1. Introduction
In recent years, with the development of the automobile industry, the vehicle has not only improved human life but also caused a significant increase in energy consumption and urban haze. Currently, the automotive industry has adopted the energy structure transformation and will gradually replace the internal combustion engine (ICE) with multiple energy sources. The existing industrial base, hybrid electric vehicles (HEV), and electric vehicles (EV) are one of the best solutions to this problem at this stage [
1,
2,
3]. Among them, HEV refers to vehicles with two or more power sources, such as the most widely used vehicles driven by a mixture of internal combustion engine and electric motor, and the type and capacity of the battery in this system will directly affect the intensity of the mixture and the space available for optimization. Compared with EV, HEV can effectively alleviate the technical bottleneck of limited battery storage and the low range of pure electric vehicles through multiple power sources. Therefore, for HEV, the question of how to design an effective EMS to reasonably distribute the power sources is the key to improving energy utilization and fuel economy [
4].
The current research on EMS can be classified into rule-based algorithms and optimization based algorithms [
5,
6]. Rule-based EMS is developed based on empirical rules summarized by logic extracted from experts or experimental data [
7,
8]. The rule-based EMS is practical because of its simple structure and reliable control performance. The most typical one is the Charge Depletion–Charge Sustaining(CD-CS) strategy [
9]. In the CD phase, the power required by the vehicle is provided by the battery and ICE is started only when the required drive power exceeds the peak output of the motor. As SOC decreases to the SOC objective value, the operating mode switches to the CS phase. In the CS phase, ICE starts, and vehicle driving is completed by the ICE and motor in conjunction while keeping SOC fluctuating around its target value. CD-CS, while offering significant improvements in fuel economy over short distances, simply distributing power as the driving distance increases does not ensure that the power source is working in the optimal efficiency zone. At the same time, the development of the strategy is completely dependent on expert experience. Therefore, to address the limitations of the above rule-based strategies, EMS based on optimization algorithms has received attention from many scholars and companies [
10].
According to the optimization objective, EMS based on optimization algorithms can be categorized into global optimization and instantaneous optimization [
11]. Global optimization is used to compute the optimal control volume by minimizing the sum of the objective functions. Dynamic programming (DP) [
7,
12,
13], Pontryagin’s minimum principle (PMP) [
14,
15], and intelligent algorithms [
16,
17] are normally used as global optimization. In EMS applications, all future driving information, such as vehicle speed profiles, road geography information, etc., is usually obtained in advance, and then the global optimization method is used to achieve the power allocation of the engine and motor by minimizing the objective function. However, the global algorithm is difficult to apply to real-time strategy optimization due to complex road condition information and a large number of iterative operations required. To reduce the computational burden and apply it to real-time optimization, instantaneous optimization strategies have been proposed and studied by many scholars [
18,
19].
Instantaneous optimization algorithms are used to determine the control variables by minimizing the instantaneous objective function. The equivalent consumption minimization strategy (ECMS) is the most representative instantaneous optimization method. This method converts electrical energy consumption into equivalent fuel consumption at each moment and uses it as an optimization objective. In the literature [
20], ECMS was used to minimize the fuel consumption of HEVs by power allocation between ICE and electric motor, and the results showed that the strategy can effectively lead to a reduction in fuel consumption. Similar studies were conducted by Gao et al. and Rousseau et al. The results showed that ECMS can produce near-optimal results for fuel consumption minimization even without driving information [
21,
22]. Mursado et al. proposed an adaptive equivalent consumption minimization strategy (A-ECMS) for real-time energy management in hybrid vehicles, which continuously changes the equivalence factor according to the road load conditions to obtain an approximately optimal control signal for maintaining the charge. By comparing the results obtained from the A-ECMS controller with those obtained from the dynamic programming optimal controller, the authors conclude that the use of an equivalent fuel consumption minimization strategy, which is much simpler than dynamic programming, can lead to a suboptimal solution that differs little from the optimal solution [
23]. However, the above strategies do not apply to the complicated driving conditions in which most vehicles are driven. Therefore, adaptive energy management strategies are more important to optimize vehicle performance under real driving conditions where there are no predefined driving cycles.
