A Neural Network Fuzzy Energy Management Strategy for Hybrid Electric Vehicles Based on Driving Cycle Recognition

: Aiming at the problems inherent in the traditional fuzzy energy management strategy (F-EMS), such as poor adaptive ability and lack of self-learning, a neural network fuzzy energy management strategy (NNF-EMS) for hybrid electric vehicles (HEVs) based on driving cycle recognition (DCR) is designed. The DCR was realized by the method of neural network sample learning and characteristic parameter analysis, and the recognition results were considered as the reference input of the fuzzy controller with further optimization of the membership function, resulting in improvement in the poor pertinence of F-EMS driving cycles. The research results show that the proposed NNF-EMS can realize the adaptive optimization of fuzzy membership function and fuzzy rules under di ﬀ erent driving cycles. Therefore, the proposed NNF-EMS has strong robustness and practicability under di ﬀ erent driving cycles.


Introduction
Hybrid electric vehicles (HEVs) combine the advantages of traditional fuel vehicles with battery electric vehicles (BEVs). However, unlike BEVs, HEVs have a more complex structure. According to the structure and energy flow of the powertrain, HEVs are mainly divided into series HEVs (SHEVs), parallel HEVs (PHEVs), and series-parallel HEVs (SPHEVs). Meanwhile, HEVs are also divided into plug-in HEVs and non-plug-in HEVs, depending on whether the battery can be charged externally. Apparently, higher requirements are imposed on the energy management strategy (EMS) of the vehicle [1,2]. The EMSs of HEVs are responsible for distributing the energy flow between the engine, the motor, and the battery under the premise of satisfying the vehicle's power demand, so as to achieve the optimal overall fuel consumption and emission performance [3,4]. At present, the EMSs applied to HEVs are mainly rule-based, including deterministic rule-based EMSs and fuzzy rule-based EMSs (F-EMSs) [5,6]. Among them, the deterministic rule-based EMSs are widely used in HEVs because of their simplicity, small computation, and strong practicability, but their disadvantages are also obvious. The logical threshold control strategy must be determined in advance, and the adaptability to different driving cycles is poor. Although F-EMSs cannot achieve global optimal energy management, fuzzy control itself has strong robustness and good real-time performance, and it can describe control rules that are difficult to quantify. Therefore, the application prospect is very good [6,7].
Up to now, many research results were achieved through the use of fuzzy control in the energy management of electric vehicles. The F-EMSs of parallel hybrid electric vehicles (PHEVs) were studied system and the reasoning of the fuzzy rules, thereby not depending on the mathematical model of the controlled object, but realizing the complex control object with uncertainty. Therefore, for high-order complex systems that are difficult to implement with traditional control methods, the advantages of fuzzy control are very evident. Fuzzy control has strong robustness and adaptability to uncertain factors, such as electric vehicle parameters and driving conditions. The control rules that are difficult to accurately quantify in the EMS of HEV can be realized by fuzzy logic [12,23,24].
The core of fuzzy control is the fuzzy control rules, which are expressed in the form of human language. For example, if an electric vehicle requires more power, the output power of the battery and engine is greater. Rules are easily accepted and understood; in addition, fuzzy control algorithms are easy to implement digitally. These distinctive features allowed fuzzy control theory to develop rapidly in recent decades and become an active field in intelligent control [8,9,25,26].
In the fuzzy system, there are two main representations of the fuzzy model: one is that the latter part of the fuzzy rule is a fuzzy set of outputs, called the standard model of the fuzzy system or the Mamdani model; the other is that the latter part of the fuzzy rule is a function of the input linguistic variable, called the Takagi-Sugeno model of the fuzzy system. Obviously, the Mamdani model should be used for the energy management or torque distribution control strategy of HEVs. The principle of the Mamdani fuzzy logic system, which is based on the standard model, is shown in Figure 1. In the process of realizing fuzzy control, a fuzzy controller is the carrier of a fuzzy control algorithm, which is mainly composed of the rule base, input fuzzification, fuzzy reasoning, and clear output [10,11].
Appl. Sci. 2020, 10, x FOR PEER REVIEW 3 of 18 order complex systems that are difficult to implement with traditional control methods, the advantages of fuzzy control are very evident. Fuzzy control has strong robustness and adaptability to uncertain factors, such as electric vehicle parameters and driving conditions. The control rules that are difficult to accurately quantify in the EMS of HEV can be realized by fuzzy logic [12,23,24]. The core of fuzzy control is the fuzzy control rules, which are expressed in the form of human language. For example, if an electric vehicle requires more power, the output power of the battery and engine is greater. Rules are easily accepted and understood; in addition, fuzzy control algorithms are easy to implement digitally. These distinctive features allowed fuzzy control theory to develop rapidly in recent decades and become an active field in intelligent control [8,9,25,26].
In the fuzzy system, there are two main representations of the fuzzy model: one is that the latter part of the fuzzy rule is a fuzzy set of outputs, called the standard model of the fuzzy system or the Mamdani model; the other is that the latter part of the fuzzy rule is a function of the input linguistic variable, called the Takagi-Sugeno model of the fuzzy system. Obviously, the Mamdani model should be used for the energy management or torque distribution control strategy of HEVs. The principle of the Mamdani fuzzy logic system, which is based on the standard model, is shown in Figure 1. In the process of realizing fuzzy control, a fuzzy controller is the carrier of a fuzzy control algorithm, which is mainly composed of the rule base, input fuzzification, fuzzy reasoning, and clear output [10,11].  The fuzzy control rules of the Mamdani fuzzy controller are represented by a series of fuzzy conditional statement with "if…, then…". The basic form of the fuzzy rule language of the fuzzy controller is as follows: If is and is and and is , then is where i x and i A (i = 1, 2, ... n) are the input fuzzy linguistic variables and their values, respectively, whereas y and B are the output linguistic variables and their values. In this paper, the fuzzy control strategy selects the SOC of the power battery and the total demand torque as the reference input of the fuzzy controller; that is, the state signals (such as SOC and demand torque) are converted into a set of fuzzy quantities expressed by membership degree, and the switching state of the engine and the magnitude of the output torque are determined by fuzzy rules according to the operating efficiency curve of the engine. The design ideas of the fuzzy EMS are described below.
(1) When the battery is fully charged and the demand torque is large, the HEV operates in the hybrid drive mode. At this point, the engine works near the optimal efficiency curve, and the residual torque is provided by the motor; when the demand torque is relatively large and less than the optimal efficiency torque of the engine, the HEV is separately driven by the engine, and the motor does not work; (2) When the battery is relatively full and the demand torque is large, the HEV operates in engine drive mode or hybrid drive mode. At this point, the engine works at optimum efficiency, and the residual torque is supplemented by the motor or used to charge the battery; The fuzzy control rules of the Mamdani fuzzy controller are represented by a series of fuzzy conditional statement with "if . . . , then . . . ". The basic form of the fuzzy rule language of the fuzzy controller is as follows: If x 1 is A 1 and x 2 is A 2 and . . . and x n is A n , then y is B, where x i and A i (i = 1, 2, . . . n) are the input fuzzy linguistic variables and their values, respectively, whereas y and B are the output linguistic variables and their values. In this paper, the fuzzy control strategy selects the SOC of the power battery and the total demand torque as the reference input of the fuzzy controller; that is, the state signals (such as SOC and demand torque) are converted into a set of fuzzy quantities expressed by membership degree, and the switching state of the engine and the magnitude of the output torque are determined by fuzzy rules according to the operating efficiency curve of the engine. The design ideas of the fuzzy EMS are described below.
(1) When the battery is fully charged and the demand torque is large, the HEV operates in the hybrid drive mode. At this point, the engine works near the optimal efficiency curve, and the residual torque is provided by the motor; when the demand torque is relatively large and less than the optimal efficiency torque of the engine, the HEV is separately driven by the engine, and the motor does not work; (2) When the battery is relatively full and the demand torque is large, the HEV operates in engine drive mode or hybrid drive mode. At this point, the engine works at optimum efficiency, and the residual torque is supplemented by the motor or used to charge the battery; (3) When the battery is low, the motor works as much as possible in the generator state. At this point, the vehicle is driven primarily only by the engine, the engine works near the optimal efficiency curve, and the remaining battery is charged.
In summary, the fuzzy rules of the F-EMS of HEVs consist of a series of "if . . . , then . . . " fuzzy conditional statements in the following form: where Rule i represents the i-th rule, m is the total number of fuzzy rule, Z is the rule output linguistic variable, here referring to the actual given engine output torque T e , X in and Y in represent the regular input linguistic variables, here referring to the SOC of power battery and the total demand torque T r , and In the F-EMS, the membership function settings of the demand torque, the battery SOC, and the given torque signal of the engine are respectively shown in Figures 2-4. According to the basic principle of fuzzy control, as more experience is learned, the division of the control rules becomes finer, and the membership function become narrower and thinner, such that the scope of control can be divided more finely and the control be more sensitive; otherwise, the control will be rougher and more stable. In the F-EMS of HEVs, if different driving cycles can be identified online and the fuzzy control rules or membership functions can be adjusted and optimized accordingly, the control effect of the F-EMS can be improved. According to the fuzzy rules, the fuzzy output rules of the HEV fuzzy controller are shown in Figure 5 through fuzzy reasoning. Appl. Sci. 2020, 10, x FOR PEER REVIEW 5 of 18

