Research on adaptive control strategy optimization of hybrid electric vehicle

Abstract

This paper is concerned with the self-adaptive control problems for the parallel hybrid electric power systems based on the fuzzy relative membership classification theory and an adaptive control strategy optimization method for HEV dynamic systems is proposed. This optimized control strategy can adaptively adjust its control parameters based on the real-time driving cycle, effectively improving the fuel economy of the HEV. Firstly, four types of representative driving cycles are constructed based on actual vehicle operating data, using principal component analysis and cluster analysis to reflect the actual vehicle running conditions. Additionally, the optimal control parameters for each type of representative driving cycle are determined. Then, a fuzzy driving cycle recognition algorithm is proposed for online recognition of the actual driving cycle type. The optimal control parameters for the identified driving cycle type are then updated in the vehicle controller, to automatically realize control strategy optimization for different driving cycles. Finally, simulation experiments are conducted to verify the accuracy of the proposed fuzzy recognition algorithm and the validity of the designed control strategy optimization method.

Keywords

Fuzzy recognition intelligent learning control strategy hybrid electric vehicle

1 Introduction

Energy conservation and environmental protection are two main themes of the current auto industry development. Hybrid electric vehicles (HEV) have been introduced as an interim solution prior to the full implementation of EVs which depend on a breakthrough in EV energy sources. With an internal combustion engine and an electric motor as the power source, HEV’s power split control strategy has become the focus of recent research [1, 2]. This control strategy is the key to determining the fuel economy and emissions of a HEV and is highly significant for improving vehicle drivability, fuel economy and reducing emissions.

At present, there are two main control strategy types in the energy management system: one uses rule-based strategies including fuzzy rule-based methods and deterministic rule-based methods, the other uses optimization-based strategies which consist of global optimization methods and real-time optimization. In [3], the authors designed an Electric Double Layer Capacitor battery power source and a corresponding power split subsystem for a parallel-series type HEV energy management system, it could compensate the drawbacks of low power density and limited charge/discharge cycle life-times, getting a considerable performance enhancement at HEV operation and also helps to suppress the short circuit fault effects. However, such type of energy management system designs only aim at the urban road conditions. Different driving cycle types are generally not considered, where a driving cycle is a series of data points representing the speed of the vehicle over time, such as the New European Driving Cycle (NEDC) and the city driving cycle FTP. Control strategies designed for urban driving cycles will not achieve the expected vehicle performance under rural driving cycles. Fuel consumption and emissions are sensitive to these driving cycle variations [4].

Therefore, it makes sense to optimize the current control strategy for different types of driving cycle, which means that the driving cycle needs to be recognized first. In [5], the authors proposed an intelligent energy management agent using a learning vector quantization (LVQ) network, which can assess the driving environment using long and short term statistical features of the driving cycle, which are then used to determine the torque distribution and charge sustenance in the power split strategy to improve the fuel economy. A multilayer perceptron neural network with a back-propagation learning algorithm is also used to classify the driving patterns in [6]. In [7], a pattern recognition method is utilized for driving cycle identification to minimize fuel consumption and engine-out NOx and PM emissions on a set of diversified driving schedules, and the Dynamic Programming (DP) technique is used to find the corresponding global optimal control actions. The Support Vector Machine (SVM) [8] is another type of method used for driving cycle identification. The authors divide driving cycles into multiple layers and use a binary SVM classifier to identify the cycle types. A genetic agency is then used to optimize the control parameters in order to reduce the fuel consumption and emissions. Other model-based approaches [9, 10] achieve good identification results for some specific routes. However, current neural network driving cycle recognition methods require a large number of suitable training samples and SVM methods only perform effectively for two-class model identification, and do not easily achieve good multi-classification results. Recently, researchers have been paying remarkable attention to the fuzzy systems [11], the combination of a set of local linear models in the fuzzy-membership functions helps to solve the nonlinear system analysis problems, while the relative membership function of the fuzzy set theory could also be utilized for the pattern recognition in the engineering optimization with a relative fast computation speed. Thus, in this paper, we aim to propose a fuzzy driving cycle recognition method with a reduced level of calculation and a better result which can be utilized for online recognition, and also optimize a commercial control strategy to increase its fuel economy.

