Puigjaner L. (ed.) Syngas from Waste

Подождите немного. Документ загружается.

Industrial Data Collection

Aarón D. Bojarski, Carlos Rodrigo Alvarez Medina,

Mar Pérez–Fortes and Pilar Coca

Abstract In this chapter, a general description of data-mining techniques is done

in the context of IGCC operation. The different control philosophies applicable to

IGCC operation are discussed together with different examples of data reconcili-

ation based on process simulation. The problem of process monitorisation, as an

example of data-mining application, is extensively discussed and an approach

based on PCA is presented.

Notation

ASU Air separation unit

CC Combined cycle

CPV Cumulative percent variance

DCS Distributed control system

DR Data reconciliation

ICA Independent component analysis

MSPC Multivariate statistical process control

NOC Normal operating condition

OTC Outlet temperature corrected

PC Principal component

A. D. Bojarski (&)  C. R. A. Medina  M. Pérez–Fortes

Universitat Politècnica de Catalunya, ETSEIB, Diagonal 647, 08028 Barcelona, Spain

e-mail: aaron.david.bojarski@upc.edu

C. R. A. Medina

e-mail: carlos.rodrigo.alvarez@upc.edu

M. Pérez–Fortes

e-mail: mar.perez-fortes@upc.edu

P. Coca

ELCOGAS, S.A.Ctra, Calzada de Calatrava a Puertollano, Km. 27 Apdo. Correos 200,

13500 Puertollano–(Ciudad Real), Spain

e-mail: pcoca@elcogas.es

L. Puigjaner (ed.), Syngas from Waste, Green Energy and Technology,

DOI: 10.1007/978-0-85729-540-8_13,  Springer-Verlag London Limited 2011

299

PCA Principal components analysis

PIMS Plant information system

PLS Partial least squares

SPE Squared prediction error

1 Introduction

A common characteristic of modern industries is the ability to generate and store a

very large amount of data that are frequently used to describe the process

behaviour. Generating and collecting data have become the most important topics

in different and diverse areas of knowledge. Nevertheless, the ability of collecting

such data has increased faster than the capabilities to analyze it. The biggest

available data collection in any area is utterly useless by itself. The ability of users

of those databases to extract useful information from them is a key part in modern

plant practices to push plant-operating conditions to improve proﬁt. The ﬁeld of

extracting useful information from databases is usually referred to as Knowledge

Discovery in Databases (KDD). KDD can be understood as the non-trivial process

of identifying novel data patterns that could prove to be useful and understandable

[1] (p. 2). In the specialised literature, KDD is understood as a sequential process

that involves the following stages:

• Analysis objective setting.

• Data selection: This selection is done according to the objectives followed by

the analysis, and is mainly related to the aspects of how to access and store the

data.

• Data pre-processing: This step is done to ensure the quality of data aiming at

eliminating: noise, outliers and dealing with lost and unreliable data.

• Data transformation is done using different techniques; in most cases, it aims at

reducing the dimensions considered (by projection), which might improve the

identiﬁcation of patterns.

• Data mining: This is the core step, where signiﬁcant and well-deﬁned patterns

are looked for.

• Interpretation and validation.

This chapter discusses the issues related to the gathering data and its subsequent

analysis in the context of KDD for the case of syngas production for energy use.

Section 2 discusses the implementation of a control system that gathers and

monitors the process variables, making special emphasis on an industrial case, the

ELCOGAS power plant. Section 3 discusses the different methodologies associ-

ated to Data Reconciliation, where variables information are used to improve the

knowledge of model parameters, while the ending Sect. 4 provides with a

300 A. D. Bojarski et al.

description of a data-mining procedure that could be implemented to data recorded

on the previously discussed control systems.

2 Process Variables Recording and Control

In general, the control system of a plant is used to keep the plant working properly

within the established limits (efﬁciency, net output, etc.), and as a consequence,

this control system is usually the source of data for any analysis. The different

measurement systems provide with signals that are stored by centralised software

that acts as the data reservoir. Data selection and pre-processing can be done using

different software, such as MS Excel, Matlab and other software suites.

The Puertollano IGCC plant is equipped with a distributed control system

(DCS) using Siemens Teleperm XP for the whole plant except for both turbines

that have a speciﬁc automation called Simadyn. This system has a modular

structure and consists of the following subsystems:

• An automatic system for automatic function implementation at the lowest control

level

• A communications network

• An operation control and monitoring system for operation processes and infor-

mation interchange

• An engineering system for planning, conﬁguration and start up

In a short time, both Teleperm and Simadyn are going to be replaced by an

updated automation system T-3000 that is going to integrate both distributed

control systems.

