You are on page 1of 6

Available online at www.sciencedirect.

com

ScienceDirect
Energy Procedia 105 (2017) 524 – 529

The 8th International Conference on Applied Energy – ICAE2016

Steam Turbine models for monitoring purposes


S. Dettoria, V. Collaa*, G. Salernob, A. Signorinib
a
Scuola Superiore Sant’Anna TeCIP Institute PERCRO Laboratory, Via Alamanni 13D, Ghezzano, 56124, Italy
b
General Electric Oil & Gas, Via F. Matteucci 2, Firenze, 50127, Italy

Abstract

The paper proposes two different approaches for steam turbine modelling for on-line monitoring applications, a
hybrid-thermodynamic method and a neural network approach. Both models can predict power and other features that
cannot be easily measured such as outlet Steam Quality, Pressure and Temperature at drums outlet. Training and
validation of both models was carried out by exploiting a dataset created by means the GE sizing design tool. The
models were tested by means of real field data of a High Pressure Turbine.

©©2017
2016TheTheAuthors. Published
Authors. by Elsevier
Published Ltd. This
by Elsevier is an open access article under the CC BY-NC-ND license
Ltd.
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Selection and/or peer-review under responsibility of ICAE
Peer-review under responsibility of the scientific committee of the 8th International Conference on Applied Energy.
Keywords: Steam turbine models; Monitoring; Steam Quality; Neural Networks; Thermodynamic Models.

1. Introduction

The monitoring and control of aging of Steam Turbines (STs) is a widely debated topic, especially in the
cases of very variable steam operating conditions or in the case of turbines which are subjected to frequent
start-up and shut-down cycles. STs with non-stables steam conditions are subjected to continuous thermal
stresses, windage effects and premature aging with consequent decrease of efficiency. In this case on-line
monitoring algorithms based on turbomachinery models have to balance a good accuracy and simplicity,
i.e. two fundamental and often counteracting features, in order to be implemented on platforms such as
PLC, which typically has stringent requirements in terms of memory and computational speed. Such
models and algorithms are applied to monitor phenomena and inaccessible variables (e.g. windage,
efficiency, steam quality (the percentage of vapor mass in a liquid-vapour mixture), pressures and
temperatures between blades) that are usually not measured through sensors.
An ST model, allowing to monitor the machine also during off-design conditions, has to keep into
account different heavy nonlinear behaviours, such as the dependence of efficiency on rotational speed

* Corresponding author. Tel.: +39-050-882507; fax: +39-050-882564.


E-mail address: colla@sssup.it .

1876-6102 © 2017 The Authors. Published by Elsevier Ltd. 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 the scientific committee of the 8th International Conference on Applied Energy.
doi:10.1016/j.egypro.2017.03.351
S. Dettori et al. / Energy Procedia 105 (2017) 524 – 529 525

and ratios between drum pressures, the mass flow (typically estimated by means of Stodola equation [1]),
the dependence of windage effect on ratios between pressures and the kinetic energy contributions which
depends on steam velocity and geometry of the blades. The approaches used for simplified ST modelling,
are typically based on semi-empirical relations [2-4] or on approximation of thermodynamic equations
[5], less frequently on the use of black box methods such as Neural Networks (NN) [6]. The work
presented in this paper compare two different approaches to steam turbine modelling, in particular an
iterative hybrid thermodynamic model and a Feed Forward NN model (FFNN).
The paper is organized as follows: Sec. 2 describes the iterative hybrid thermodynamic model focusing
on physical equations and relative approximation useful to estimate the features of interest. Sec. 3
describes the NN-based approach. Sec.4 depicts the methods for designing the training and validations
datasets for both models. Finally, Sec. 5 provides some concluding remarks and hints for future work.

