Professional Documents
Culture Documents
Research Paper
Performance evaluation and prediction for electric vehicle heat pump using T
machine learning method
⁎
Yufeng Wanga, Wanyong Lia, Ziqi Zhanga, Junye Shia,b, Jiangping Chena,b,
a
Institute of Refrigeration and Cryogenics, Shanghai Jiaotong University, Shanghai, China
b
Shanghai High Efficient Cooling System Research Center, Shanghai, China
A R T I C LE I N FO A B S T R A C T
Keywords: A method to predict the performance of R134a heat pump with EVI (Economized Vapor Injection), is proposed in
Electrical vehicle the present study. Models using SVR (Support Vector Regression) as the base estimator and Adaboost.R2 as the
Heat pump ensemble method, are established to predict the heating capacity and COP of the heat pump. Different feature
Refrigerant injection sets for the model input are formed, based on the working principle of the heat pump system and correlation
Machine learning
analysis. Parameters of the models are optimized to improve prediction performance. The simulation results are
SVR
compared with the experimental results, and the relative errors for heating capacity and COP prediction are
Adaboost.R2
within 8.5%. Moreover, the impacts of injection pressure on the EVI heat pump system are discussed and si-
mulated using the model established. The optimum injection pressure of the heat pump system can be obtained
from the model under different working conditions.
⁎
Corresponding author at: Institute of Refrigeration and Cryogenics, Shanghai Jiaotong University, Shanghai, China.
E-mail address: jpchen_sjtu@163.com (J. Chen).
https://doi.org/10.1016/j.applthermaleng.2019.113901
Received 1 January 2019; Received in revised form 7 May 2019; Accepted 1 June 2019
Available online 01 June 2019
1359-4311/ © 2019 Elsevier Ltd. All rights reserved.
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
Subscripts
air air-side
Rasmussen [15] used Linear regression, random forest, and SVR (Sup- [20,21]. However, former applications of machine learning method in
port Vector Regression) algorithms to forecast the electrical load of the refrigeration field mainly focus on domestic or commercial re-
supermarket refrigeration systems. Shi [16] developed a model for frigeration and heating systems. When it comes to automotive air
detecting refrigerant charge fault using Bayesian neural network. Guo conditioning or heat pump, few researches have been found.
[17] compared the performance of SVR, BPNN (Back Propagation In this study, models to predict the heating capacity and COP of
Neural Network), ELM (Extreme Learning Machine) algorithms in R134a EVI heat pump for electric vehicles are established. The models
forecasting energy demand of building heating system. Apart from use SVR as the base estimator and Adaboost.R2 as the ensemble
predicting system performances, the machine learning method is also method. First, different feature sets for the model input are formed
capable of investigating the detailed working status of the refrigeration based on the working principle of the heat pump system and correlation
system [18,19] and finding an optimal control strategy for the system analysis. Then parameters of the models are optimized to improve
prediction performance. The simulation results computed from the
optimized model are compared with the experimental result. Finally,
the impacts of injection pressure on the EVI heat pump system are
discussed and simulated using the model established. Hence optimum
injection pressure of the heat pump system can be achieved from the
model.
2. Experimental setup
The structure of the EVI heat pump system is shown in Fig. 2. The
system is capable of both air conditioning and heating. When in AC (Air
Conditioning) mode, condenser inside the HVAC (Heating, Ventilation,
and Air Conditioning) module is blocked on the air side. BV1 is open
and the refrigerant flows through it without throttling. The outside HX
functions as a condenser. BV3 is closed and the refrigerant has to be
Fig. 1. Schematic diagram of EVI heat pump system [2]. throttled by EXV3. The inner evaporator is in charge of cooling the air
2
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
entering HVAC. When in heating mode, BV1 is closed and the re- Uncertainty analysis is performed in order to verify the measured data
frigerant is throttled by EXV1. The outside HX functions as an eva- of the cooling/heating capacity and COP using Eq. (5) [22]. Propagated
porator this time. The opening of BV3 and EXV3 depend on whether uncertainties are estimated while calculating the final results from the
dehumidification is required. When BV2 is closed, there is no vapor experimental measurements. Overall uncertainty is represented as δE,
injection into the compressor. When BV2 is open, the refrigerant split and δXi represents the uncertainty of its affecting factors. After calcu-
from the condenser outlet is first throttled by EXV2 and then heated by lation, the relative uncertainties for Q and COP are 5.5% and 6.3%
the economizer. The opening of EXV2 can control the injection pressure respectively.