This paper is organized as follows. The next section presents the construction of four types of representative driving cycles using principal component analysis and cluster analysis. In Section 3, fuzzy driving cycle recognition is realized by calculating the membership grade between the sample to be recognized and each representative driving cycle. The fourth section presents the optimization of the commercialized electric assistive control strategy. Finally, the simulation results are presented in the last section to verify the accuracy of the proposed fuzzy recognition algorithm and the validity of the optimized control strategy.

2 Fuzzy recognition for driving cycles

To ensure that the control strategy can adjust online to different actual driving cycles, advance recognition of the driving cycles is required. The first step is to build four types of representative driving cycles. Additionally, primary and secondary factors that impact the characteristic parameters of the actual driving cycle are analyzed using orthogonal optimization theory. Finally, formula derivation with fuzzy theory is used to calculate the relative membership degrees in order to accomplish driving cycle recognition.

2.1 Construction of four types of driving cycles

The four types of driving cycles were constructed based on independent research and a development data acquisition system, which has been operating for nearly four years, as shown in Fig. 1. The system consists of a subordinate computer vehicle-mounted terminal and a supervisory computer remote monitoring terminal. The subordinate computer is responsible for data collection from the CAN (Controller Area Network) bus and sends data to the Internet over the GPRS network. The supervisory computer is responsible for receiving and storing data, including key parameters such as vehicle speed, engine speed, battery SOC (the state of charge) and vehicle load information.

This data acquisition system is the basis of the parameter acquisition and fuzzy recognition, providing hardware support for on-line recognition of the driving cycle. The data has been collected from hybrid electric vehicles within the Energy Efficient & New Energy Vehicle Demonstration Project in Dalian over a period of four years. A large volume of reliable and efficient real-time operational data was chosen, which can be divided into four representative types by cluster analysis: the Stopngo, the Urban, the Suburban and the Rural.

The steps to build each type of representative driving cycle are as follows. Firstly, the short trip of the original data is defined (a short trip is defined as the route between two idle cycles), and each short trip is divided continuously into micro-trips (a micro-trip is defined as the running process between an idle start-point and the next idle start point), and thus a database of short trips is formed. Secondly, the characteristic parameters of these separate segments of kinematic sequences are analyzed by principal component analysis using Matlab. Based on the score results obtained by the principal component analysis, the micro-trips can be classified to obtain excellent examples of micro-trips using the cluster analysis technique. Finally, these micro-trips examples are combined linearly to obtain this type of driving cycle.

2.2 Selection of recognition parameters for driving cycle

The recognition of a driving cycle is based on the characteristic parameters of the actual driving cycle [12]. Common conditional characteristic parameters currently used are as follows: average speed of the cycle, mean driving speed, maximum speed, maximum acceleration, maximum deceleration, mean acceleration, mean deceleration, idle time, number of stops, percentage of idle time and so on. However, it is not necessary to analyze all parameters in the process. If the characteristic parameters that can best characterize the driving cycles are selected from the parameters described above, the difficulty of online fuzzy recognition will be greatly reduced. According to mathematical statistical theory, we can assume that the emergence of a maximum speed, maximum acceleration and maximum deceleration is random in one cycle. Based on actual engineering experience and related literature [13 –15], these three quantities are not deemed to be useful as a measure of the eigenvalues of the driving cycle. The corresponding maximum and minimum value of the average acceleration and deceleration can also be determined within the process of computation. Many other characteristic parameters, including the average traveling speed, idle time, number of stops, and the percentage of idle time, are closely related to the idle time. Therefore, the average speed of the cycle, the average acceleration, the average driving speed, and the percentage idle time were selected as the eigenvalues of the driving cycle.

The five characteristic parameters were analyzed for each cycle using the orthogonal optimization design method. Using vehicle fuel economy as the evaluation index, the primary and secondary impacting factors were analyzed with orthogonal optimization theory, providing a theoretical basis for selection and extraction of the characteristic parameters.