Driving condition recognition methods can be roughly divided into three categories, which include driving condition recognition based on neural network theory, driving condition recognition based on cluster analysis, and driving condition recognition based on fuzzy controller, and the driving condition recognition technology is mainly combined with the energy management strategy of rules and the energy management strategy of instantaneous optimization. In reference [
24], driving condition recognition consists of two parts: (1) driving condition information extractor and (2) driving environment recognizer. In the literature, the driving condition information extractor extracts 16 feature parameters, and the driving environment recognizer includes a road type recognizer, a driving style recognizer, a driving trend recognizer, and a driving pattern recognizer. Among these parameters, 11 typical driving conditions were selected and constructed by the learning vector quantization (LVQ) neural network algorithm in the road type recognizer. The latter three recognizers (environment, driving trend, and driving style) were implemented by a fuzzy controller. In the literature [
25], a learning vector quantization (LVQ) neural network is introduced for designing driving pattern recognizers based on the driving information of the vehicle. This multimodal strategy can automatically switch to a genetic algorithm optimized strategy for a specific driving condition depending on the difference in road condition recognition results. In reference [
26], a new hierarchical clustering method is used to divide the duty cycle data into four groups, which are extracted from a sample of historical driving condition cycles. A support vector machine approach is then used to predict the current driving cycle based on the classification results. Finally, a switchable drive controller is built based on the current operating cycle and slope information. Several papers have investigated the relationship between computational accuracy and complexity. In reference [
27], the k-nearest neighbor algorithm is used to study the speed extracted from facility-based driving cycles. Cross-validation techniques have been used to evaluate the effect of window length on classification accuracy. The final selection was made with 10 driving modes and a window length of 60 s. In the literature [
28], a K-means clustering algorithm was used to classify the driving blocks. A novel driving pattern recognition method was designed by combining variational pattern score (VMD) and extreme learning machine (ELM).
In summary, the rule-based energy management strategy does not require high hardware conditions for the controller and has good real-time and robustness. However, various thresholds in the strategy are dependent on expert experience, and the driving distance is mostly greater than the maximum distance of pure electricity in the actual environment. The driving conditions are complex and variable, and it is difficult for the simple allocation strategy to play a good fuel-saving potential. The global optimization strategy in the optimization-based energy management strategy can certainly achieve the overall optimal results. However, entire driving conditions cannot be predicted and its computation is huge; thus, it is only suitable for offline optimization of fixed driving conditions and cannot be controlled online in real-time. Although the transient-based optimization strategy can achieve real-time optimization, the determination of the fuel-electricity conversion factor is less adaptable to complex driving conditions. Moreover, there is little literature that considers the influence of driving distance on energy management strategy in the above strategies. Therefore, if an energy strategy can be designed to achieve real-time optimization and adapt to different driving conditions and driving distances, the fuel economy of the vehicle can be further improved.
Motivated by this, this paper proposes a real-time energy management strategy based on FFELM driving condition recognition by combining the energy management strategy based on transient optimization and driving condition recognition technology. The proposed strategy can realize real-time online control and adapt to different driving conditions, and good fuel economy can be achieved. Firstly, feature data of different typical driving conditions are collected and counted, and an extreme learning machine based on feature fusion is used to classify and identify the data. Then, ECMS is used to optimize the different typical driving conditions offline, and initial SOC, driving distance, and power distribution under the optimization results of each typical condition are fitted and quantified. Based on the above, the FFELM-based driving condition identifier and the fitted quantized results are applied to real-time energy management. A real-time control strategy that can dynamically adjust the power distribution to adapt to different driving conditions is designed.