and mass flow rate into the compressor. In our case, we investigated the
heat pump performance with EVI working. δE N ∂E
E
= 2 ∑i =1 ( ∂Xi δXi )2 (5)
2.2. Test facility
2.3. Test conditions
The test facility is shown in Fig. 2. The inner condenser and inner
evaporator are packaged in an HVAC module, installed in the indoor
The test conditions are designed to investigate the EVI heat pump
chamber. Compressor, separator, and outside HX are installed in the
system under different outdoor temperatures, indoor temperatures, and
outdoor chamber. The location of sensors is shown in Fig. 2. The ca-
compressor speeds. There are totally 14 test conditions, as is concluded
pacity (Q) of the tested EVI heat pump was determined by the average
in Table 3. In each condition, the injection pressure is adjusted by EXV2
of air-side and refrigerant side heat transfer rate using Eq. (1). The air
to 5 levels to investigate the impact of injection pressure on the system
side heat transfer was calculated using Eq. (2). The refrigerant side heat
heating performance.
transfer rate was obtained by the enthalpy difference calculation, using
Eq. (3). The properties of the R134a refrigerant were calculated ac-
cording to the NIST REFPROP 9.1. The work consumption is the pro- 3. Methodology
duct of compressor voltage and input current, and the overall system
COP was determined by Eq. (4). Analysis of the experimental results 3.1. Feature selection method
indicated that the errors between the heat transfer rate for the air side
and the refrigerant side were within ± 5%. Choosing appropriate features as the input is vital for the machine
3
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
Table 2
Component information.
Components Information
Table 3
Test conditions.
Index Outside HX facing velocity (m/s) Indoor air volume (m3/h) Tout door (°C) Tin door (°C) Compressor speed (RPM) Injection pressure level
1 4 300 0 0 5500 5
2 0 20 5500
3 −10 0 4000/5500/7000
4 −10 20 4000/5500/7000
5 −20 −7 4000/5500/7000
6 −20 20 4000/5500/7000
4
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
MFR_inj
R134a Q P_inj
R134a COP
T_sc T_inj
ʌ_ inj Q
T_sh,inj EXV2-Opening
COP T_suc
T_dis T_sc
ij_inj EXV1-Opening
T_cond,out P_suc
EXV1-Opening T_sh,inj
ʌ_ main P_evap,in
P_dis ij_inj
EXV2-Opening MFR_main
T_inj T_sh,dis
T_sh,dis MFR_inj
P_inj ʌ_ inj
T_suc T_dis
MFR_main P_dis
P_suc T_cond,out
P_evap,in ʌ_ main
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Absolute Correlation Coefficient Absolute Correlation Coefficient
Table 4
Feature sets for the model input.
Name Including features Feature numbers Collected feature numbers
0.25
1.00
0.20
0.95
RMSE
Score R134a Test
0.90 RMSE R134a Train
RMSE R134a Test 0.10
0.85
0.05
0.80 0.00
1 2 3 4 5
Feature Sets
Fig. 6. Performances of different feature sets for (a) heating capacity prediction
and (b) COP prediction (b).
5
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
Simulation COP
2.5 2.5
2.0 2.0
1.5 1.5
1.0 1.0
1.0 1.5 2.0 2.5 3.0 1.0 1.5 2.0 2.5 3.0
Experiment Capacity (kW) Experiment COP
Fig. 7. Comparison between simulation result and experiment result (a) Heating capacity (b) COP.
