy systems in the world, demands control mechanisms to secure their safe and nonstop operation. In the current paper a new intelligent system for controlling complex energy systems is presented. More particularly an anticipatory system is introduced for prediction making of values of variables, which are essential for decisions, and then perform decision making. The presented anticipatory system utilized the synergism of a set of support vector regressor models and

fuzzy logic tools to perform robust control of complex energy systems. All the SVR models are equipped with the same kernel function, namely, the Gaussian kernel. The presented system was tested on a set of real-world experimentally taken datasets regarding the turbine degradation evolution. The system aims at predicting the crack length on the turbine blades and subsequently making decisions with respect to “maintenance” or “replacement” actions. Results demonstrated its potentiality over the use of single support vector regressors or mean of SVR predictions. On the average, the presented system performs accurate predictions, and promotes appropriate decisions.

DISCLAIMER This information was prepared as an account of work sponsored by an agency of the U.S. Government. Neither the U.S. Government nor any agency thereof, nor any of their employees,



makes any warranty, expressed or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness, of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. References herein to any specific commercial product, process, or service by trade name, trade mark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement, recommendation, or favoring by the U.S. Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the U.S. Government or any agency thereof.

REFERENCES Agarwal, V., Lybeck, N., Pham, B., Rusaw, R., & Bickford, R. (2013). Online Monitoring of Plant Assets in the Nuclear Industry. Prognostics and Health Management Society Annual Conference, New Orleans, LA. Alamaniotis, M., Agarwal, V., & Jevremovic, T. (July 2014). Anticipatory monitoring and control of complex energy systems using a fuzzy based fusion of support vector regressors. The 5th International Conference on Information, Intelligence, Systems and Applications, IISA 2014, (pp. 33-37). Chania, Greece. Alamaniotis, M., Gao, R., Tsoukalas, L. H., & Jevremovic, T. (November 2009). Expert System for Decision Making and Instructing Nuclear Resonance Fluorescence Cargo Interrogation. 21st International Conference on Tools with Artificial Intelligence, 2009. ICTAI’09. (pp. 666-673). Newark, New Jersey. Alamaniotis, M., Grelle, A., & Tsoukalas, L. H. (2014, October). Regression to fuzziness method for estimation of remaining useful life in power plant components. Mechanical Systems and Signal Processing, 48(1-2), 188–198. doi:10.1016/j. ymssp.2014.02.014 Alamaniotis, M., Ikonomopoulos, A., & Tsoukalas, L. H. (2011, August). Online Surveillance of Nuclear Power Plant Peripheral Components using Support Vector Regression. Proceedings of the International Symposium on Future I&C for Nuclear Power Plants, Cognitive Systems Engineering on Process Control, and International Symposium on Symbiotic Nuclear Power Systems (ICI 2011), Daejeon, Korea, (pp. 1230(1-6)).

Alamaniotis, M., Ikonomopoulos, A., & Tsoukalas, L. H. (2012). Optimal assembly of support vector regressors with application to system monitoring. International Journal of Artificial Intelligence Tools, 21(06), 1250034. doi:10.1142/S0218213012500340 Basak, D., Pal, S., & Patranabis, D. C. (2007). Support vector regression. Neural Information ProcessingLetters and Reviews, 11(10), 203–224. Bishop, C. M. (2006). Pattern recognition and machine learning (Vol. 1, p. 740). New York: Springer. Chiu, S., Chand, S., Moore, D., & Chaudhary, A. (1991). Fuzzy logic for control of roll and moment for a flexible wing aircraft. Control Systems, IEEE, 11(4), 42–48. doi:10.1109/37.88591 Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector machines and other kernel-based learning methods. Cambridge university press. doi:10.1017/CBO9780511801389 Harris, C. J., Moore, C. G., & Brown, M. (1993). Intelligent control: Aspects of fuzzy logic and neural networks. World Scientific Press (Robotics and Automated Systems), 6. Ikonomopoulos, A., Tsoukalas, L. H., & Uhrig, R. E. (1993). Integration of neural networks with fuzzy reasoning for measuring operational parameters in a nuclear reactor. Nuclear Technology, 104(1), 1–12. Kusiak, A., Song, Z., & Zheng, H. (2009). Anticipatory control of wind turbines with data-driven predictive models. IEEE Transactions on Energy Conversion, 24(3), 766–774. doi:10.1109/TEC.2009.2025320 Liu, G. L., Hu, N. S., Zhao, Y., & Wu, J. F. (2002). Parameter simulation in performance monitoring system of steam turbine unit for a fossil-fuel power plant. 2002. Proceedings. 2002 International Conference on Machine Learning and Cybernetics, Vol. 3. (pp. 1249-1252). Beijing, China. doi:10.1109/ ICMLC.2002.1167402 Rasmussen, C. E. (2006). Gaussian processes for machine learning. MIT Press. ReliaSoft Corp. website, Retrieved August 2014. http://www.reliasoft.com Rosen, R. (2012). Anticipatory systems (pp. 313–370). Springer New York. doi:10.1007/978-14614-1269-4_6 Ross, T. J. (2013). Fuzzy logic with engineering applications (Vol. 686). Wiley.