The orthogonal experiment is based on an experimental design method, and selects representative and typical sites from a large number of test points using mathematical statistics theory methods and the orthogonal principle in multi-factor optimization experiments. Scientific tests are organized by applying the “orthogonal table” to achieve the optimal results using as few tests as possible [16]. This experimental design is a basic discrete optimization method which uses Latin square, uniform orthogonal tables as a tool to design the experimental scheme and seek the optimal points directly. The scheme for the orthogonal experiment is as shown in Table 1, where A is the average speed of the cycle, B is the average driving speed, C is the percentage idle time, D is the average acceleration, E is the average deceleration and y_i is the fuel consumption.

The results of the orthogonal design were analyzed to determine the primary and secondary factors of the experiment, the optimal level (value) of each experimental factor and the optimal factor combination for the test, in preparation for fuzzy recognition of the driving cycle. Based on a comprehensive comparison of the orthogonal table, a range analysis method (R method or visual analysis method) is applied to analyze the results of the driving cycle characteristic parameters in the orthogonal tests. The results can be analyzed intuitively and simply to determine the primary and secondary factors and the optimal combination of the experimental factors.

The range analysis method analyzes the problem using the average range for each factor. The range here means the difference between the maximum and minimum values for average effects. Within this range, the main factors influencing the indicator y (fuel consumption) can be discovered to determine the best combination of factors.

Table 2 shows the results of the orthogonal experiment. y_jk is the total test indicator corresponding to the jth factor and kth level and ${\bar{y}}_{jk}$ is the average value judged by the value of ${\bar{y}}_{jk}$ ; the value of k corresponding to a minimum value of ${\bar{y}}_{jk}$ is the optimal level for the jth factor in this experiment. The optimal design is a combination of the optimal level of various factors. In the table, R_j is the range of the jth factor, which is calculated as follows: $R_{j} - max [{\bar{y}}_{j 1}, {\bar{y}}_{j 2}, {\bar{y}}_{j 3} \dots] - min [{\bar{y}}_{j 1}, {\bar{y}}_{j 2}, {\bar{y}}_{j 3} \dots]$ (1)

It can be seen from Table 2 that the primary and secondary relationship between the five factors can be determined intuitively by comparing the values of the range R, as R_A > R_C > R_B > R_E > R_D. Therefore, the most important factor affecting the fuel economy is the average speed V_m, followed by the percentage idle time η and the average driving speed V_mr. The average deceleration A_d and average acceleration A_a are less important factors. This conclusion provides a theoretical basis for later characteristic parameters selection for fuzzy recognition of the driving cycle.

2.3 Fuzzy recognition of driving cycles

2.3.1 The theory of fuzzy recognition

The theory of fuzzy recognition is the mathematical basis for the optimum fuzzy set of system. The relative membership function of the fuzzy set theory for pattern recognition and adaptive fuzzy feedback control have been widely used in Agriculture, Chemicals and flood control fields [17 –20]. The objective function in Equation (2), which gives the minimum square sum of the weighted generalized Euclidean distances between all of the samples clustering sets to all categories, is used to establish a unified theory and cyclical iteration model of fuzzy recognition.

The expression for the constructed objective function is: $min {Y = \sum_{j = 1}^{n} \sum_{h = 1}^{c} {u_{h j}^{2} \sum_{i = 1}^{m} [ω_{i} (r_{ij} - s_{ih})]^{2}}}$ (2)

Given a set of samples consisting of n samples to be recognized: $X = [X_{1}, X_{2}, \dots \dots, X_{n}]$ (3)

Assuming that there are m indicators in each sample, then the feature vector for the jth sample is: $x_{j} = [x_{1 j}, x_{2 j}, \dots, x_{m j}]^{T}$ (4)

So the indicators for the n samples that need to be identified can be presented using the characteristic value matrix: $X = [\begin{matrix} x_{11} & x_{12} & \dots & x_{1 n} \\ x_{21} & x_{22} & \dots & x_{2 n} \\ \dots & \dots & \dots & \dots \\ x_{m 1} & x_{m 2} & \dots & x_{mn} \end{matrix}] = (x_{ij})_{m \times n}$ (5) where, x_ij is the characteristic value of indicator i for sample j (i = 1, 2, …… , m ; j = 1, 2, …… , n).