0.5 50
They have to be fine-tuned to improve model performance. The C
2.4 Q(kW) parameter is regarded as the regularization coefficient. The γ parameter
ij_inj defines how far the influence of a single training example reaches, with
0.4 45
T_cond,in (
small values meaning ‘far’ and large values meaning ‘close’. The larger
C and larger γ have the same influence on the model: more support
T_cond,in (°C)
2.0 0.3 40 vectors, more complexity, and more chances of overfitting. If they are
Q (kW)
ij_inj
too small, the model is too ‘flat’, and underfitting occurs. The ε para-
meter defines the maximum deviation that can be tolerated by the SVR.
0.2 35
When ε is small, it has the same effect as a large C and large γ .
1.6
0.1 30
3.2.2. Adaboost.R2 algorithm
Boosting is an ensemble method often used in machine learning to
1.2 0.0 25 build a committee of regressors that may be superior to a single re-
0.20 0.25 0.30 0.35 0.40 gressor. Adaboost.R2 [26] is a famous boosting algorithm for regres-
Injection Pressure (Mpa) sion, and it is applied in this paper for regression tasks.
Briefly, the Adaboost.R2 algorithm is intended to adjust the weights
Fig. 8. Effect of injection pressure on the heat pump system (@-10/20 ℃,
4000RPM).
for training samples. If the predicted value of a certain sample deviates
from its true value, then the weight of it will be increased in the next
round, making the sample paid more attention to by the model. The
m ∗
⎧ ∑i = 1 (αi −αi ) = 0 ⎫ machines are trained sequentially in this algorithm. The calculating
Constraints:
⎨ 0 ≤ αi∗,αi ≤ C ⎬ procedure can be expressed as followed.
⎩ ⎭
In the first round, the weight of each sample in the training set is
And the function that we obtain from the optimization problem is: initialized as 1. And the probability that training sample i is in the
training set is regarded as pi.
m
f (x ) = ∑ (αi∗ −αi) k〈xi , x〉 + b wi = 1i = 1, 2, ⋯, m (16)
i=1 (15)
wi
pi =
The penalty parameter C, the deviation ε and the kernel coefficient m
∑i = 1 wi (17)
γ are 3 important hyper-parameters during the SVR calculating process.
EXP@4000RPM EXP@4000RPM
SIM@4000RPM SIM@4000RPM
3.2
EXP@5500RPM (a) EXP@5500RPM (b)
SIM@5500RPM 2.4 SIM@5500RPM
EXP@7000RPM EXP@7000RPM
2.8
SIM@7000RPM SIM@7000RPM
2.0
2.4
Q (kW)
Q (kW)
2.0
1.6
1.6
1.2
1.2
0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.10 0.15 0.20 0.25 0.30 0.35
P_inj (Mpa) P_inj (Mpa)
Fig. 9. Effect of injection pressure on the heating performance (a) @-10/20 ℃ (b) @-20/20 ℃.
6
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
A machine is trained using the weights. And a mapping from input prevent one feature from far outweighing another one. In regression
features xi to outcome f(xi) is established using the machine in each tasks, standardization can help reduce multicollinearity issues for
round. The loss of each training sample is Li, and the maximum de- models containing interaction terms.
viation of all the training data is D. And linear loss function is used in
this study. 4.2. Feature selection & model optimization
D = sup |yi − f (x i )| i = 1, 2, ⋯, m (18)
The method for feature selection can be referred to in Section 3.1.
|y − f (x i )| Five feature sets as described in Table 4 are selected as the model input
Li = i
D (19) to test their performances. The model used for each feature set is op-
The average loss L̄ and the weight for each machine β can be de- timized, and the optimization target is maximum r2score of the test set.
termined. The optimized parameters are C, ε, γ and n_round respectively, and the
optimization range of them are presented in Table 5.