2.1 Variables Recording

Main features of the process—apart from pressure, temperature and ﬂows—that

are monitorised during operation are:

• In the gasiﬁer, quality of gasiﬁcation process, which is evaluated taking into

account syngas composition, such as CO

, CO, energy exchange in the

immersion shaft and energy gasiﬁcation reaction.

• In the combined cycle (CC), clean gas LHV, OTC (outlet temperature corrected)

of the gas turbine exhaust gases and combustion stability. The OTC is a measure

of exhaust gas temperature that reﬂects the temperature in the combustion

chamber and is represented as the difference between the average temperature of

the ﬂue gas (measured at line) and outside air temperature.

• In the air separation unit (ASU), purity of oxygen, pure and waste nitrogen.

Industrial Data Collection 301

The former variables are fed to a plant thermo-economic model. Advanced

thermo-economic diagnosis of the plant equipment performance is intended to

improve the plant operation and maintenance procedures. Additionally, steady-

state simulation tools assess the potential for design improvements and the impact

of plant modiﬁcations.

The diagnosis system focuses on the effective management of both ﬁxed inputs

(e.g., plant layout, instrumentation, etc.) and operation and management (O&M)

data (e.g., process data, operating modes, maintenance schedule, etc.) to screen the

equipment performances and operating costs at steady-state conditions. The Plant

Information Management System (PIMS), with its own interface to the DCS

(Distributed Control System), laboratory and other data sources provide raw data

that is sorted, validated and stored. The cost evaluation module calculates the

potential savings that can be achieved by comparing the performance tests (based

on actual plant measurements) and the state of reference model (best possible

operating case), under given operating constraints.

The system conceptual design enables online applications and easy module

update and/or modiﬁcation. The main applications are:

• Plant raw process data validation and reconciliation ahead of energy/exergy

calculations. Dynamic monitoring of plant instrumentation.

• Comprehensive map of resources thorough the plant at the process stream level.

• Costs evaluation: actual versus best possible operation. Allocation and causes

of performance degradation.

• Assess the extra cost because of performance degradation versus corrective

O&M actions impact.

• Online monitoring: plant heat rate, cost targets compliance and instrumentation

status.

2.2 Control Levels

The current system’s automation and control levels (Fig. 1) of the Teleperm XP

are as follows:

• Field level: The lowest level where sensors are located and data are withdrawn.

Its function is to receive signals from sensors and to transmit them to the higher

levels or the actuators.

• Automation level: The automation level has two sub-levels:

– Individual control level: Basic control of operations with analogue and binary

signals. Actuators control in open loop and individual control in closed loop.

– Group control level: Automatic functions, such as closed loop regulations,

open loop control and signal protection management.

• Process level: Storage of process data and transfer of the dynamic information

to the man–machine interface.

302 A. D. Bojarski et al.

• Operation and control level: Association of the man–machine interface with

the interaction supervision and conﬁguration systems.

2.3 Plant Control Philosophy

The ELCOGAS power control operates on gas turbine leading mode, so both

gasiﬁer and ASU follow the gas turbine load. Therefore, the clean gas pro-

duction by the gasiﬁcation island depends on the gas turbine demand, being the

objective to consume as much gas as produced to match the power demand of

the grid. The ASU produces the necessary oxygen to obtain the required clean

gas. The steam turbine works in slide control, taking the vapour produced by

both the CC and gasiﬁer boilers. However, a signiﬁcant time delay (a couple of

minutes) is needed to accordingly adjust the feed streams to the gasiﬁer and

particularly the coal ﬂow rate. It is a major task of the proposed control concept

to harmonize these diverging facts, taking into account the storage capacity of

the gas path.

One of the main objectives of the control philosophy is to ensure stable

operating conditions for the gasiﬁer. For this target, the gasiﬁer pressure is used

as a principal controlled variable and is kept at a ﬁxed value. The differential

pressure along the gas path is dependent on gas production, and the gas

inventory constitutes a reservoir that can be loaded and unloaded. The gasiﬁer

and the sulphur removal units contain a certain gas inventory determined by the

geometrical volume, temperature and pressure. In case of a short-term gap

between clean gas production and demand of the gas turbine, part of this

inventory can be used.

Fig. 1 Subsystems of the

Teleperm XP

Industrial Data Collection 303

2.4 Coordinated Control

The coordinated control between the CC and the gasiﬁer is based on the following

principles:

• The load set point of the gas turbine is sent to the load controller.

• The gasiﬁer is controlled by a pressure control and uses the set point of the gas

turbine load as an anticipating signal.