Shawe-Taylor, J., & Christanini, N. (2000). Kernel methods for pattern analysis. Cambridge, UK: Cambridge University Press. Tsoukalas, L., & Reyes-Jimenez, J. (1990, June). A hybrid expert system-neural networks methodology for anticipatory control in a process environment. Proceedings of the 3rd international conference on Industrial and engineering applications of artificial intelligence and expert systems-Volume 2 (pp. 10451053). ACM. doi:10.1145/98894.99121 Tsoukalas, L. H. (1989). Anticipatory systems using a probabilistic-possibilistic formalism. Doctoral dissertation, University of Illinois at UrbanaChampaign. Tsoukalas, L. H. (1998). Neurofuzzy approaches to anticipation: a new paradigm for intelligent systems. Systems, Man, and Cybernetics, Part B: Cybernetics. IEEE Transactions on Cybernetics, 28(4), 573–582. doi:10.1109/3477.704296 Tsoukalas, L. H., & Uhrig, R. E. (1996). Fuzzy and neural approaches in engineering. John Wiley & Sons, Inc.

Van Dam, K. H., Ottjes, J. A., Lodewijks, G., Lukszo, Z. V., & Wagenaar, R. W. (2004, October). Intelligent infrastructures: distributed intelligence in transport system control-an illustrative example. 2004 IEEE International Conference on Systems, Man and Cybernetics. Vol. 5. (pp. 4650-4654). The Hague, Netherlands. doi:10.1109/ICSMC.2004.1401265 Vapnik, V., Golowich, S. E., & Smola, A. (1997). Support vector method for function approximation, regression estimation, and signal processing. Advances in Neural Information Processing Systems, 281–287. Vapnik, V. N. (1995). The nature of statistical learning theory. Springer. doi:10.1007/978-1-4757-2440-0 Xinqing, L., Tsoukalas, L. H., & Uhrig, R. E. (1996, September). A neurofuzzy approach for the anticipatory control of complex systems. FUZZ-IEEE ‘96, Proceedings of the Fifth IEEE International Conference on Fuzzy Systems Vol. 1 (pp. 587-593). New York: Institute of Electrical and Electronics Engineers. doi:10.1109/FUZZY.1996.551806

Uhrig, R. E., & Tsoukalas, L. H. (2003). Multi-agentbased anticipatory control for enhancing the safety and performance of Generation-IV nuclear power plants during long-term semi-autonomous operation. Progress in Nuclear Energy, 43(1), 113–120. doi:10.1016/S0149-1970(03)00003-9 Uhrig, R. E., Tsoukalas, L. H., & Ikonomopoulos, A. (1994). Application of neural networks and fuzzy systems to power plants. Neural Networks, 1994. IEEE World Congress on Computational Intelligence. 1994 IEEE International Conference on Neural Networks. Vol. 6 (pp. 3703-3718).


How to Submit Proof Corrections Using Adobe ...

How to Submit Proof Corrections Using Adobe ...

Suggest Documents