m
L¯ = Fig. 6 shows the performances of different feature sets for heating
∑ Li pi
i=1 (20) capacity and COP prediction. A higher r2score and a lower RMSE suggest
better prediction performance. As indicated in Fig. 6, with growing
L¯ numbers of feature input from FS1 to FS5, the r2score of the train set
β=
1 − L¯ (21) tend to increase, which states that more feature input is likely to in-
In the next round, the weight for each sample can be updated, and crease the ability of the model to fit the train set. In heating capacity
the new weights are used to train a new machine. prediction, the r2score of FS2 is the highest. In COP prediction, the r2score
of FS3 is the highest, but the r2score of FS2 is only a bit lower. This
wi = wi β1 − Li (22) suggests that FS2 is an appropriate feature set for both heating capacity
The Adaboost.R2 algorithm will not stop until the maximum and COP prediction. Also, FS2 requires the least features that have to be
training rounds n_round (exactly the number of trained machines) that collected from the system (same as FS1) among these feature sets,
we set is reached or L̄ exceeds 0.5. The final prediction of the model is which makes it more applicable when it comes to practical use. A larger
the weighed median of all the values predicted by machines, instead of number of features seems disadvantageous to the model performance
the weighted mean. for the test set. This may be attributed to the fact that more features
The reason why we choose Adaboost.R2 as the ensemble method in increase the model complexity, and overfitting may occur. The opti-
performance prediction of refrigeration or heat pump system is that the mized parameters (C, ε, γ and n_round) for the heating capacity pre-
system operates under very different conditions. Data obtained under diction model (FS2 as the input) are 20, 0.01, 0.02 and 25. For the COP
regular conditions (regular compressor speed, regular ambient tem- prediction model, parameters are 5, 0.01, 0.05 and 25.
perature, etc.) are much more than those obtained under irregular ones.
The Adaboost.R2 algorithm can cope with data nonuniformity and 5. Result and discussion
sparsity issues. With the application of Adaboost.R2, data obtained
under irregular conditions can be more attended to, and the overall 5.1. Comparison with the experiment results
prediction errors can be decreased.
A comparison between the simulation results and experiment results
3.3. Model performance evaluation (including both train set and test set) is carried out. As illustrated in
Fig. 7, the simulation results are in good agreement with the experi-
Two performance indexes are used to evaluate model performance ment results. The maximum errors for capacity and COP prediction are
in this study, namely r2score and RMSE (Root Mean Square Error). 8.25% and 8.33% respectively. This indicates that the models estab-
m lished are applicable under almost all the experiment conditions, which
∑i = 1 (f (x i ) − yi )2
r 2score = 1 − m
are quite complex.
∑i = 1 (y¯ − yi )2 (23)
m 5.2. Effect of the injection pressure
∑i = 1 (f (x i ) − yi )2
RMSE =
m (24) The major difference between a conventional heat pump and an EVI
The r2score is less than 1. The closer value of r2score to 1, the better heat pump lies in the second loop described in Fig. 3. Hence, the in-
model regression performance is achieved. Meanwhile, the smaller the jection parameters along the second loop should be studied to under-
value of RMSE also indicates better model prediction performance. stand their influence on the system performance. Among them, the most
critical one is the injection pressure Pinj [27]. The injection pressure has
4. Machine learning modeling an impact on the condenser inlet temperature, as well as the injection
mass flow rate. The experiment shows the effect of the injection pres-
The machine learning procedure can be summarized as the flow sure, as described in Fig. 8. φinj indicates the injection ratio, and the
chart shown in Fig. 5. expression of it is shown in Eq. (25).
φinj = MFRinj/MFRmain (25)
4.1. Data preprocessing
The expansion valve on the second loop (EXV2 in Fig. 2) controls
As is described in Section 2.3, there are totally 14 test conditions. the injection pressure. When the opening of the valve is larger, the
Under each test conditions, the injection pressure is adjusted to 5 levels. injection pressure grows, and as a result the injection mass flow rate
Therefore, 70 samples are available. 80% of the samples are randomly increases, which is beneficial to the heating capacity. A larger injection
separated into the train set, and the remaining 20% are separated into ratio is also observed due to this effect. However, as the injection ratio
the test set. grows to some point, the compressor may be further cooled down and
Data standardization is needed before we can proceed with machine the discharge temperature decreases. This is disadvantageous to the
learning modeling. Standardization means to rescale your data to have heating capacity because the temperature difference between the re-
a mean of zero and a standard deviation of one. Standardization can frigerant side and air side in the condenser is hence decreased.