2. The Hybrid-thermodynamic Model

The hybrid-thermodynamic approach aims at extending a physic-based approach through the use of
approximated relationships which describe the nonlinear behaviour of efficiency, pressure ratios between
drums at varying of operating conditions of steam and rotational speed of the machine. Therefore, the
simple equations, such as Stodola or the computation of enthalpy drop between drums, are modified in
order to reproduce the real behaviour of the machine. With respect to classical thermodynamic models,
which describe the steam flow along the stator and the rotoric blades, the point of view is extended to the
sections of machine, i.e. the drums.
Each drum of the steam turbine is modelled in order to estimate the fundamental parameters such as
mass flow, pressure, temperature, enthalpy and power. In particular, the mass flow is related to the inlet
and outlet conditions using the Stodola equation, expressed as follows:

Pin2  Pout
2
m k (1)
Tin

and extended by means a polynomial with order greater than 3, according to the following equation:

Pin2  Pout
2
k s1m 3  k s 2 m 2  k s 3m  k s 0 (2)
Tin

The identification of coefficients k si is performed for a wide range of operating conditions, by


minimizing the mean squared error between predicted and actual mass flow. The i-th coefficient k si
depends on the rotational speed Z and on the stroke of the inlet steam valve through a nonlinear
relationship that is matched by representing the k si coefficient through a lookup table. An example of the
outcome of the identification procedure is presented in Figure 1.
Once the coefficients k si have been identified, the drum exit pressure is estimated by means Eq. (2),
knowing the inlet conditions of each drum (pressure, temperature and mass flow). Other quantities (such
as enthalpy, entropy etc.) are evaluated by means of steam tables functions [7].
The relationship between the real exit enthalpy and the isentropic exit enthalpy, typically described as:

H in  H out 'H
K (3)
H in  H out
iso
'H iso
526 S. Dettori et al. / Energy Procedia 105 (2017) 524 – 529

Example of Stodola Eq.Fit


1.8

1.6

1.4

1.2

0.8 Internal Sizing Tool


0.6 Fitting Curve

0.4

0.2

0
0 10 20 30 40 50 60 70 80
Mass Flow Rate

Fig. 1. Example of Stodola Equation fit.

has been approximated by means of a polynomial, as follows:

H out
H in  K3'H iso
3
 K2 'H iso
2
 K1'H iso  K0 (4)

where K i coefficients are identified with the same approach used with Eq. (2). In particular, a set of
coefficient was identified as a function of Z. This dependency is matched by expressing the K i
coefficients as lookup tables.
Outlet temperature and vapour quality are evaluated by means of steam tables functions [7].
The power of each drum is computed as a function of mass flow, the enthalpy drop on each drum of the
steam turbine and the kinetic energy contributions, as follows:

KWout
m H in  H out  Einkin  Eout
kin
(5)

where the kinetic contributions are computed by taking into account the mass flow conservation law and
the triangle of velocity at the exit of drum [1].
In order to match the pressure at the exit of the last drum of the turbine (input of the model), an
iterative approach has been implemented: the inlet mass flow is computed iteratively in order to match the
actual exit pressure with the one coming from the thermodynamic model (see Figure 2).
The logic of the iteration algorithm is based on the Eq. (1), assuming that, for each operating condition,
a k exists linking the mass flow to some boundary conditions, according to the following basic steps:
1. Compute the pressure error, check if the estimated exit pressure matches the actual exit pressure;
if yes, the current mass flow is correct; if not go to the next step;
2. If the estimate of the exit pressure coming from Eq. (2) is a complex number, the mass flow is
computed as 5%, otherwise it is computed as follows.

Pin2  Pact
2
m i 1 m i (6)
Pin2  Pout
2
i

3. if the sign of the pressure error varies and the exit pressure computed as in Eq. (2) is a complex
number, the mass flow is computed as the average value between the mass flow computed at
step i and i+1. otherwise it is computed through Eq. (6).
4. If the error on pressure lower that 0.1% of the exit pressure the mass flow is correct
S. Dettori et al. / Energy Procedia 105 (2017) 524 – 529 527

Fig. 2. Flow diagram of the iterative procedure.