Neural Network Fuzzy EMS Based on DCR
The fuzzy control system relies on experience to reason, but its speed is slow, precision is low, self-learning ability is poor, and membership function and fuzzy rules are difficult to determine. In addition, the neural network uses expert knowledge to simulate the thinking function of the human brain structure, which has strong self-learning ability and less manual intervention, but cannot make full use of empirical knowledge to deal with fuzzy information. It can be seen that it is necessary to use a combination of fuzzy logic and neural networks to learn from others in order to achieve complementarity [27,28].
The fuzzy control can analyze the driving intention based on experience, and then distribute the output torque of the motor and the engine. However, in reality, the driver's driving mode is determined based on the actual driving conditions, which vary greatly and significantly affect the fuel economy and emissions of HEV. The traditional HEV fuzzy control strategy apparently does not adjust the control strategy according to the change in driving conditions, which seriously affects the control effect and the popularization application of F-EMS. If the sample information of the neural network training condition can be used, then the neural network can be used to learn the online recognition of the driving cycle category, and the recognition result and the SOC, demand torque, and other signals become blurred, which are then used together as a reference input to the fuzzy controller, whereby the membership function and fuzzy rules of the fuzzy controller can be adjusted online based on the driving cycle information to improve the fuel economy and emissions of the HEV vehicle.