7
Y. Wang, et al. Applied Thermal Engineering 159 (2019) 113901
Therefore, there exists an optimum injection pressure, where the system temperatures, Appl. Therm. Eng. 142 (2018) 656–664.
heating capacity reaches its maximum. [5] W. Yufeng, W. Dandong, Y. Binbin, S. Junye, C. Jiangping, Experimental and nu-
merical investigation of a CO2 heat pump system for electrical vehicle with Series
However, it’s difficult to find the optimum injection pressure be- Gas Cooler configuration, Int. J. Refrig. 100 (2018).
cause it varies greatly under different working conditions. Also, the [6] Y.U. Choi, M.S. Kim, G.T. Kim, M. Kim, M.S. Kim, Performance analysis of vapor
working conditions of EVI are quite complex, and the operating para- injection heat pump system for electric vehicle in cold startup condition, Int. J.
Refrig. 80 (2017).
meters of the system are heavily coupled, making us hard to find which [7] Z. Zhang, D. Wang, C. Zhang, J. Chen, Electric vehicle range extension strategies
parameters we should count on to find the optimum injection pressure. based on improved ac system in cold climate – a review, Int. J. Refrig. (2018)
The model established in this study may find a solution to this problem. S0140700718300033.
[8] F. Qin, G. Zhang, Q. Xue, H. Zou, C. Tian, Experimental investigation and theore-
As illustrated in Fig. 9, the simulated results of injection pressure’s ef- tical analysis of heat pump systems with two different injection portholes com-
fect share a similar trend with the experiment results: initially, in- pressors for electric vehicles, Appl. Energy (2016) S0306261915016050.
creasing injection pressure leads to a larger heating capacity; when the [9] X. Xu, Investigation of vapor injection heat pump system with a flash tank utilizing
r410a and low-gwp refrigerant r32, Dissertations Theses - Gradworks (2012).
injection pressure is beyond the optimum one, the heating capacity
[10] X. Wang, Y. Hwang, R. Radermacher, Two-stage heat pump system with vapor-
starts to decrease. Also, there is not much difference between the op- injected scroll compressor using r410a as a refrigerant, Int. J. Refrig. 32 (6) (2009)
timum injection pressure simulated by the model and found in the ex- 1442–1451.
periment. Hence the optimum injection pressure can be roughly esti- [11] F. Qin, Q. Xue, G.M. Albarracin Velez, G. Zhang, H. Zou, C. Tian, Experimental
investigation on heating performance of heat pump for electric vehicles at -20°c
mated using the machine learning model. ambient temperature, Energy Convers. Manage. 102 (2015) 39–49.
[12] F. Qin, Q. Xue, G. Zhang, H. Zou, C. Tian, Experimental investigation on heat pump
6. Conclusion for electric vehicles with different refrigerant injection compressors, Energy
Procedia 75 (6) (2015) 1490–1495.
[13] J. Jongho, J. Yongseok, L. Hoseong, K. Yongchan, Numerical study of the effects of
A method to predict the performance of an EVI heat pump is pro- injection-port design on the heating performance of an R134a heat pump with
posed in the present study. The procedure of this method is as the vapor injection used in electric vehicles, Appl. Therm. Eng. 127 (2017) 800–811.
[14] K. Chunkyu, K. Mo Se, C. Younguk, K. Min Soo, Performance evaluation of a vapor
following: data standardization, train/test split, feature selection, injection heat pump system for electric vehicles, Int. J. Refrig. 74 (2017) 138–150.
model optimization, and performance prediction. SVR and Adaboost.R2 [15] L.B. Rasmussen, P. Bacher, H. Madsen, H.A. Nielsen, C. Heerup, T. Green, Load
are applied in the modeling process. The conclusions are summarized forecasting of supermarket refrigeration, Appl. Energy 163 (2016) 32–40.
