Professional Documents
Culture Documents
ScienceDirect
j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / i j r e f r i g
A R T I C L E I N F O A B S T R A C T
Article history: This work presents the empirical study of a reciprocating compressor using Artificial In-
Received 22 May 2015 telligence to model it. Several artificial neural networks were used to model three energy
Received in revised form 10 August parameters of the compressor with high precision. The number of neurons in each ANN
2015 was optimized to use the minimum number of neurons without compromising accuracy;
Accepted 12 August 2015 very few neurons were used when comparing with other works. Computer simulations show
Available online 18 August 2015 that the ANN model for the mass flow rate has the highest accuracy when compared with
the models for the discharge temperature and power consumption. These simulations also
Keywords: illustrate that the ANN model for the discharge temperature presented the lowest accuracy.
Artificial neural network Using the ANN model, 3D plots were built to analyze the energy behavior of the compressor.
Vapor compression system © 2015 Elsevier Ltd and International Institute of Refrigeration. All rights reserved.
Neuron
Energy
Simulated annealing
Optimization
damage the ozone layer, see Secan (2011). Li et al. (2012) de-
Nomenclature veloped an ANN-based dynamic model for the experimental
variable speed direct expansion air conditioning system.
h Neural network weight in the hidden layer
Another model using steady state input data of an automo-
ṁ Mass flow rate [kg s−1]
N Rotation compressor speed [rpm] tive air conditioning system was proposed by Haslinda et al.
(2013). Belman-Flores et al. (2013) modeled a variable speed
Pdischarge Discharge pressure [kPa]
Power Power consumption [kW] vapor compression system using R134a; this method accu-
Psuction Suction pressure [kPa] rately estimates the number of neurons in the hidden layer,
Tdischarge Discharge temperature [K] and the model can predict the energetic performance with good
Tsuction Suction temperature [K] accuracy. Later, the authors built energetic maps using the re-
w Neural network weight in the output layer frigerant R1234yf, see Ledesma and Belman-Flores (2014).
In the literature, there are few publications related to arti-
xi Neuron input
ficial intelligence applied to reciprocating compressors. For
y Activation function input
instance, Yang et al. (2009) developed a neural network loss–
z Neuron output
efficiency model to simulate the performance of positive
μ Number of neurons in the hidden layer
displacement compressors. Yang et al. (2005) created a practical
classification system to select faulty products at semi-product
stage in an automatic mass production line of reciprocating
analysis (Barroso-Maldonado et al., 2015; Farzaneh-Gord et al., compressors. The extracted features from the original and
2014; Navarro et al., 2007; Yang et al., 2013), using semi- wavelet-transformed signals are used as inputs to both: clas-
empirical methods (Wenhua, 2013), through dynamic simulation sifiers based on ANNs, and SVM. Qin et al. (2012) presented a
(Castaing-Lasvignottes and Gibout, 2010; Ndiaye and Bernier, scheme for fault detection of reciprocating compressor valves
2010), analysis with a computational fluid dynamics, CFD established on: basis pursuit, wave matching, and SVM.
(Pereira et al., 2008), development of a new generic network Ghorbanian and Gholamrezaei (2009) designed a fast, reliable
model (Hu et al., 2014). and robust neural network tool to smoothly predict a com-
Based on this literature review, it may be said that differ- pressor performance map at design stage. The capability of
ent topologies of reciprocating compressor models exist with different types of ANNs was investigated. One paper in the lit-
diverse levels of complexity. Models with physical founda- erature applies an ANN to model a rotary type compressor see
tions are basic, other models use correlations but with no Sanaye et al. (2011); they proposed two ANNs and non-linear
physical meaning and cannot be applied outside the speci- regression models to analyze the experimental data, for pre-
fied range of operating conditions. In terms of applications, dicting the refrigerant mass flow rate and the refrigerant
these models go beyond design, optimization, or simple pre- discharge temperature for R134a. In addition, they used error
diction of the energy behavior of the compressor. Additionally, analysis methods to compared their models. Tian et al. (2015)
the use of CFD has been a countless tool in the analysis of these applied a hybrid ANN, and Partial Least Squares, PLS, regres-
systems. Among the most relevant parameters to consider are: sion to predict the thermodynamic performance of a scroll
refrigerant mass flow rate, input power, and discharge tem- compressor.
perature of the refrigerant. Based on the review of the state of the art, the main con-
However, there has been a recent influx of modern tech- tributions of this paper into this field are:
niques for modeling. In recent years, the application of artificial
intelligence in different fields of mechanical engineering has • Extend the use of ANNs to analyze and model reciprocat-
been increasing. For instance, the use of Artificial Neural Net- ing compressors that are employed in refrigeration and air
works, ANNs, in thermal analysis of heat exchanger (Monharaj conditioning systems.
et al., 2015), solar thermal energy (Yaïci and Entchev, 2014), and • Instead of using one single ANN to model the compressor,
in particular way according to this work, the application in the each energy parameter was separately modeled in order to
field of refrigeration, air conditioning and heat pump systems analyze which parameter can be best characterized by the
(Mohanraj et al., 2012). Also, the application of ANNs, Support model. These parameters are: mass flow rate, discharge tem-
Vector Machines, SVMs, Classification and Regression Tree, CART, perature, and power consumption.
Multiple Regression, MR, Generalized Linear Regression, GLR, • The number of neurons in the hidden layer was opti-
and Chi-squared Automatic Interaction Detector, CHAID, were mized to model the compressor and avoid machine learning
applied for predicting refrigeration system performance, see over-fitting.
Chou et al. (2014). Esen and Inalli (2010) compared an ANN and • 3D plots were build using the ANN model to analyze the
Adaptive Neuro-Fuzzy Inference System, ANFIS, for predic- energy behavior of the compressor.
tion performance a heat pump system. Secan et al. (2011) used
an ANN to determine liquid and vapor thermophysical prop-
erties of alternative refrigerants. Also, the applicability of an
ANN to predict the performance of an experimental ground- 2. Description of the experimental facility
coupled heat pump system has been studied (Esen and Inalli,
2009; Esen et al., 2008). Other authors applied ANNs as a comple- The experimental facility of the vapor compression system is
ment to investigate the performance of a single-stage composed of: a compressor, a condenser, an expansion valve,
refrigeration system using different refrigerants which do not and an evaporator. The installation is setup to control the
192 international journal of refrigeration 59 (2015) 190–197
x1 Neuron
Electric Motor w11
w12 y
x2 ∑ f(y) z
w13
Inductive Sensor w
x3 bias
1
Fig. 2 – Internal structure of a neuron.
Psuction h22
hμ2 w2 . -6
mse
Hidden x Output
Neuron 2 2 Neuron m
h13
. training-mse
h2B . .
Tsuction h23
1 . wB
hμ3
. 1
h14
h24 . -7
Pdischarge hμ 4 wμ
xμ 0 3 6 9 12 15
Hidden Number of Neurons in Hidden Layer ( )
1 Neuron μ
hμB
Fig. 6 – Progress of the mse for the mass flow rate of the
compressor.
Fig. 5 – ANN to model the mass flow rate of the
compressor.
0.5
0
validation-mse
mse
-0.5
training-mse
-1
0 3 6 9 12 15
Fig. 8 – Effect of N and Psuction on the mass flow rate.
Number of Neurons in Hidden Layer ( )