Since the characteristic values of the m indicators have different physical quantities and dimensions, normalization is required, which can be done using the following calculation: $r_{ij} = \frac{x_{ij} - x_{i min}}{x_{i max} - x_{i min}} or r_{ij} = \frac{x_{i max} - x_{ij}}{x_{i max} - x_{i min}}$ (6) where x_imax is the largest characteristic value of the ith indicator; x_imin is the smallest characteristic value of the ith indicator; and r_ij is the normalized value of x_ij which has a range 0 ≤ r_ij ≤ 1.

The relative membership degree matrix is calculated by substituting Equations (6) into (5): $R = [\begin{matrix} r_{11} & r_{12} & \dots & r_{1 n} \\ r_{21} & r_{22} & \dots & r_{2 n} \\ \dots & \dots & \dots & \dots \\ r_{m 1} & r_{m 2} & \dots & r_{mn} \end{matrix}] = (r_{ij})_{m \times n}$ (7)

Setting the n samples for fuzzy recognition in accordance with class c on the basis of the characteristic value of indicator m, then the fuzzy recognition matrix is $U = [\begin{matrix} u_{11} & u_{12} & \dots & u_{1 n} \\ u_{21} & u_{22} & \dots & u_{2 n} \\ \dots & \dots & \dots & \dots \\ u_{c 1} & u_{c 2} & \dots & u_{cn} \end{matrix}] = (u_{hj})_{c \times n}$ (8) where u_hj is the relative membership degree of sample j attributed to the hth class, which should meet: $0 \leq u_{hj} \leq 1, \sum_{h = 1}^{c} u_{hj} = 1, \sum_{j = 1}^{n} u_{hj} > 0$ (9)

Setting that indicator m’s characteristic value of category h as the clustering center of the class, the indicator’s characteristic values of the c categories can be expressed by the fuzzy clustering center m × c matrix as follows: $S = [\begin{matrix} s_{11} & s_{12} & \dots & s_{1 c} \\ s_{21} & s_{22} & \dots & s_{2 c} \\ \dots & \dots & \dots & \dots \\ s_{m 1} & s_{m 2} & \dots & s_{mc} \end{matrix}] = (s_{ih})_{m \times c}$ (10) where s_ih is the normalized number of characteristic values for category h and indicator i, with a range 0 ≤ s_ih ≤ 1.

The weight of the different indicators within fuzzy recognition needs to be considered, so W is used to express the indicator weight vector, which has the following form: $W = (w_{1}, w_{2}, \dots, w_{m}), \sum_{i = 1}^{m} w_{i} = 1$ (11)

To solve optimal fuzzy matrix, we need to determine the fuzzy clustering center matrix and the indicator weight vector.

Then the relative membership function of the fuzzy recognition theory model can be obtained by derivation: $u_{hj} = \frac{1}{\sum_{k = 1}^{c} \frac{\sum_{i = 1}^{m} {[w_{i} (r_{ij} - s_{ih})]}^{2}}{\sum_{i = 1}^{m} {[w_{i} (r_{ij} - s_{ik})]}^{2}}}$ (12)

2.3.2 Design of fuzzy driving cycle recognition algorithm

In this paper’s HEV driving cycle recognition, the sample matrix to be identified consists of five characteristic parameters which are achieved by receiving, unpacking and extracting the original data. So the sample contains five indicators and the indicator vector is x = { V_m, V_mr, η, A_a, A_d } ^T. Since the matrix requires normalization, more than one set of characteristic parameters in the matrix are required. Therefore, four groups of parameters relating to the typical driving cycle are added in order to identify the first group when the identification succeeds. The indicator characteristic matrix that needs to be identified can be expressed as: $X = (x_{ij})_{5 \times 5} = [\begin{matrix} V_{m} & 5.69 & 10.66 & 20.67 \\ V_{mr} & 19.59 & 16.55 & 23.22 \\ η & 0.7093 & 0.3562 & 0.1096 \\ \begin{matrix} A_{a} \\ A_{d} \end{matrix} & \begin{matrix} 1.34 \\ - 0.88 \end{matrix} & \begin{matrix} 0.70 \\ - 0.76 \end{matrix} & \begin{matrix} 0.71 \\ - 0.65 \end{matrix} \end{matrix} \begin{matrix} 24.55 \\ 27.27 \\ 0.0998 \\ \begin{matrix} 0.94 \\ - 1.13 \end{matrix} \end{matrix}]$ (13)