Basic Test Conditions and Characteristics
Different countries and regions have different standard driving cycles for vehicle emission tests. The representative driving cycles are the Federal Test Procedure, commonly known as FTP-75 for the city driving cycle, defined by the United States (US) Environmental Protection Agency (EPA), and the New European Driving Cycle (NEDC) of the Economic Commission of Europe (ECE). They are both a series of tests for measuring tailpipe emissions and fuel economy of passenger cars (excluding light trucks and heavy-duty vehicles); however, FTP75 is mainly used in the United States, Canada, South America, etc., while NEDC is mainly used in Europe, China, Australia, etc. [29,30]. The local driving velocity in NEDC is set to a constant value, as shown in Figure 6. The NEDC consists of two parts: Urban Driving Cycle ECE-15 (or simply UDC), repeated four times, is plotted from 0 s to 780 s; Extra-Urban Driving Cycle ECE R101 (or simply EUDC) cycle is plotted from 780 s to 1180 s. FTP75 is a mandatory dynamometer test on tailpipe emissions of a car that represents the driving conditions of the city, which was developed from the Urban Dynamometer Driving Schedule (UDDS, also known as FTP-72). As shown in Figure 7, FTP75 is identical to UDDS plus the first 505 s of an additional UDDS cycle.

Neural Network Fuzzy EMS Based on DCR
The fuzzy control system relies on experience to reason, but its speed is slow, precision is low, self-learning ability is poor, and membership function and fuzzy rules are difficult to determine. In addition, the neural network uses expert knowledge to simulate the thinking function of the human brain structure, which has strong self-learning ability and less manual intervention, but cannot make full use of empirical knowledge to deal with fuzzy information. It can be seen that it is necessary to use a combination of fuzzy logic and neural networks to learn from others in order to achieve complementarity [27,28].
The fuzzy control can analyze the driving intention based on experience, and then distribute the output torque of the motor and the engine. However, in reality, the driver's driving mode is determined based on the actual driving conditions, which vary greatly and significantly affect the fuel economy and emissions of HEV. The traditional HEV fuzzy control strategy apparently does not adjust the control strategy according to the change in driving conditions, which seriously affects the control effect and the popularization application of F-EMS. If the sample information of the neural network training condition can be used, then the neural network can be used to learn the online recognition of the driving cycle category, and the recognition result and the SOC, demand torque, and other signals become blurred, which are then used together as a reference input to the fuzzy controller, whereby the membership function and fuzzy rules of the fuzzy controller can be adjusted online based on the driving cycle information to improve the fuel economy and emissions of the HEV vehicle.

Basic Test Conditions and Characteristics
Different countries and regions have different standard driving cycles for vehicle emission tests. The representative driving cycles are the Federal Test Procedure, commonly known as FTP-75 for the city driving cycle, defined by the United States (US) Environmental Protection Agency (EPA), and the New European Driving Cycle (NEDC) of the Economic Commission of Europe (ECE). They are both a series of tests for measuring tailpipe emissions and fuel economy of passenger cars (excluding light trucks and heavy-duty vehicles); however, FTP75 is mainly used in the United States, Canada, South America, etc., while NEDC is mainly used in Europe, China, Australia, etc. [29,30]. The local driving velocity in NEDC is set to a constant value, as shown in Figure 6. The NEDC consists of two parts: Urban Driving Cycle ECE-15 (or simply UDC), repeated four times, is plotted from 0 s to 780 s; Extra-Urban Driving Cycle ECE R101 (or simply EUDC) cycle is plotted from 780 s to 1180 s. FTP75 is a mandatory dynamometer test on tailpipe emissions of a car that represents the driving conditions of the city, which was developed from the Urban Dynamometer Driving Schedule (UDDS, also known as FTP-72). As shown in Figure 7, FTP75 is identical to UDDS plus the first 505 s of an additional UDDS cycle. The comparison and driving cycles of NEDC and FTP75 are shown in Table 2. Obviously, their basic characteristic parameters are different; in particular, the NEDC and the FTP75 that are steadystate and transient driving cycles, respectively, with an obvious characteristic difference. Under different driving cycles, it is apparently not appropriate to adopt a constant EMS. According to the characteristic parameters of different driving cycles, a small section of the driving cycle that runs continuously between the start and stop of the vehicle is defined as a "driving cycle block". In this way, no matter how complicated the driving cycles are, they can be divided into specific small driving cycle blocks.

Driving Cycle Block and Its Definition
Driving cycle blocks can be distinguished according to different operating parameters, such as maximum velocity, maximum acceleration, maximum deceleration, average driving velocity, and driving distance, thereby effectively reducing the complexity of driving cycle analysis. In addition, the driving cycle blocks are independent of each other, and any complicated actual driving conditions can be divided into a plurality of different driving cycle blocks, which improves the reuse rate of the blocks [15,17]. The driving cycle blocks are used for the recognition of driving conditions, and the block recognition library of different scales and different levels of complexity can be established according to the control requirements, thereby designing an EMS that is highly targeted [31,32]. For example, a driving cycle library for different cities and different road sections can be established separately.
Taking NEDC as an example, driving cycle blocks can be divided into five basic categories [17].
(1) Block ① is defined for the driving condition in the city center, where the vehicle runs and stops frequently with heavy traffic; (2) Block ② is defined for driving conditions in urban areas, where the vehicle runs at low velocity; (3) Block ③ is defined for the driving condition in the suburbs, where the vehicle runs at medium and low velocity; The comparison and driving cycles of NEDC and FTP75 are shown in Table 2. Obviously, their basic characteristic parameters are different; in particular, the NEDC and the FTP75 that are steady-state and transient driving cycles, respectively, with an obvious characteristic difference. Under different driving cycles, it is apparently not appropriate to adopt a constant EMS. According to the characteristic parameters of different driving cycles, a small section of the driving cycle that runs continuously between the start and stop of the vehicle is defined as a "driving cycle block". In this way, no matter how complicated the driving cycles are, they can be divided into specific small driving cycle blocks.