The main contributions of this paper are as follows: (1) To improve the accuracy and robustness of the driving cycle identification technique, an extreme learning machine that can fuse features is used to classify and identify different types of typical driving condition data. (2) To analyze the effect of driving condition type and driving distance on energy management strategy, ECMS is used for offline optimization of each typical driving condition, and the results are analyzed and quantified. (3) To reduce the computational burden and improve the fuel economy in real-time situations, a real-time energy management strategy is designed that can be adapted to different driving conditions. The simulation results demonstrate that the proposed strategy can achieve real-time power allocation and improve energy utilization under complicated driving conditions.
4. EMS Based on Driving Condition Identification
According to previous studies, the initial values of battery SOC before driving, the type of driving conditions, and the driving distance have a great influence on the real-time energy management strategy of the vehicle. Therefore, this paper proposes a real-time energy management strategy that integrates three factors to improve fuel economy. As shown in
Figure 8, the proposed strategy is divided into three parts: (1) offline optimization strategy, (2) online optimization strategy, and (3) real-time driving condition identification.
According to what is discussed in
Section 3.1, driving conditions are classified into six categories based on the characteristic parameters of each condition, and FFELM is used for training classification recognition. In addition, the initial values of battery SOC and driving distance are very important for vehicle energy strategies; therefore, in order to combine these two, the equivalent distance parameter is introduced according to the literature [
14]. Considering that there is a relationship between the remaining power and the remaining distance, the expression is as follows:
where
is the remaining distance,
is the maximum distance in pure electric mode,
indicates the remaining charge,
is the instantaneous value of the battery charge during vehicle driving, and
is the range of the available SOC.
and
are the upper and lower battery limits, respectively.
indicates that the battery can complete the remaining distance in pure electric mode. When
, the battery has insufficient power left to complete the rest of the distance in pure electric mode, and the engine is required.
4.1. Offline EMS Optimization
In order to adapt the control strategy to different driving conditions, ECMS is applied offline to optimize the optimal allocation for each typical driving condition in preparation for the application of the online strategy.
4.1.1. Objective Function
According to the actual driver demand power of the vehicle, the actual output power of the engine and motor is reasonably allocated within the power range of the engine and motor such that the sum of the equivalent fuel consumption of the instantaneous fuel consumption of the engine and the power consumption of the motor is minimized [
35]; thus, the objective function can be shown as follows:
where is
J the total consumption,
and
denote the initial time and final time, respectively,
x and
u are the system state and control variables respectively,
is the instantaneous fuel consumption of the engine, and
is the instantaneous equivalent fuel consumption after power conversion. In order to calculate equivalent fuel consumption more accurately, the equivalent charging factor and equivalent discharging factor are introduced, and the equivalent fuel consumption of battery power can be expressed as follows:
where
is the fuel mass calorific value.
Since battery SOC’s balance is not well maintained by ECMS alone, a penalty function is introduced to correct the equivalent fuel consumption to maintain it close to the objective SOC [
14]. The basic value of the penalty coefficient is taken as 1. When battery SOC is close to the target SOC, the equivalent fuel consumption of the motor power is basically not corrected so that the strategy can reasonably power the engine and motor according to the lowest equivalent fuel consumption. When the battery SOC is higher than the target SOC value, the penalty coefficient is less than 1, and the equivalent fuel consumption of motor power consumption is reduced by the penalty coefficient so that the control strategy is more inclined to use electric power. When battery SOC is lower than the target SOC value, the penalty coefficient is greater than 1, and the equivalent fuel consumption of motor power consumption is increased by the penalty coefficient such that the control strategy is more inclined to use fuel. Moreover, the penalty coefficient in the target SOC near the value of the change should be relatively gentle so that the penalty coefficient on the power distribution can reduce impact; battery SOC deviates from the target SOC, and the speed of change of the penalty coefficient should be intensified by speeding up the response as soon as possible to maintain power balance. In order to make better and effective use of SOC, the penalty function is introduced:
where
is the empirical coefficient; here, 0.2 is chosen;
is the objective value of SOC,
is the minimum value of SOC, and
is the SOC at time
t; as shown in the figure, when the penalty factor decreases with the increase in SOC, the modified equivalence factor is as follows.