Since the characteristic values of the five indicators differ in their physical quantity dimension, this influence needs to be eliminated by performing normalization by using Equation (6) to calculate the optimal relative matrix R. The five characteristic values of the indicators in each of the four categories become the clustering center of the class, which can be expressed by the fuzzy clustering center matrix S. $S = [\begin{matrix} 1.0000 & 0.7365 & 0.2057 & 0.0000 \\ 0.7478 & 1.0000 & 0.3778 & 0.0000 \\ 0.0000 & 0.5793 & 0.9839 & 1.0000 \\ \begin{matrix} 0.0000 \\ 0.2500 \end{matrix} & \begin{matrix} 1.0000 \\ 0.7708 \end{matrix} & \begin{matrix} 0.9844 \\ 1.0000 \end{matrix} & \begin{matrix} 0.6250 \\ 0.0000 \end{matrix} \end{matrix}] = (s_{ih})_{5 \times 4}$ (14)

The different weights of the indicators in the fuzzy recognition then need to be determined. From the conclusions in Section 3, the degree of impact on the driving cycle is in the following order: the average speed, the percentage idle time, the percentage running time, the average deceleration, the average acceleration, and the given weight w_i. The indicator weight vector is W = (w₁, w₂, … , w₅), and the weight of each indicator should satisfy w₁ + w₂ + w₃ + w₄ + w₅ = 1 and w₁ > w₃ > w₂ > w₅ > w₄.

The weight of each indicator needs to be set, and the fourth cycle is selected as the standard in this paper. The similarity of each indicator in the other three cycles to the standard indicator is calculated using the equation:

$similarity = 1 - \frac{| similarity value - standard value |}{max (standard value, similarity value)}$ . The results are shown in Table 3.

The iteration procedures are written using the C# language and satisfy the above conditions. Three experiments were performed to calculate the similarity coefficient in each group, and the indicator weight vector result is W = (0.44, 0.34, 0.10, 0.09, 0.03).

The detection matrix can be calculated using Equation (12). For the required recognition matrix, the relative class characteristic value can be obtained using Equation (15) of the class characteristic value to determine which value the sample belongs to. So the type of driving cycle is recognized by H (x), $H (x) = \sum_{h = 1}^{c} h \cdot μ_{h} (x)$ (15) where μ_h (x) is the relative membership degree of sample x to category h (h = 1, 2, …, c).

3 Optimization design of control strategy

After identification of the driving cycle, control strategy optimization of each driving cycle should be performed. The optimization design for the control strategy is based on a commercialized electric assistive control strategy [21]. The core objective of this method is to reduce the power of the engine, and there is a threshold value to ensure that the engine is working in the high efficiency area. The motor drive system works using auxiliary power during low vehicle speed conditions or where there is a high power requirement, and regenerates power when braking to charge the battery, thereby improving the fuel economy and reducing the gas emissions of the HEV.

When studying the influence of the driving cycle on control strategy, the authors have noted that different engine off torque coefficients (denoted by ‘engine_off’) and different velocities of pure electric mode (denoted by ‘electric_launch_spd’) have the greatest influence on the vehicle fuel economy in different cycles, and the charging/discharging torque coefficient and the upper and lower limit value of SOC (state of charge) parameters have a very small impact on the vehicle fuel economy. In other words, there exists a specific parameter value of ‘engine_off’ that ensures that the vehicle will achieve the minimum fuel consumption per hundred kilometers within a single cycle; similarly, there exists a specific parameter value of ‘electric_launch_spd’ that also ensures that the vehicle achieves the minimum fuel consumption per hundred kilometers within a single cycle. Therefore, the parameters ‘engine_off’ and ‘electric_launch_spd’ are selected as the optimal control parameters for the control strategy in this paper. For the four types of driving cycles constructed in Section 2.1 above, there is a group of parameter values for ‘engine_off’ and ‘electric_launch_spd’ that can ensure that the HEV achieves the best fuel economy in each driving cycle. For each different type of driving cycles, a simulation has been performed by selecting a range of different parameter values for ‘engine_off’ and ‘electric_launch_spd’ using the advanced vehicle simulator (ADVISOR) software to determine the optimal parameters corresponding to the best fuel economy under each cycle’s conditions (an example for Stopngo is shown in Fig. 2). The selected optimal parameters ‘engine_off’ and ‘electric_launch_spd’ are chosen as the optimal parameters in this cycle, as shown inTable 4.