3. Neural Network-based modelling approach

NNs were investigated for steam turbine modelling, in order to predict different characteristics of the
machine. Four different NN were designed: NNs were used to predict the overall power of the turbine
(NN1) and the power of each drum (NN2). NN were also applied to the prediction of temperatures and
pressure at exit of each drums (NN3) and the estimate of steam quality at turbine outlet (NN4).
FFNN are a typical choice for approximation of heavy nonlinear input-output relationships. The
number of weights and hidden layers was determined through an optimization study. All the NNs were
trained through the standard Levenberg-Marquardt backpropagation procedure [8] and early stopping
algorithm. The target function to minimize was the MSE between prediction and the target output.
The input of each NN is composed by inlet steam pressure and temperature, mass flow of each
bleeding or extraction and outlet pressure.

4. Numerical Results

In order to test the proposed approaches, a real High Pressure (HP) steam turbine was modelled, which
is composed by an impulse stage and 3 reaction drums, with 1 bleeding and a nominal power of 18 MW.
The dataset used for the training of both approaches was designed through the GE internal sizing tool, by
creating a large set of operating conditions, ranging from nominal to those most off-design ones; 80% of
data were used for training, the remaining ones for validation. An additional dataset was created starting
from field data measured during on full day, useful to test the approaches.
The accuracy of the models is evaluated through the validation dataset by exploiting 4 indexes: The
Mean Absolute Error (eav), the Maximum Absolute Error (EM), the Maximum Percentage Absolute Error
respect to target value (EP) and the Standard Deviation (STD).
The NN models provide on the overall power a value of the first index eav=8.3 kW over 18 MW, with
EP=0.5%. The power of each drum is estimated with EP in the range 0.3-1.2%. Pressures and
528 S. Dettori et al. / Energy Procedia 105 (2017) 524 – 529

Temperatures at exit of each drum, are estimated with EP in the range 0.8-1.3% and 0.1-0.8%,
respectively. The outlet steam quality is estimated with EP=0.1%. The thermodynamic model estimates
the overall power with eav=571 kW over values around 18 MW, with EP=5.9%. The power of each drum
is estimated with EP in the range 4.6-34%, being the value EP=34% obtained for the impulsive drum.
Pressures and Temperatures at exit of each drum are estimated, respectively, with EP in the range 4.7-
4.9% and 0.9-1.9%.
Fig. 3(a) shows the daily profile of pressures and temperatures of HP turbine (the test dataset,
continuous line) and their esteem obtained through NN3 models (dashed line). Fig. 3(b) shows the daily
profile of the power generation of each HP drum (the test dataset) and the esteem of power obtained
through NN2 model. For confidentiality constraints all variables are normalized with respect to their
maximum value in the day and the estimate of turbine outlet steam quality cannot be reported.

Table 1. Test results for HP Steam Turbine Models.