The corrected power equivalent fuel consumption is as follows.
4.1.2. Control and State Variables
Since this paper is aimed at a single-axis parallel structure where the engine and motor are located on the same shaft and the engine speed and motor speed are the same, the distribution of driver demand power can be translated into an equation for the distribution of engine and motor torque. In order to better achieve torque distribution, the power distribution factor is expressed as the formula of the ratio of engine torque to demand torque, and
is introduced. When
, the engine and the motor drive together, and when
, the engine drives and it drives motor to charge the battery. When
, the engine torque is the same as the demand torque and the vehicle is in the engine drive-alone mode. When
, the engine is not involved in vehicle driving:
where
is the demanded power,
and
are the engine power and motor power, respectively,
is the demanded torque, and
the demanded speed. Therefore, it can be further expressed as follows.
Optimal control variables are stated as follows.
Changes in state variables with battery power SOC are defined as follows.
The optimization of different initial values of SOC under different driving conditions is conducted by ECMS. As shown in
Figure 9a, the optimized data including the power distribution factor, the rate of change of battery SOC, and the driving distance are sorted. The variation of battery SOC and driving distance under a driving condition can be used to obtain the optimal power distribution factor under this condition.
4.2. Real-Time Driving Condition Recognition and Calculation
Based on the current driving conditions, FFELM is used to identify the conditions and to calculate the equivalent distance parameters based on the current battery’s SOC and the distance to the destination. As shown in
Figure 9b, the parameters of the operating conditions characterized by 120 periods are calculated, including maximum speed, maximum acceleration, maximum deceleration, etc. These parameters are used as the basis. FFELM identifies the current operating conditions. The equivalent distance is calculated based on the Equation (
23), and it is then input to the vehicle controller during the vehicle driving process.
4.3. Online Strategy Optimization
Online torque distribution is achieved by identifying the type of driving condition and the equivalent distance parameters according to the required torque under current driving condition. As shown in
Figure 10, online torque distribution is achieved.
6. Conclusions
In this paper, a real-time energy management strategy based on driving condition recognition is proposed. First, an FFELM identification method with high stability and accuracy is used to train six typical driving conditions, which provides the basis for identification under combined driving conditions. Second, the six typical driving conditions are optimized offline using ECMS. By collating the data of optimization results under each driving condition, the optimization results were analyzed based on the equivalent distance coefficients of battery SOC and driving distance. The optimal power distribution factor for each condition optimization is also provided by fitting the data to help the real-time management strategy. Finally, to demonstrate the effectiveness of the proposed strategy, the real-time energy management strategy under mixed driving conditions is used and compared with CD-CS strategy and ECMS strategies.
The comparison between ELM and FFELM for raw data identification shows that FFELM is more accurate and stable than ELM. In the simulation experiments for verifying the effectiveness of the proposed strategy, it is shown that the proposed strategy improves fuel economy by 10.21% and the computation time is slightly longer than that of CD-CS. This means that the proposed strategy has a great potential to save fuel compared to the rule-based strategy and is also effective in real-time control. The equivalent fuel consumption of the proposed strategy is 7.12 L/km compared to the equivalent fuel consumption of ECMS of 7.31 L/km, which is a 2.5% improvement. This shows that the proposed strategy can be adjusted in real-time to achieve optimal fuel economy. Therefore, the proposed strategy has great advantages in fuel economy in real-time and practicality for different driving conditions.
Although the proposed strategy has greatly improved real-time fuel economy and the application of complicated driving conditions, the extraction of typical driving cycle data in this paper are all from the European standard driving cycles and the selection of the features of the road condition data has a great influence on the recognition accuracy of the driving conditions. Therefore, in future work, we will collect the driving condition data that meet the actual driving characteristics of Chinese roads and establish the driving condition recognition model on this basis to improve recognition accuracy.