The control strategy is a general solution to the problem of coordinating switching and matching between multiple energy sources for HEVs. Effective and reasonable allocation of energy between the engine and the motor is realized by adjusting and controlling the power flow from the different components, to achieve the highest efficiency within the overall vehicle system in order to obtain the vehicle’s maximum fuel economy, lowest emissions and smoothest driving performance. An optimized control strategy is designed in this paper as shown in Fig. 3, to achieve the objective of greatly improving vehicle performance by designing control strategies that can be adaptively adjusted for different driving cycles, and confirming the actual driving cycle. Firstly, the vehicle’s typical driving cycle is identified in real-time by the online fuzzy recognition system; then using the speed, the torque requirements, and the battery’s SOC along with the identified driving cycle, the optimized control parameters are obtained using the method in the previous paragraph and stored in the controller; and finally these parameters are used by the vehicle controller to obtain the optimized engine torque and the motor torque, thus achieving optimal vehicle control.

4 Simulation and analysis

4.1 Construction of simulation model

A simulation model was built and a simulation experiment was conducted to verify the effectiveness of the driving cycle fuzzy recognition algorithm and the effectiveness of the control strategy proposed in this paper. The driving cycle recognition model for the control strategy designed in this paper is displayed in Fig. 4. The vehicle’s driving cycle (Stopngo, Urban, Suburban or Rural) was identified by the driving cycle recognition system which is realized by a MATLAB functional module; and then the optimal control parameters corresponding to the identified driving cycle were selected to update the control parameters to achieve the objective of automatically updating the control strategy with the driving cycle.

The model for the control strategy was embedded into the ADVISOR and the simulation was performed. The test cycle used in the process of simulation was the Dalian cycle (as is shown in Fig. 5), which was constructed based on data collected from hybrid electric vehicles on the Energy Efficient & New Energy Vehicle Demonstration Project in Dalian for four years by an independent research and development data acquisition system for HEV. The total time of the Dalian cycle is 1235 s and the cycle is similar to the actual traffic characteristics of the Dalian area. Therefore this cycle was selected for the experiment to ensure the accuracy of the experimental results.

4.2 Analysis of simulation experiment

The driving cycle recognition system operated over a period of 10 s, and the recognition results are shown in Fig. 5. The blue line is the Dalian city driving cycle used in the simulation and the red dashed line represents the identified type of driving cycle. Class 1 represents ‘Stopngo’ Class 2 represents ‘Urban’; Class 3 represents ‘Suburban’; and Class 4 represents ‘Rural’. As can be seen from the graph, in addition to individual time points (red overlapping region), the driving cycle recognition system accurately identified the cycle type. By lengthening the recognition time period, great improvements were seen in the red areas. Based on the proportion of different types of driving cycles shown in the figure below, it can easily be seen that the lowest proportion are of Urban type and the highest proportion are of Rural type, which is in accordance with the real driving conditions in Dalian. In order to verify the reliability and accuracy of the proposed algorithm, the typical driving cycles of New York Bus and UDDS are also used for testing, whose results are shown in Fig. 6. Compared with the Dalian Driving cycle, the New York Bus cycle’s average speed is relatively lower, and shows the frequent start-and-stop feature of the city bus. From the recognition result, we found that there was nearly no delay, even when only using the speed over the past 10 seconds as the input to our recognition algorithm, and the classification accurately reflected the average speed feature (which is given by the largest weight at the data processing stage). As shown in the pie chart, Stopngo is the dominant driving cycle, with a 59% proportion. UDDS, another American city standard driving cycle, is also tested and the results show 80% of the cycles are suburban or rural cycles. This is reasonable, because the average speed, acceleration and deceleration are much higher than the other two tested cycles. Therefore the fuzzy recognition algorithm can feasibly get an accurate classification result for the various cycles.