Model Section Output var. Description eav EM EP STD


Wimp [kW] Power of impulsive drum 466 606 34 108
Impulse Timp [°C] Temperature at exit of impulsive drum 2.6 3.2 0.9 0.5
Pimp [BarA] Pressure at exit of impulsive drum 2.5 3.3 4.7 0.6
WD1 [kW] Power of drum 1 143 203 6.9 42
D1 TD1 [°C] Temperature at exit of drum 1 2.5 3.3 1.1 0.6
Thermodynamic PD1 [BarA] Pressure at exit of drum 1 1.8 2.6 4.9 0.5
model WD2 [kW] Power of drum 2 143 204 4.6 46
D2 TD2 [°C] Temperature at exit of drum 2 1.3 2.5 1.0 0.9
PD2 [BarA] Pressure at exit of drum 2 1.0 4.9 4.8 0.3
WD3 [kW] Power of drum 3 135 267 5.4 65
D3
TD3 [°C] Temperature at exit of drum 3 0.4 3.5 1.9 0.8
Overall Woverall [kW] Overall Power 571 704 5.9 130
NN1 Overall Woverall [kW] Overall Power 7.8 44 0.4 9.3
Impulse Wimp [kW] Power of impulsive drum 2.1 6.1 0.9 2.2
D1 WD1 [kW] Power of drum 1 2.2 11 0.7 2.7
NN2 D2 WD2 [kW] Power of drum 2 1.8 9.2 0.3 2.1
D3 WD3 [kW] Power of drum 3 5.3 23 1.2 5.3
Overall Woverall Overall Power 8.3 44 0.5 10
Pimp [Bar A] Pressure at exit of impulsive drum 0.02 0.2 0.8 0.03
Impulse
Timp [°C] Temperature at exit of impulsive drum 0.05 0.3 0.1 0.04
PD1 [Bar A] Pressure at exit of drum 1 0.04 0.1 1 0.04
D1
NN3 TD1[°C] Temperature at exit of drum 1 0.1 0.6 0.2 0.1
PD2 [Bar A] Pressure at exit of drum 2 0.03 0.08 1.3 0.02
D2
TD2 [°C] Temperature at exit of drum 2 0.08 0.7 0.3 0.1
D3 TD3 [°C] Temperature at exit of drum 3 0.1 1.4 0.8 0.2
Turbine
NN4 xout Steam quality at turbine outlet 2e-4 1e-3 0.1 3e-4
Outlet

5. Conclusions

Two different modelling approaches are presented, which provide an estimate of the power generated
by the overall turbine or by each drum, of the temperatures and pressures at exit of each drum and steam
quality at turbine outlet. Both models were trained with datasets created with GE Turbomachinery design
tool but were tested also with field data measured in a real power plant. The results are very interesting
S. Dettori et al. / Energy Procedia 105 (2017) 524 – 529 529

and encouraging: in particular, the NN-based models provide satisfactory results also considering
extremely off-design conditions. On the other hand, the hybrid thermodynamic model incorporates the
monitoring of windage effect, which is highlighted, but its performances are not fully satisfactory.
Future work will focus on designing of more accurate algorithms for the Hybrid thermodynamic
model, and to explore more complex network structures and hybrid approaches.

Fig. 3. Estimates of Pressures and temperatures for HP ST (left) and Power of each drum of HP ST (right) .

References

[1] Dixon, S. Larry, and Cesare Hall: Fluid mechanics and thermodynamics of turbomachinery. Butterworth-Heinemann, 2013.
[2] Medina-Flores J.M., Picn-Nez M. Modelling the power production of single and multiple extraction steam turbines. Chemical
Engineering Science 65.9; 2010, p. 2811-2820.
[3] Chaibakhsh Ali, Ali Ghaffari. Steam turbine model. Simulation Modelling Practice and Theory 16.9; 2008, p. 1145-1162.
[4] Varbanov P. S., S. Doyle, R. Smith. Modelling and optimization of utility systems. Chemical Engineering Research and
Design 82.5; 2004, p. 561-578.
[5] Ray, Asok. Dynamic modelling of power plant turbines for controller design. Appl. Mat. Modelling 4.2; 1980, p. 109-112.
[6] Lu S., B. W. Hogg. Dynamic nonlinear modelling of power plant by physical principles and neural networks. International
Journal of Electrical Power & Energy Systems 22.1; 2000, p. 67-78.
[7] Wagner, W., et al. The IAPWS industrial formulation 1997 for the thermodynamic properties of water and steam. Journal of
Engineering for Gas Turbines and Power 122.1; 2000, p. 150-184.
[8] M. T. Hagan and M. B. Menhaj. Training feedforward networks with the marquardt algorithm, IEEE Transactions on Neural
Networks, vol. 5, no. 6; 1994, p. 989-993.

Biography
Valentina Colla is currently responsible of the Center of ICT for Complex Industrial
Systems and Process of the TeCIP Institute. Her expertise concerns simulation,
modelling, control and optimization technologies for industrial applications as well as
data processing and mining through traditional and AI-based techniques.

You might also like