You are on page 1of 7

Prediction of Power Output of a Combined Cycle Power Plant using

Multi-layer Feed Forward Neural Network

Md. Zia Khurshid

Roll 17ME91R25

Mobile 980511851

Email ziakhurshid1@gmail.com

Department Mechanical Engineering

Abstract Full load electrical power output prediction is very important for maximizing profit and
saving resources. This project uses multi-layer feed forward neural network to develop a predictive
model which can predict hourly full load electrical power output of a combined cycle power plant. The
base load operation of a power plant is influenced by four main parameters, which are used as input
variables in the dataset, such as ambient temperature, atmospheric pressure, relative humidity, and
exhaust steam pressure. These parameters affect electrical power output, which is considered as the
target variable. The data set is collected from a paper and is collected over a six-year period.
Levenberg-Marquardt algorithm has been used to update the weight.

Keywords neural network; combined cycle power plant; multi-layer feed forward neural
network; levenberg-marquardt method; energy output prediction

1. Introduction

Predicting thermodynamical processes in real life is an extremely different task


consisting of solving thousands of nonlinear equations and it takes too much time and effort to
solve. This problem is solved with the advent of soft computing, as soft computing is getting
better and better, efficient and more accurate predictions of difficult systems is becoming easier
if sufficient practical data is available [1].

Soft computing is widely used for regression analysis by which the output of complex
systems is predicted. Many real-life problems can be solved as regression problems, and
evaluated using machine learning approaches to develop predictive models [2].

In this project prediction analysis of combined cycle power plant (CCPP) has been done
which has two gas turbines, one steam turbine, and two heat recovery systems. Predicting
electrical power output of a power plant has been considered a critical real-life problem to
construct a model using neural network. A power plant cannot be operated efficiently without
the prediction of full load power output. It is useful in order to maximize the income from the
available megawatt hours (MW h). For a gas turbine to be reliable and sustainable, the power
generation prediction should be accurate.

Gas turbine power output primarily depends on the ambient parameters which are
ambient temperature, atmospheric pressure, and relative humidity. Steam turbine power output
has a direct relationship with vacuum at exhaust. Previously artificial neural network has been
used in power output prediction. The effects of ambient-pressure and temperature, relative
humidity, wind-velocity and direction on the plant power are investigated based on the
measured data from the plant [1]. Effect of varying local ambient conditions is analyzed to
predict the operational and performance parameters of a gas turbine [3]. Not only full load but
steam turbine behavior at other operating points have also been analyzed [4] even at startup
stage [5]. Neural Network has been proven to be a better predictor model than dynamic linear
models [6]. Neural network has also been used for performance analysis, anomaly detection,
fault detection and isolation of gas turbine engines [7][8].

2. Aims & Objectives

To predict the energy output of a combined cycle power plant.

3. Data Collection

Data of the Combined Cycle Power Plant has been collected in the span of 6 years, from 2006-
2011when the power plant is set to full load [9]. The input of the data consists of hourly average
ambient variables temperature (T), ambient pressure (AP), relative humidity (RH) and exhaust
vacuum (V) to predict the net electrical energy output (EP). A two fold data set is created which
is shuffled five times. Only 200 values from the original dataset have been taken. The data has
been normalized so that all data lies between (0.0, 1.0).

4. Proposed Technique

The analysis is performed in MATLAB using the Neural Network toolbox. The method used is
multi-layer feed forward neural network in which it was assumed to have an input layer having
4 inputs, only 1 hidden layer and an output layer with only one output as shown in Figure 1.
Figure 1: Neural network of the problem with N number of hidden neurons

Also, a set of 50 data is set apart to test the validity of the chosen neural network. The transfer
functions used at the input is linear,

All connections from input to hidden layer has a weight summed in a matrix as

11 1 1
21 2 2
[] = [ 3 3 ]
31
41 4 4

1


Similarly, all connections from hidden layer to output has a weight summed as [] =

[ ]

Now the input to the j-th hidden neuron is given by,

= 1 1 + 2 2 + 3 3 + 4 4

The hidden layer has tan-sigmoid transfer function given by,



= , where is the co-efficient of the transfer function.
+

The output of the network is given by, = 1 1 + 2 2 + 3 3 + 4 4

For updating the weights of the network, Levenberg-Marquardt algorithm. It is a non-linear


iterative optimization algorithm given by,
+1 = [ + ]1

Where, is Jacobian matrix, is identity matrix, is eigen value and is the adaptive value.
When the scalar is zero, this is just Newtons method. When is large, this becomes gradient
descent with a small step size. Newtons method is faster and more accurate near an error
minimum, so the aim is to shift toward Newtons method as quickly as possible. Thus, is
decreased after each successful step (reduction in performance function) and is increased only
when a tentative step would increase the performance function. In this way, the performance
function is always reduced at each iteration of the algorithm.

Now neural networks are made with hidden layers ranging from 2 to 20 and the input data is
fed to them, out of 200 data, 70% is kept as training data while 15% as test data 15% as
validation data. is set as 0.0001.

Termination criteria is set such as any of this occur:

The maximum number of epochs (repetitions) is reached, set to 100.


The performance gradient falls below min_grad, set to 10-7.
Validation performance has increased more than max_fail times since the last time it
decreased (when using validation), set to 6.

The data obtained is given in Table X.