• To prevent any depressurization in case of any problems with gasiﬁer load, the

gasiﬁer deﬁnes a maximum set point of clean gas ﬂow to be taken by the gas

turbine, which differs with the actual ﬂow controls of the CC clean gas valve

when the set point is exceeded.

The set point of the power plant load is given by the control room operator and

is sent by the controller to each island. This set point is elaborated with a maxi-

mum gradient of 3% load/min, according to the ASU limitation and is used:

• By the CC to elaborate the load set point of the gas turbine and to determine the

gas ﬂow consumption actuating on the position of control elements (clean gas

valve and Inlet Guide Vanes). This set point is adjusted, taking into account the

steam turbine contribution.

• By the gasiﬁer to calculate, with a proportional value, the clean gas ﬂow and the

oxygen ﬂow. The set point of oxygen and nitrogen ﬂows is elaborated by the

gasiﬁer for the ASU.

• On the other hand, the steam turbine is considered in case of uncontrolled steam

turbine load variation (i.e., steam turbine trip or during load ramping), the

control is designed to by-pass the steam turbine load inﬂuence in the elabo-

ration of the gas turbine set point.

So, the power plant control operates on gas turbine leading mode, the gasiﬁer—

working in pressure control—and the ASU follows the gas turbine load, even if

because of operational problems the actual load does not match the external set

point.

In general, the former primary control loops try to use most of the information

available in the process historical, however, they do not perform an efﬁcient use of

them and other techniques such as the ones described in next sections that could

allow for extracting other information and knowledge from the raw control data.

3 Data Reconciliation

The problem of data reconciliation (DR) was ﬁrst stated by Kuehn and Davidson

[2] for a total material balance where all variables were measured. Several works

in both data reconciliation and gross error detection areas have been published

since then. Important improvements to the basic DR technique were presented in

304 A. D. Bojarski et al.

those years including treatment of unmeasured variables using a graph-theoretic

approach and the introduction of the concepts of observability and redundancy.

Nonlinear data reconciliation was ﬁrst addressed by Knepper and Gorman [3]

using an iterative technique proposed originally with parameter estimation that

employs a nonlinear regression. In general, because the nonlinear data reconcili-

ation problems are basically nonlinear optimization problems, some well-known

constrained nonlinear optimization methods have been used to solve them, for

instance, Tjoa and Biegler [4] used sequential quadratic programming (SQP)

technique to solve a combined data reconciliation and gross error detection

problem. Several other approaches can be seen in Romagnoli and Sanchez [5] and

Alvarez-Medina [6].

3.1 Data Reconciliation Basics

Generally speaking, data reconciliation can be deﬁned as the process of adjusting

(reconciling) process measurements, such as ﬂow rates, temperatures, composi-

tions, etc., to obtain a new set of estimates that are consistent with mass and energy

balances, and sometimes, with some thermodynamic equilibrium equations as

well. In the context of KDD, DR is used for data transformation, given that DR

techniques receive pre-processed data and produce ‘reconciled data’, which could

improve the detection of trends in data.

From the data reconciliation’s point of view, in the absence of gross errors, a

process measurement can be ‘modelled’ as follows:

y ¼ x þ e; y; x; e 2 R

ð1Þ

where y is a (n 9 1) measurement vector, x is a (n 9 1) vector of true variables

values and e is a vector of random measurement errors.

The statistical basis in DR relies basically in the assumptions made regarding

vector e. It is usually assumed that:

1. Vector e has a null expected value, E (e) = 0, and it is normally distributed.

2. Random errors for successive measurements are independent, i.e., E e



0; for i = j.

3. Cov eðÞ¼W ¼ E e



is known (or estimable) and positive deﬁnite.

The data reconciliation problem can be generally stated as the following

constrained weighted least-squares estimation problem deﬁned in Eq. 2.

min

x;u

1

s:t: uðx; uÞ¼0

x x

u u

ð2Þ

Industrial Data Collection 305

In general, the constraints sets are presented as a set of nonlinear algebraic

equations, such as:

uðx; uÞ¼0 x 2 R

u 2 R

; ð3Þ

where u indicates the (m 9 1) vector of unmeasured variables and uðx; uÞ is the

constraints set. Because random errors have been assumed normally distributed, by

solving the above problem, the maximum likelihood estimates of process variables

are obtained. Different methodologies have been proposed for solving these

problems depending on whether the constraints constitute a linear, bilinear or a

nonlinear set of equations.