Driving Cycle Block and Its Definition
Driving cycle blocks can be distinguished according to different operating parameters, such as maximum velocity, maximum acceleration, maximum deceleration, average driving velocity, and driving distance, thereby effectively reducing the complexity of driving cycle analysis. In addition, the driving cycle blocks are independent of each other, and any complicated actual driving conditions can be divided into a plurality of different driving cycle blocks, which improves the reuse rate of the blocks [15,17]. The driving cycle blocks are used for the recognition of driving conditions, and the block recognition library of different scales and different levels of complexity can be established according to the control requirements, thereby designing an EMS that is highly targeted [31,32]. For example, a driving cycle library for different cities and different road sections can be established separately.
Taking NEDC as an example, driving cycle blocks can be divided into five basic categories [17].
(1) Block 1 is defined for the driving condition in the city center, where the vehicle runs and stops frequently with heavy traffic; (2) Block 2 is defined for driving conditions in urban areas, where the vehicle runs at low velocity; (3) Block 3 is defined for the driving condition in the suburbs, where the vehicle runs at medium and low velocity; (4) Block 4 is defined for the driving condition in the suburbs, where the vehicle runs at medium and high velocity; (5) Block 5 is defined for the driving condition on the highway between cities and suburbs, where the vehicle runs at high velocity.
In fact, the NEDC basically consists of the above five basic driving conditions. The "driving cycle block" for NEDC defined in this paper is shown in Figure 8. Of course, these blocks are divided by different characteristic parameters of the driving conditions, which are closely related to complicated driving conditions. The block division of FTP75 is much more complicated than that of NEDC.
Appl. Sci. 2020, 10, x FOR PEER REVIEW 8 of 18 (4) Block ④ is defined for the driving condition in the suburbs, where the vehicle runs at medium and high velocity; (5) Block ⑤ is defined for the driving condition on the highway between cities and suburbs, where the vehicle runs at high velocity.
In fact, the NEDC basically consists of the above five basic driving conditions. The "driving cycle block" for NEDC defined in this paper is shown in Figure 8. Of course, these blocks are divided by different characteristic parameters of the driving conditions, which are closely related to complicated driving conditions. The block division of FTP75 is much more complicated than that of NEDC.

Neural Network Fuzzy EMS (NNF-EMS) Design Based on DCR
In this paper, the backpropagation (BP) neural network is used for DCR. The BP neural network is a multi-layer feed-forward network trained based on the error backpropagation algorithm. It can store and learn a large number of input-output mappings without having to mathematically describe these mappings. It embodies a nonlinear functional relationship between the input and output of the network. The BP neural network generally consists of an input layer, a hidden layer, and an output layer. The basic structure of the network is shown in Figure 9, which has M input nodes, L output nodes, and q neurons, wherein the actual input of the network is 1 2 , , , M x x x  , and the actual output is 1 2 , , , M y y y  . In the calculation process, learning and training are continuously performed according to the input sample network, and the response values of the neurons are transmitted from the input layer to the output layer through the hidden layer. According to the obtained network response, the output layer continuously adjusts the connection weights and thresholds of the network in the direction of the greatest reduction of error; the square of the error of the network is minimized by backpropagation, thereby improving the accuracy of the network output [27,28].