Figure 7 is the actual speed of the hybrid electric vehicle which accurately tracks the Dalian driving cycle. This illustrates that the control strategy that has been designed in this paper can meet dynamic performance requirements. Figure 8 is the history curve of the current for the power battery. It can be seen from Fig. 8 that the current curve of the power battery during the simulation process shows both positive and negative values. When the current is positive, this illustrates that the power battery is discharging to provide power to the motor by changing its energy from electrical energy to mechanical energy; when the current is negative, this illustrates that the power battery is charging, supplied from the motor by changing its energy from mechanical energy to electrical energy; when the current is zero, the power battery has stopped actively working, and the battery SOC fluctuates around 0.66, which helps to reduce the damage to the power battery, thus extending its life. When the power battery is discharging, the SOC value decreases. This mainly happens during periods of high speeds (i.e. at every peak interval, the battery is discharging), which means that the vehicle is working at a high load, and the motor can provide assistive torque to reduce the engine load. In contrast, during periods of low speeds, the motor tends to charge the battery, and the SOC value increases, so the additional charging torque can help the engine to increase its power load and improve its working efficiency. This chipping peak off and filling valley up effect can ultimately help the whole power system by improving its overall efficiency leading to a lower fuel consumption. Figure 9 is one history curve of the control parameter ‘engine_off’ during the simulation process. As can be seen from the graphs, the controller will choose the optimal control parameters of the identified driving cycle to optimize the control strategy, ensuring that the control parameters are updated in real-time during vehicle operation. The fuel consumption of the control strategy designed in this paper for the Dalian cycle is 27.975 L/100 km (the type of vehicle is CA6124SH8). Compared to the fuel consumption of the original control strategy used in the actual vehicle (31.112 L/100 km, which is the average value of 20 vehicles), the fuel economy is improved by 10.1%.

5 Conclusion

In this paper, we originally introduce the fuzzy relative membership theory to solve the HEV driving cycle pattern recognition problem and the control strategy is further researched using an independent research and development data acquisition system, and an adaptive control strategy optimization method for HEV is proposed. This optimized control strategy can adaptively adjust the control parameters based on the actual driving cycle, thus effectively improving the vehicle fuel economy of the hybrid electric vehicle. The main contributions of this paper are as follows. Firstly, using vehicle operating data obtained over a period of four years, four types of driving cycles have been built using principal component analysis and cluster analysis technology. Secondly, the optimal control parameters for each type of driving cycle have been determined, which laid a foundation for real-time optimization of the control strategy. Thirdly, the five characteristic parameters of real-time driving cycle data were inputted to calculate the relative fuzzy membership function for each representative cycle type, realizing the online driving cycle recognition based on robust fuzzy theory. Finally, a control strategy simulation model has been built and the validity and accuracy of this model has been verified by simulation experiments. From the results of the online driving cycle recognition, it has been shown that the vehicle controller adjusts to the corresponding optimal control parameters, thereby automatically updating the control strategy and thus optimizing each of the different driving cycles. The simulation experimental results show that the adaptive control strategy optimization method for HEV proposed in this paper improves vehicle fuel-consumption and emission performance effectively.

Footnotes

Acknowledgments

This project is supported by the National Natural Science Foundation of China (Grant No. 61203171, 61473057), China Postdoctoral Science Foundation (Grant No. 2012M511139, 2013T60278) and the China Fundamental Research Funds for the Central Universities (Grant No. DUT15LK13).

References

Ngamroo

, Specified structure mixed H2/H∞ control-based robust frequency stabilization in a smart grid by plug-in hybrid electric vehicles, Int J Innov Comput Inf Control9(1) (2013), 81–97.

Brown

, Pyke

and Steenhof

, Electric vehicles: The role and importance of standards in an emerging market, Energy Policy38 (2010), 3797–3806.