No. of Hidden Mean Square Epoch


Layers Error (X 10-5)
2 1.2276 8
3 1.5044 3
4 1.9490 3
5 1.4047 5
6 2.3508 2
7 1.0619 2
8 1.0287 4
9 1.5618 9
10 0.8629 9
11 1.9678 11
12 1.8406 5
13 1.7113 10
14 1.8478 12
15 3.0031 3
16 3.2547 2
17 1.9312 2
18 3.1310 2
19 2.1110 4
20 1.4111 4

Table 1: Mean Square Error for each number of hidden neurons

From Table 1 it can be seen that the least mean square error is when the hidden layer has 10
neurons for that number is finalized.

The 50 valued data set aside for simulation is then fed into the network of 10 hidden neurons.
Now if be the target output then the total error in prediction is given by,

1 ||
= 50
=1
100, where is the number of data used for simulation.

By feeding the simulation data we get the results shown in Table 2,

Experimental Predicted | | Experimental Predicted | |


Output (P) Output (O) 100 Output (P) Output (O) 100

0.44407 0.52188 17.52201 0.46216 0.52188 12.92193
0.46957 0.52188 11.13998 0.47105 0.52188 10.79079
0.45989 0.52188 13.47931 0.48471 0.52188 7.668503
0.47959 0.52188 8.817949 0.44634 0.52188 16.92432
0.44092 0.52188 18.36161 0.46902 0.52188 11.27031
0.48087 0.52188 8.528292 0.43212 0.52188 20.77201
0.44190 0.52188 18.09912 0.46728 0.52188 11.68464
0.43020 0.52188 21.31102 0.42966 0.52188 21.46348
0.46516 0.52188 12.19365 0.46949 0.52188 11.15892
0.47132 0.52188 10.72732 0.48587 0.52188 7.411448
0.48543 0.52188 7.508807 0.48195 0.52188 8.285092
0.49535 0.52188 5.355809 0.47903 0.52188 8.94516
0.44912 0.52188 16.20057 0.43450 0.52188 20.11047
0.48053 0.52188 8.605082 0.46490 0.52188 12.25640
0.45707 0.52188 14.17945 0.45271 0.52188 15.27910
0.44367 0.52188 17.62797 0.42974 0.52188 21.44087
0.47752 0.52188 9.289663 0.45709 0.52188 14.17445
0.47295 0.52188 10.34570 0.44677 0.52188 16.81178
0.47254 0.52188 10.44144 0.46076 0.52188 13.26504
0.46917 0.52188 11.23473 0.47195 0.52188 10.57951
0.43521 0.52188 19.91452 0.45329 0.52188 15.13159
0.47778 0.52188 9.230190 0.44161 0.52188 18.17667
0.47589 0.52188 9.663998 0.46473 0.52188 12.29746
0.48390 0.52188 7.848729 0.46468 0.52188 12.30955
0.47620 0.52188 9.592608 0.43059 0.52188 21.20114

Table 2: Predicted output (O) vs experimental output (P)

5. Results and discussion

The total error E in the simulation turned out to be 13.191%.

This shows that the multi-layer feed forward neural network using Levenberg-Marquardt
algorithm can give good enough predictions for the energy output of combined cycle power
plant.

In this project different algorithms have been tested and Levenberg-Marquardt algorithm has
been found to be the most suitable and the number of hidden neurons in the layer is also checked
and fount that for this data set 10 neurons in the hidden layer is the optimum choice.

6. Conclusion

In place of thermodynamical approaches, involving so many nonlinear equations for real life
application, soft computing approach is taken here for the prediction of electrical power output
of a base load operated combined cycle power plant at full load for accurate results because its
saves time and effort. In comparison thermodynamical approaches give unreliable and
unsatisfactory results due to many assumptions taken into account and nonlinearity of
equations.

7. References

[1] Kesgin U, Heperkan H. Simulation of thermodynamic systems using soft computing


techniques. Int J Energy Res 2005;29:581611.

[2] Gvenir HA. Regression on feature projections. Knowl-Based Syst 2000;13:20714.


[3] Fast M, Assadi M, Deb S. Development and multi-utility of an ANN model for an
industrial gas turbine. Appl Energy 2009;86(1):917.

[4] Rahnama M, Ghorbani H, Montazeri A. Nonlinear identification of a gas turbine


system in transient operation mode using neural network. In: 4th Conference on thermal
power plants (CTPP), IEEE Xplore; 2012.

[5] Refan MH, Taghavi SH, Afshar A. Identification of heavy duty gas turbine startup
mode by neural networks. In: 4th Conference on thermal power plants (CTPP), IEEE
Xplore; 2012.

[6] Yari M, Shoorehdeli MA. V94.2 gas turbine identification using neural network. In:
First RSI/ISM international conference on robotics and mechatronics (ICRoM), IEEE
Xplore, 2013.

[7] Kumar A, Srivastava A, Banerjee A, Goel A. Performance based anomaly detection


analysis of a gas turbine engine by artificial neural network approach. In: Procee.
annual conference of the prognostics and health management society; 2012.

[8] Tayarani-Bathaie SS, Sadough Vanini ZN, Khorasan K. Dynamic neural network-
based fault diagnosis of gas turbine engines. Neurocomputing 2014;125(11):15365.

[9] http://archive.ics.uci.edu/ml/datasets/combined+cycle+power+plant

You might also like