Neural Network Fuzzy EMS (NNF-EMS) Design Based on DCR
In this paper, the backpropagation (BP) neural network is used for DCR. The BP neural network is a multi-layer feed-forward network trained based on the error backpropagation algorithm. It can store and learn a large number of input-output mappings without having to mathematically describe these mappings. It embodies a nonlinear functional relationship between the input and output of the network. The BP neural network generally consists of an input layer, a hidden layer, and an output layer. The basic structure of the network is shown in Figure 9, which has M input nodes, L output nodes, and q neurons, wherein the actual input of the network is x 1 , x 2 , . . . , x M , and the actual output is y 1 , y 2 , . . . , y M . In the calculation process, learning and training are continuously performed according to the input sample network, and the response values of the neurons are transmitted from the input layer to the output layer through the hidden layer. According to the obtained network response, the output layer continuously adjusts the connection weights and thresholds of the network in the direction of the greatest reduction of error; the square of the error of the network is minimized by backpropagation, thereby improving the accuracy of the network output [27,28].
BP neural networks are used for pattern recognition and classification, but are not suitable for expressing rule-based knowledge. In the fuzzy system, knowledge extraction and rule expression are more convenient and suitable for fuzzy or qualitative knowledge, but lack the capacity for self-learning and self-adaptation. It can be seen that the neural network fuzzy EMS (NNF-EMS) of HEV combines the neural network-based cycle recognition and F-EMS, and their advantages complement each other to realize the DCR-based fuzzy control reasoning process. Although the actual driving conditions are random and undeterminable, if there are sufficient circulating conditions, the driving cycle blocks with similar characteristics can be found, and the control parameters can be adjusted or the corresponding control strategies can be found to improve control performance through real-time analysis. Appl. Sci. 2020, 10, x FOR PEER REVIEW  BP neural networks are used for pattern recognition and classification, but are not suitable for expressing rule-based knowledge. In the fuzzy system, knowledge extraction and rule expression are more convenient and suitable for fuzzy or qualitative knowledge, but lack the capacity for selflearning and self-adaptation. It can be seen that the neural network fuzzy EMS (NNF-EMS) of HEV combines the neural network-based cycle recognition and F-EMS, and their advantages complement each other to realize the DCR-based fuzzy control reasoning process. Although the actual driving conditions are random and undeterminable, if there are sufficient circulating conditions, the driving cycle blocks with similar characteristics can be found, and the control parameters can be adjusted or the corresponding control strategies can be found to improve control performance through real-time analysis.
The main advantage of a neural network for DCR is that the sample data can adapt to the driving conditions, and the neural network can also function normally even when the acquired driving condition data contain noise. In addition, the neural network provides simple tools for automatic feature selection, generates useful data representations, and can also serve as a front-end preprocessor for fuzzy controllers. Therefore, this paper uses a neural network to realize DCR, and the recognition result is combined with the input signals such as SOC and demand torque, and then used as input to the fuzzy controller to optimize the fuzzy membership function and fuzzy rules in overcoming the poorly targeted shortcomings of the F-EMS of HEVs.
The BP neural network algorithm is used as the training algorithm, and the number of training times is 50. The input node of the neural network selects the relevant characteristic parameters of the current driving velocity, average vehicle velocity, average acceleration, average deceleration, maximum vehicle velocity, maximum acceleration, maximum deceleration, traveled distance, and the travel time as the input nodes of the BP network. By adjusting the connection weighting coefficient, the role of different characteristic parameters in the identification of the working cycle can be adjusted, and the output value of the network node can be obtained by the weighted sum [16]. In this paper, the graphical interface editor (Anfis editor) of the adaptive neuro-fuzzy inference system was used provided by MATLAB Fuzzy Logic Toolbox to train and test the driving conditions of NEDC, and we selected parameters such as BP algorithm, error accuracy, and training times to train the neural network system. It can be seen that the error accuracy of the test data was reduced to a certain degree after a certain number of trainings. Figure 10 displays the training data and test data for training NEDC. After 50 training sessions, the error curve is shown in Figure 11. The main advantage of a neural network for DCR is that the sample data can adapt to the driving conditions, and the neural network can also function normally even when the acquired driving condition data contain noise. In addition, the neural network provides simple tools for automatic feature selection, generates useful data representations, and can also serve as a front-end preprocessor for fuzzy controllers. Therefore, this paper uses a neural network to realize DCR, and the recognition result is combined with the input signals such as SOC and demand torque, and then used as input to the fuzzy controller to optimize the fuzzy membership function and fuzzy rules in overcoming the poorly targeted shortcomings of the F-EMS of HEVs.
The BP neural network algorithm is used as the training algorithm, and the number of training times is 50. The input node of the neural network selects the relevant characteristic parameters of the current driving velocity, average vehicle velocity, average acceleration, average deceleration, maximum vehicle velocity, maximum acceleration, maximum deceleration, traveled distance, and the travel time as the input nodes of the BP network. By adjusting the connection weighting coefficient, the role of different characteristic parameters in the identification of the working cycle can be adjusted, and the output value of the network node can be obtained by the weighted sum [16]. In this paper, the graphical interface editor (Anfis editor) of the adaptive neuro-fuzzy inference system was used provided by MATLAB Fuzzy Logic Toolbox to train and test the driving conditions of NEDC, and we selected parameters such as BP algorithm, error accuracy, and training times to train the neural network system. It can be seen that the error accuracy of the test data was reduced to a certain degree after a certain number of trainings. Figure 10 displays the training data and test data for training NEDC. After 50 training sessions, the error curve is shown in Figure 11.  The neural network is used to train the driving cycle samples and solve the DCR problem by learning and extracting characteristic parameters of driving cycles. Then, the recognition results are used as control parameters to optimize the fuzzy membership function and fuzzy rules of the fuzzy controller. Taking the recognition process of the driving cycle blocks ② and ④ as an example, when the vehicle is recognized to be driving in the urban area (block ②), the traffic is relatively smooth and the vehicle runs at a low velocity. Therefore, as shown in Figures 12a and 13a, the membership function of the engine output torque signal is optimized to be relatively narrow and thin, the control scope is finely divided, and the control accuracy is high. When it is recognized that the vehicle is driving in the suburb (block ④), the traffic is smooth, and the vehicle drives at medium and high velocity. At this time, as shown in Figures 12b and 13b, the membership function of the engine output torque signal is optimized to be relatively wide and fat, and the control is relatively rough and stable.  The neural network is used to train the driving cycle samples and solve the DCR problem by learning and extracting characteristic parameters of driving cycles. Then, the recognition results are used as control parameters to optimize the fuzzy membership function and fuzzy rules of the fuzzy controller. Taking the recognition process of the driving cycle blocks ② and ④ as an example, when the vehicle is recognized to be driving in the urban area (block ②), the traffic is relatively smooth and the vehicle runs at a low velocity. Therefore, as shown in Figures 12a and 13a, the membership function of the engine output torque signal is optimized to be relatively narrow and thin, the control scope is finely divided, and the control accuracy is high. When it is recognized that the vehicle is driving in the suburb (block ④), the traffic is smooth, and the vehicle drives at medium and high velocity. At this time, as shown in Figures 12b and 13b, the membership function of the engine output torque signal is optimized to be relatively wide and fat, and the control is relatively rough and stable. The neural network is used to train the driving cycle samples and solve the DCR problem by learning and extracting characteristic parameters of driving cycles. Then, the recognition results are used as control parameters to optimize the fuzzy membership function and fuzzy rules of the fuzzy controller. Taking the recognition process of the driving cycle blocks 2 and 4 as an example, when the vehicle is recognized to be driving in the urban area (block 2 ), the traffic is relatively smooth and the vehicle runs at a low velocity. Therefore, as shown in Figures 12a and 13a, the membership function of the engine output torque signal is optimized to be relatively narrow and thin, the control scope is finely divided, and the control accuracy is high. When it is recognized that the vehicle is driving in the suburb (block 4 ), the traffic is smooth, and the vehicle drives at medium and high velocity. At this time, as shown in Figures 12b and 13b, the membership function of the engine output torque signal is optimized to be relatively wide and fat, and the control is relatively rough and stable.