Sefik

and Asfani

D.A.

, Performance analysis of hybrid electric vehicle with electric double layer capacitor under short circuit fault, 11(1) (2015), 257–280.

Sharer

, Leydier

and Rousseau

, Impact of drive cycle aggressiveness and speed on HEVs fuel consumption sensitivity, 2006 (2007), 02–17.

Langari

and Won

J.S.

, Intelligent energy management agent for a parallel hybrid vehicle-part I: System architecture and design of the driving situation identification process, Vehicular Technology, IEEE Transactions on54 (2005), 925–934.

V.L.

and Calabrese

, NN-based measurements for driving pattern classification, Instrumentation and Measurement Technology Conference, 2009 I2MTC’09 IEEE, 2009, pp. 259–264.

Lin

C.C.

, Jeon

and Peng

, Driving pattern recognition for control of hybrid electric trucks, 42 (2010), 41–58.

Zhang

, Zhang

, Tian

and Zhang

, Intelligent Energy Management for Parallel HEV Based on Driving Cycle Identification Using SVM, Proceedings of the International Workshop on Information Securit, 2009, pp. 457–.

Larsson

, Johannesson

, Egardt

and Lassson

, Benefit of route recognition in energy management of plug-in hybrid electric vehicles, 2012, pp. 1314–1320.

10.

Ravey

, Blunier

and Lukic

, Control strategy of fuel cell hybrid electric vehicle based on driving cycle recognition, Transportation Electrification Conference and Expo (ITEC), 2012, pp. 1–5.

11.

, Shi

, Wu

and Zhang

, Fuzzy-model-basedD-stability and nonfragile control for discrete-time descriptor systems with multiple delays, IEEE Transactions on Fuzzy Systems22(4) (2014), 1019–1025.

12.

Savvidis

, Li

, Andrews

and Ioakimidis

, Analysis of various driving parameters and emissions for passenger cars driven with and without stops at intersections under different test cycles, Diesel Engine2011, 08–19.

13.

Montazeri

M.G.

, Poursamad

and Ghalichi

, Application of genetic algorithm for optimization of control strategy in parallel hybrid electric vehicles, Journal of the Franklin Institute343 (2006), 420–435.

14.

Huang

Y.J.

, Yin

C.L.

and Zhang

J.W.

, Design of an energy management strategy for parallel hybrid electric vehicles using a logic threshold and instantaneous optimization method, International Journal of Automotive Technology10 (2009), 513–521.

15.

Zhou

, Wang

Q.N.

and Wang

P.Y.

, The driving cycle characteristic parameters research for hybrid electric vehicle, Advanced Materials Research462 (2012), 271–276.

16.

Qazi

M.D.

and Linshu

, Nearly-orthogonal sampling and neural network metamodel driven conceptual design of multistage space launch vehicle, Computer-Aided Design38 (2006), 595–607.

17.

Tong

S.C.

, Li

Y.M.

and Liu

Y.J.

, Observer -based adaptive fuzzy backstepping control for a class of stochastic nonlinear strict-feedback systems, IEEE Transactions on Systems, Man and Cybernetics, Part B41(6) (2011), 1693–1704.

18.

Tong

S.C.

, Huo

B.Y.

and Li

Y.M.

, Observer-based adaptive decentralized fuzzy fault-tolerant control of nonlinear large-scale systems with actuator failures, IEEE Transactions on Fuzzy Systems22(1) (2014), 1–15.

19.

Tong

S.C.

and Li

Y.M.

, Adaptive fuzzy output feedback tracking backstepping control of strict-feedback nonlinear systems with unknown dead zones, IEEE Transactions on Fuzzy Systems20(1) (2012), 168–180.

20.

Tong

S.C.

, He

X.L.

and Zhang

H.G.

, A combined backstepping and small-gain approach to robust adaptive fuzzy output feedback control, IEEE Transactions on Fuzzy Systems17(5) (2009), 1059–1069.

21.

Johnson

V.H.

, Wipke

K.B.

and Rausen

D.J.

, HEV control strategy for real-time optimization of fuel economy and emissions, SAE Transactions109 (2000), 1677–1690.