One of the most common calculations made in chemical engineering is total

mass balances, because they enable the engineer to get a general idea of the

process state. If the purpose is to reconcile total mass ﬂows, or molar ﬂows if

reactions are not considered, a linear data reconciliation problem results. Two

different situations can be found depending on if dealing with only measured or

both measured and unmeasured variables.

In the case of having all the ﬂowrates measured, then Eq. 2 can be rewritten as

in Eq. 4, given that Eq. 3 is simply a set of linear relationships among the different

measured ﬂows.

min

ðy  xÞ

1

ðy  xÞ

s:t: Ax ¼ 0

; ð4Þ

where A is a (i 9 n) incidence matrix, which shows the relationship among

ﬂowrates and plant equipment. Matrix A elements (a

) are 1 or -1 if the nth ﬂow

enters or leaves equipment i, whereas zero if the ﬂow is not related. Because of the

simplicity of the restrictions, the former problem can be solved applying the

Lagrange multipliers method, the improved estimate of the process variable value

(

x) can be obtained as:

x ¼ y  WA

ðAWA

1

Ay ð5Þ

For the case of unmeasured ﬂow rates, the reader is referred to Crowe et al. [7]

and Sánchez and Romagnoli [8], where the reader will ﬁnd different implemen-

tations of linear and bi-linear DR problems.

In the general case, process operation is commonly modelled by nonlinear

systems of algebraic equations. This nonlinearity comes from components or

energy balances and it may also include thermodynamic relationships to explain

some particular system’s behaviour. Sometimes, process control systems require

an accurate process model, and the most common linear or bilinear data recon-

ciliation strategies are not applicable anymore. To these cases, the most general

mathematical formulation of an optimization problem is as in Eq. 2. The con-

straints set (uðx; uÞ) is in general nonlinear and customarily difﬁcult to solve.

The necessary conditions for an optimal solution come from the deﬁned Lagrange

function and are known as Karush–Kuhn–Tucker (KKT) conditions. The sufﬁcient

306 A. D. Bojarski et al.

condition for obtaining a global minimum of the nonlinear problem is that both

the objective function and the constraint set must be convex. Otherwise, there is

no guarantee that the reached local optimum will be the global optimum.

As discussed in chapter ‘‘Modelling Superstructure for Conceptual Design of

Syngas Generation and Treatment’’, most commonly used optimization software,

such as GAMS or MatLab packages, includes several well-known optimization

algorithms as SQP. Other techniques include solvers such as NPSOL which is

especially effective for nonlinear problems whose functions and gradients are

expensive to evaluate.

Because of the complexity included in the set of constraints of Eq. 3, process

simulation environments are used to solve them with ease, as discussed in

chapters ‘‘Modelling Syngas Generation, Main Puriﬁcation Operations and

Modelling Superstructure for Conceptual Design of Syngas Generation and

Treatment’’. Consequently, the application of DR techniques requires frame-

works that use the process simulation data together with other optimisation

algorithms. In such frameworks, the simulation environment usually acts as a

server of the optimisation algorithm by providing the value of Eq. 3 for the

different values of the variables being optimised. The following examples try to

clarify on this sense.

3.2 Examples of Data Reconciliation in obtaining Synthesis Gas

3.2.1 Case A: DR Applied to a Claus Desulphurisation Process

Recovery of sulphur from industrial waste gases is an important problem from the

environmental point of view, and as it was discussed in chapter ‘‘Main Puriﬁcation

Operations’’, it is of paramount importance in the case of IGCC power plant

operation. The rise of sulphur volumes in waste gases together with tightening

emission regulations leads to the increase of sulphur-recovering needs [9].

In gasiﬁcation plants, the main source of sulphur recovered is hydrogen sulphide

produced in the gasiﬁcation step as a sub-product. The most widely used method to

treat H

S is based on oxidation of hydrogen sulphide into sulphur by adding oxy-

gen, further details are discussed in chapter ‘‘Main Puriﬁcation Operations’’ .

The global sulphur recovering in Claus plants containing recirculation is around

99.8% of the sulphur content in the input gas. All the equipments and process units

involved in the model are put together as a whole operational unit. Six streams are

considered to link the Claus plant to the rest of the whole process equipments and

facilities, see Fig. 2.

The proposed strategy described in Eq. 2 has been applied to the above

described Claus plant to reduce the noise level in the inputs and output streams.

The set of constraints are modelled using the process simulator (in this case,

AspenHysys). These constraints include: mass and energy balances together with

the corresponding thermodynamic models for the calculation of phase distribution.

Clearly, the simulation environment eases the implementation of such framework

Industrial Data Collection 307

Puigjaner L. (ed.) Syngas from Waste - Emerging Technologies

Подождите немного. Документ загружается.