Simulation Results and Analysis
The designed PHEV model in AVL Cruise is shown in Figure 14. Models of motors, engines, and power batteries are mainly built based on the look-up tables, which can greatly improve the calculation speed. In practice, the data in the look-up tables of these components can be obtained experimentally. According to the configuration, the drive and powertrain structure of the HEV is a parallel drive system, which is a torque-coupled, single-shaft hybrid drive system whose transmission is located behind the motor. The engine and the motor are required to have the same speed range, because the torque and speed transmitted by them to the drive axle are adjusted by the continuously variable transmission (CVT). The motor performs multiple functions, such as driving the vehicle, starting the engine, and generating regenerative braking.

Simulation Results and Analysis
The designed PHEV model in AVL Cruise is shown in Figure 14. Models of motors, engines, and power batteries are mainly built based on the look-up tables, which can greatly improve the calculation speed. In practice, the data in the look-up tables of these components can be obtained experimentally. According to the configuration, the drive and powertrain structure of the HEV is a parallel drive system, which is a torque-coupled, single-shaft hybrid drive system whose transmission is located behind the motor. The engine and the motor are required to have the same speed range, because the torque and speed transmitted by them to the drive axle are adjusted by the continuously variable transmission (CVT). The motor performs multiple functions, such as driving the vehicle, starting the engine, and generating regenerative braking.

Simulation Results and Analysis
The designed PHEV model in AVL Cruise is shown in Figure 14. Models of motors, engines, and power batteries are mainly built based on the look-up tables, which can greatly improve the calculation speed. In practice, the data in the look-up tables of these components can be obtained experimentally. According to the configuration, the drive and powertrain structure of the HEV is a parallel drive system, which is a torque-coupled, single-shaft hybrid drive system whose transmission is located behind the motor. The engine and the motor are required to have the same speed range, because the torque and speed transmitted by them to the drive axle are adjusted by the continuously variable transmission (CVT). The motor performs multiple functions, such as driving the vehicle, starting the engine, and generating regenerative braking. Appl. Sci. 2020, 10, x FOR PEER REVIEW 12 of 18 The driving cycles of NEDC and FTP75 were selected correspondingly to test the fuel consumption, CO, NOx, and hydrocarbon (HC) emission performance of the vehicle. The fuzzy rulebased F-EMS and the cycle recognition-based NNF-EMS were established through the co-simulation of AVL Cruise and Matlab/Simulink. Additionally, the co-simulation was implemented in the form of a dynamic link library called the Matlab DLL.

Results Analysis of Power Performance
The vehicle velocity performance shows the dynamic response of the actual vehicle velocity following the driver's given velocity, which is also the most basic condition for judging whether EMS is feasible. From the vehicle velocity conditions shown in Figures 15 and 16, it can be seen that, during the entire cycle of NEDC and FTP75, the vehicle had good dynamic performance and could respond quickly to a given vehicle velocity, indicating that the vehicle model is rationally modeled and the control strategy is feasible. It can be seen from the following situation after enlargement in Figures  17 and 18 that the NNF-EMS based on DCR has the best performance, whose dynamic response time is the fastest with no overshoot. In addition, it can be seen from the following situation that it is obviously better in the steady-state driving cycle NEDC than in the transient driving cycle FTP75, which is close to the actual operation. The driving cycles of NEDC and FTP75 were selected correspondingly to test the fuel consumption, CO, NOx, and hydrocarbon (HC) emission performance of the vehicle. The fuzzy rule-based F-EMS and the cycle recognition-based NNF-EMS were established through the co-simulation of AVL Cruise and Matlab/Simulink. Additionally, the co-simulation was implemented in the form of a dynamic link library called the Matlab DLL.

Results Analysis of Power Performance
The vehicle velocity performance shows the dynamic response of the actual vehicle velocity following the driver's given velocity, which is also the most basic condition for judging whether EMS is feasible. From the vehicle velocity conditions shown in Figures 15 and 16, it can be seen that, during the entire cycle of NEDC and FTP75, the vehicle had good dynamic performance and could respond quickly to a given vehicle velocity, indicating that the vehicle model is rationally modeled and the control strategy is feasible. It can be seen from the following situation after enlargement in Figures 17  and 18 that the NNF-EMS based on DCR has the best performance, whose dynamic response time is the fastest with no overshoot. In addition, it can be seen from the following situation that it is obviously better in the steady-state driving cycle NEDC than in the transient driving cycle FTP75, which is close to the actual operation. The driving cycles of NEDC and FTP75 were selected correspondingly to test the fuel consumption, CO, NOx, and hydrocarbon (HC) emission performance of the vehicle. The fuzzy rulebased F-EMS and the cycle recognition-based NNF-EMS were established through the co-simulation of AVL Cruise and Matlab/Simulink. Additionally, the co-simulation was implemented in the form of a dynamic link library called the Matlab DLL.