[16] S. Shi, G. Li, H. Chen, J. Liu, Y. Hu, L. Xing, et al., Refrigerant charge fault diagnosis
below: in the vrf system using bayesian artificial neural network combined with relieff
filter, Appl. Therm. Eng. 112 (2017) 698–706.
(1) FS2 (including N, Tout, Tin, MFRair,in, Pdis, Pinj, πinj) is the best feature [17] Y. Guo, J. Wang, H. Chen, G. Li, J. Liu, C. Xu, et al., Machine learning-based thermal
response time ahead energy demand prediction for building heating systems, Appl.
set for model input.
Energy 221 (2018) 16–27.
(2) The models established can predict the heating capacity and COP [18] M.M. Rashidi, A. Aghagoli, R. Raoofi, Thermodynamic analysis of the ejector re-
within an error margin of 8.25% and 8.33% respectively. frigeration cycle using the artificial neural network, Energy 129 (2017) 201–215.
(3) The effects of injection pressure on the heat pump system can be [19] J.M. Belman-Flores, J.M. Barroso-Maldonado, S. Ledesma, V. Pérez-García,
A. Gallegos-Muñoz, J.A. Alfaro-Ayala, Exergy assessment of a refrigeration plant
simulated and optimum injection pressure for the system can be using computational intelligence based on hybrid learning methods, Int. J. Refrig.
estimated using the model established. 88 (2018).
[20] M. Laidi, S. Hanini, Optimal solar cop prediction of a solar-assisted adsorption re-
frigeration system working with activated carbon/methanol as working pairs using
In conclusion, the machine learning model is a powerful tool to direct and inverse artificial neural network, Int. J. Refrig. 36 (1) (2013) 247–257.
predict the performance of EVI heat pump system. Meanwhile, it poses [21] M. Liangyu, G. Yinping, Superheated steam temperature predictive optimal control
some potential in finding the optimum operating parameters of the based on external time-delay BP neural network and a simpler PSO algorithm, in:
Control Conference, IEEE, 2012.
system. [22] R.J. Moffat, Describing the uncertainties in experimental results, Exp. Therm Fluid
Sci. 1 (1) (1988) 3–17.
References [23] J.L. Rodgers, W. Alannicewander, Thirteen ways to look at the correlation coeffi-
cient, Am. Statist. 42 (1) (1988) 8.
[24] A.J. Smola, B. Schölkopf, A tutorial on support vector regression, Statist. Comput.
[1] H. Khayyam, A.Z. Kouzani, E.J. Hu, S. Nahavandi, Coordinated energy management 14 (2004) 199–222.
of vehicle air conditioning system, Appl. Therm. Eng. 31 (5) (2011) 750–764. [25] V. Vapnik, The Nature of Statistical Learning Theory, Springer, New York, 1995.
[2] Z. Zhang, W. Li, C. Zhang, J. Chen, Climate control loads prediction of electric [26] H.D. Monmouth, H. Drucker, Improving regressors using boosting techniques,
vehicles, Appl. Therm. Eng. 110 (2016) 1183–1188. Fourteenth International Conference on Machine Learning, Morgan Kaufmann
[3] L. Cichong, Z. Yun, G. Tianyuan, S. Junye, C. Jiangping, W. Tianying, et al., Publishers Inc, 1997.
Performance evaluation of propane heat pump system for electric vehicle in cold [27] X. Xu, Y. Hwang, R. Radermacher, Refrigerant injection for heat pumping/air
climate, Int. J. Refrig. (2018) S0140700718303165-. conditioning systems: literature review and challenges discussions, Int. J. Refrig. 34
[4] W. Dandong, Y. Binbin, L. Wanyong, S. Junye, C. Jiangping, Heating performance (2) (2011) 402–415.
evaluation of a co 2, heat pump system for an electrical vehicle at cold ambient