Results Analysis of Power Performance
The vehicle velocity performance shows the dynamic response of the actual vehicle velocity following the driver's given velocity, which is also the most basic condition for judging whether EMS is feasible. From the vehicle velocity conditions shown in Figures 15 and 16, it can be seen that, during the entire cycle of NEDC and FTP75, the vehicle had good dynamic performance and could respond quickly to a given vehicle velocity, indicating that the vehicle model is rationally modeled and the control strategy is feasible. It can be seen from the following situation after enlargement in Figures  17 and 18 that the NNF-EMS based on DCR has the best performance, whose dynamic response time is the fastest with no overshoot. In addition, it can be seen from the following situation that it is obviously better in the steady-state driving cycle NEDC than in the transient driving cycle FTP75, which is close to the actual operation.

Results Analysis of Fuel Economy and Emission Performance
Cycle economy tests of vehicle fuel economy and emission performance under NEDC and FTP75 were carried out through AVL Cruise. The experimental results are compared in Table 3. In comparison with the F-EMS, the NNF-EMS considers that the DCR is more sensitive to the driving conditions and the control fuzzy rules are more detailed; hence, the control is more sensitive and the vehicle performance of the HEV is further improved. Specifically, compared to the F-EMS, the fuel consumption of the NNF-EMS based on the cycle recognition under the NEDC cycle was reduced from 7.093 L/100 km to 5.576 L/100 km, and the fuel optimization rate reached 21.39%; CO emissions decreased from 5.362 g/km to 4.419 g/km, a 17.59% reduction; NOx emissions decreased from 1.910 g/km to 1.524 g/km, a 20.21% reduction; HC emissions decreased from 0.5702 g/km to 0.4054 g/km, a

Results Analysis of Fuel Economy and Emission Performance
Cycle economy tests of vehicle fuel economy and emission performance under NEDC and FTP75 were carried out through AVL Cruise. The experimental results are compared in Table 3. In comparison with the F-EMS, the NNF-EMS considers that the DCR is more sensitive to the driving conditions and the control fuzzy rules are more detailed; hence, the control is more sensitive and the vehicle performance of the HEV is further improved. Specifically, compared to the F-EMS, the fuel consumption of the NNF-EMS based on the cycle recognition under the NEDC cycle was reduced from 7.093 L/100 km to 5.576 L/100 km, and the fuel optimization rate reached 21.39%; CO emissions decreased from 5.362 g/km to 4.419 g/km, a 17.59% reduction; NOx emissions decreased from 1.910 g/km to 1.524 g/km, a 20.21% reduction; HC emissions decreased from 0.5702 g/km to 0.4054 g/km, a

Results Analysis of Fuel Economy and Emission Performance
Cycle economy tests of vehicle fuel economy and emission performance under NEDC and FTP75 were carried out through AVL Cruise. The experimental results are compared in Table 3. In comparison with the F-EMS, the NNF-EMS considers that the DCR is more sensitive to the driving conditions and the control fuzzy rules are more detailed; hence, the control is more sensitive and the vehicle performance of the HEV is further improved. Specifically, compared to the F-EMS, the fuel consumption of the NNF-EMS based on the cycle recognition under the NEDC cycle was reduced from 7.093 L/100 km to 5.576 L/100 km, and the fuel optimization rate reached 21.39%; CO emissions decreased from 5.362 g/km to 4.419 g/km, a 17.59% reduction; NOx emissions decreased from 1.910 g/km to 1.524 g/km, a 20.21% reduction; HC emissions decreased from 0.5702 g/km to 0.4054 g/km, a

Results Analysis of Fuel Economy and Emission Performance
Cycle economy tests of vehicle fuel economy and emission performance under NEDC and FTP75 were carried out through AVL Cruise. The experimental results are compared in Table 3. In comparison with the F-EMS, the NNF-EMS considers that the DCR is more sensitive to the driving conditions and the control fuzzy rules are more detailed; hence, the control is more sensitive and the vehicle performance of the HEV is further improved. Specifically, compared to the F-EMS, the fuel consumption of the NNF-EMS based on the cycle recognition under the NEDC cycle was reduced from 7.093 L/100 km to 5.576 L/100 km, and the fuel optimization rate reached 21.39%; CO emissions decreased from 5.362 g/km to 4.419 g/km, a 17.59% reduction; NOx emissions decreased from 1.910 g/km to 1.524 g/km, a 20.21% reduction; HC emissions decreased from 0.5702 g/km to 0.4054 g/km, a 28.90% reduction. In addition, the reduction in fuel consumption and the improvement in emission performance did not occur at the expense of consuming more battery power. The power consumption only increased from 6.0488 kWh to 6.0645 kWh. This shows that the NNF-EMS based on DCR better optimizes the working zone of the engine, improves fuel efficiency, and reduces emissions. The simulation output results of NNF-EMS in AVL Cruise under NEDC cycles are shown in Figure 19. 28.90% reduction. In addition, the reduction in fuel consumption and the improvement in emission performance did not occur at the expense of consuming more battery power. The power consumption only increased from 6.0488 kWh to 6.0645 kWh. This shows that the NNF-EMS based on DCR better optimizes the working zone of the engine, improves fuel efficiency, and reduces emissions. The simulation output results of NNF-EMS in AVL Cruise under NEDC cycles are shown in Figure 19.  Fuel consumption is ultimately caused by differences in engine operating points. Figure 20 shows the distribution of key operating points of the engine under the NEDC cycle. According to the comparison, in Figure 20a,b, under the F-EMS, the engine can work closer to the rated torque, with higher fuel efficiency, thereby significantly improving fuel economy and emissions. NNF-EMS further optimizes the engine operating points, thereby further reducing fuel consumption and improving emission performance. Fuel consumption is ultimately caused by differences in engine operating points. Figure 20 shows the distribution of key operating points of the engine under the NEDC cycle. According to the comparison, in Figure 20a,b, under the F-EMS, the engine can work closer to the rated torque, with higher fuel efficiency, thereby significantly improving fuel economy and emissions. NNF-EMS further optimizes the engine operating points, thereby further reducing fuel consumption and improving emission performance. Table 4 shows the simulation results of fuel economy and emissions comparison under the FTP75 cycle. It can be seen that the performance is quite different under NEDC and FTP cycles, indicating that the driving conditions have a great impact, and the control strategy should be adjusted based on the driving conditions. The fuel economy and emission performance of the F-EMS are relatively stable under different NEDC and FTP75 cycles, indicating that the fuzzy control itself is robust to changes in driving conditions. The NNF-EMS performance of DCR-based HEV is always the best under different driving conditions, indicating that the impact changes in driving conditions on the EMS optimization of HEVs cannot be ignored.
Simultaneously, it can be seen from Tables 3 and 4 that, compared to NEDC driving conditions, DCR-based NNF-BMS improves the fuel consumption and emission performance of the vehicle by more than 30% under the FTP75 driving cycles, but not the power consumption, which was reduced from 9.6751 kWh to 9.3127 kWh. It shows that the application of DCR-based NNF-EMS is more obvious under real transient conditions.  Table 4 shows the simulation results of fuel economy and emissions comparison under the FTP75 cycle. It can be seen that the performance is quite different under NEDC and FTP cycles, indicating that the driving conditions have a great impact, and the control strategy should be adjusted based on the driving conditions. The fuel economy and emission performance of the F-EMS are relatively stable under different NEDC and FTP75 cycles, indicating that the fuzzy control itself is robust to changes in driving conditions. The NNF-EMS performance of DCR-based HEV is always the best under different driving conditions, indicating that the impact changes in driving conditions on the EMS optimization of HEVs cannot be ignored.
Simultaneously, it can be seen from Tables 3 and 4 that, compared to NEDC driving conditions, DCR-based NNF-BMS improves the fuel consumption and emission performance of the vehicle by more than 30% under the FTP75 driving cycles, but not the power consumption, which was reduced from 9.6751 kWh to 9.3127 kWh. It shows that the application of DCR-based NNF-EMS is more obvious under real transient conditions.

Conclusions
The driving pattern recognition results affect the EMS directly, and the driving condition is complex in NEDC, UDDS, etc. Some results are now shown in Figures 10 and 11, because the "driving cycle block" is defined differently under different conditions, as shown in Figure 8 for the definition of NEDC; however, the NNF-EMS is still efficient for other conditions. Although there are other existing EMS techniques in the literature that were not designed in this paper in comparison with our proposed method, such as dynamic programming, it can be used as a benchmark for evaluating the

Conclusions
The driving pattern recognition results affect the EMS directly, and the driving condition is complex in NEDC, UDDS, etc. Some results are now shown in Figures 10 and 11, because the "driving cycle block" is defined differently under different conditions, as shown in Figure 8 for the definition of NEDC; however, the NNF-EMS is still efficient for other conditions. Although there are other existing EMS techniques in the literature that were not designed in this paper in comparison with our proposed method, such as dynamic programming, it can be used as a benchmark for evaluating the proposed method; however, it is not useable in real-time, and the focus of this article is how to improve current fuzzy control methods with high availability. This paper does not excessively emphasize and compare global and real-time optimization methods of energy management.
Looking at the literature, the concept of using driving data and traffic information in an adaptive fuzzy inference system for energy management of HEV is not new. In this paper, by analyzing the characteristic parameters of the driving cycles and learning using neural network samples, the DCR is realized, and the recognition results are used as the reference input for the fuzzy control strategy. Therefore, the membership function in the fuzzy controller is optimized to solve the problem of poorly targeted driving cycles. The proposed method can adopt the optimized control strategy in different driving cycles, and its robust substitution and performance are better than the traditional fuzzy control strategy. The proposed NNF-EMS can realize the adaptive optimization of fuzzy membership function and fuzzy rules under different driving cycles. Different from intelligent algorithms such as global optimization and real-time optimization, from the perspective of real vehicle application, fuzzy control is more practical. Thus, it will have good practical application value in improving the fuel economy of hybrid electric vehicles.
For electric vehicles, especially battery electric vehicles with limited energy, the optimization of energy management is crucial. With the vigorous development of electric vehicles toward unmanned vehicles and connected vehicles, on-board electronic devices such as sensors and high-performance computers that consume large power put forward higher requirements for vehicle energy management. Road condition information not only affects driving safety but also directly affects fuel and/or electricity consumption. Intelligent energy management strategies that take into account information prediction of road/operating condition will become a hot research topic in the area of vehicle energy management in the future. Artificial intelligence methods are also widely applied, such as deep learning, reinforcement learning, and convolutional neural networks, and further research is urgently needed.