Energy Management Strategy for Plug-In Hybrid Electric Vehicles Based on Driving Condition Recognition: A Review

Abstract

Appropriate energy management strategies (EMSs) have been selected for plug-in hybrid electric vehicles (PHEVs) based on driving-condition recognition (DCR) according to the acquired driving information, so as to minimize the target parameters. With online control and offline optimization, the strategy is suitable for real-time applications and is of great significance to repetitive routes, owing to its simplicity and ease of implementation. This paper aims to identity the DCR-based EMSs, develop efficient EMSs, and invite researchers involved in this field to discover new solutions. This paper presents a comprehensive analysis of EMSs based on DCR in terms of principles, the scope of application as well as their advantages and disadvantages, and provides a thorough survey of the latest progress in EMSs. We concluded that DCR-based EMSs show an improvement in energy saving and the pollutant-discharging effect.

1. Introduction

PHEVs, which are transition vehicles between conventional vehicles and electric vehicles, possess great potential in real-life cycles. There is a larger margin of efficiency improvement from conventional vehicles to PHEVs, which could also compensate for the technical shortcomings of fuel cells. As vehicles powered by at least two power sources, PHEVs usually involve an internal combustion engine and an electric motor, with external power supplies. There are different intrinsic characteristics between an engine and electric motor, such as frequency response characteristics, torque characteristics and working efficiency. Therefore, it is difficult to collocate two power sources to meet the requirements for vehicle performance and energy economy and to keep the state of charge (SOC) at no less than the pre-set value through online implementation. All the above mentioned problems should be addressed by EMSs.

According to the control methods, EMSs can be classified into two categories. The EMS framework is illustrated in Figure 1. Deterministic rule-based strategies were first applied to the energy management of PHEVs. Banvait et al. [1] detailed a charge depletion–charge sustaining (CD–CS) strategy instead of the original electric assistance strategy. The speed-switching power, following the auxiliary power unit cooperative control strategy, which is forced on the runs along with the optimal brake specific fuel consumption obtain good curve, was presented by Shen et al. [2]. Fuzzy logic belongs to intelligent control methods, dispenses with accurate mathematical models of controlled systems, possesses a self-learning ability, as well as high adaptivity and robustness. So, it is widely utilized in complex nonlinear problems [3,4,5]. Denis et al. [6] designed a Sugeno-type fuzzy logic controller by taking the moving average of the past speed and the current global discharge rate as the inputs so as to benefit from the trip information. Likewise, an adaptive-equivalent consumption minimization strategy blending a fuzzy inference system to improve the self-adaptation was provided by Li et al. [7]. For the fulfillment of global optimization, various algorithms have emerged. The main optimization algorithms encompass dynamic programming (DP) [8,9,10], Pontryagin’s minimum principle (PMP), stochastic dynamic programming (SDP), genetic algorithm (GA), particle swarm optimization (PSO), and an aerodynamic traffic flow model. Gong et al. [11] developed a two-stage dynamic programming algorithm, which could improve computational efficiency and fuel economy, with the historical and real-time traffic-flow parameter detection system. Du et al. [12] optimized the logic-threshold value of the rule-based EMS using the improved Non-dominated Sorting Genetic Algorithm-II (NSGA-II) to obtain the optimal logic threshold value. Real-time optimization-based EMSs mainly refer to equivalent consumption-minimization strategies (ECMSs) [13,14,15] and improved ECMSs. An adaptive, simplified EC MS-based strategy was detailed by Zeng et al. to reduce calculation load [16]. A real-time control, based on DP-ECMS, was proposed by Wang et al. to achieve the suboptimal solution [17]. Model Predictive Control (MPC), a feedback control method belonging to local optimization-based EMSs, compensates for the lack of online application of the global optimization-based EMSs [18]. Lin et al. [14] modelled the power demand from the driver as a random Markov process. Then, the optimal control strategy was obtained by SDP. A novel hierarchical-control framework was addressed by Zhang et al. to achieve the instantaneous vehicle-environment cooperative control [15].

With the development of computer science and technology, artificial intelligence (AI) is increasingly used in the energy management of PHEVs. Widely used AI technologies include Artificial Neural Networks (ANN), deep learning [19], deep reinforcement learning [20], etc. **s without describing them mathematically. Its learning rule is to adopt the steepest descent method in which the back propagation is used to regulate the weight value and threshold value of the network to achieve the minimum error sum of square. The topology of the BP neural network model includes the input layer, hidden layer, and output layer, as shown in Figure 6 [66].

Due to its adoption of the gradient method, the BP neural network experiences problems such as slowly learning convergent velocity and easily converging to the local minimum. The researcher has put forward many improved algorithms to solve these defects. Among these methods, the more typical ones are the additional momentum method, the variable learning-rate method, and the optimization of initial weight [64]. To overcome the shortcomings of a slow convergence rate, an adaptive back propagation algorithm which could update the learning rate and inertia factor automatically based on dynamical training error rate of change was introduced [67].

In the study by Zhang et al. [68], 50 s was nominated as the training times of the BP neural network, and similarly relevant characteristic parameters such as the current driving velocity, average vehicle velocity, average acceleration, average deceleration, maximum vehicle velocity, maximum acceleration, maximum deceleration, travel distance, and the travel time were regarded as the input nodes of BP network. The influences of different characteristic parameters can be adjusted by adjusting the connection weighting coefficients. The accuracy of test data of the simulation can meet the demand of experiment. In another case investigated by Zhang et al. [69], 100 was nominated as the maximum number of iterations. Seventy percent of sample data were utilized for learning and training, fifteen percent for validation, and the remaining fifteen percent for testing. After 88 iterations, the error was reduced to less than 0.0018, and the results demonstrated the accuracy of the used neural network. The relationship between the sampling period and recognition accuracy was studied and the author discovered that the maximum recognition accuracy reached 99.99% when the sampling period was 275 s. The most successful achievements of the BP neural network were made while the accuracy of the network decreased due to problems such as the slow convergence rate and the possibility of falling into local minimum caused by the randomness of its initialization. Moreover, Jiang [70] took advantage of the PSO algorithm (simple principle, few required parameters, fast convergence speed and easy implementation) to optimize the initialization process of the BP network. The simulation results demonstrated that the accuracy was improved to 92% by the optimized BP network.

From the abovementioned studies, it can be seen that there are many factors, such astraining times, the connection–weighting coefficients, sampling period, prediction period and the initialization process, that are able to influence the accuracy of the BP neural network. Before applying the BP neural network to DCR, it is necessary to set the parameter values above properly.

3.3. DCR Based on PNN

PNNs are evolved from radial basis networks, belonging to feed-forward networks. Its advantages include a simple learning process, fast training speed, accurate classification, fault tolerance and so on. In a PNN, there is no need for massive back-propagation training computations. Instead, each data pattern is represented with a unit that measures the similarity of the input patterns to the data pattern. It is composed of input layer, pattern layer, summation and decision layer, as shown in Figure 7 [71].

However, its advantage of a fast training speed comes at the cost of requiring a large memory capacity for operation. Researchers explored ways to overcome this shortcoming. In [72,73], PCA was used to reduce the dimension of the input vectors. Moreover, in order to eliminate the defect of experience value, the key parameter of PNN was optimized by the improved single-target optimization genetic algorithm [74]. A PSO algorithm was proposed to optimize the smoothing factor in PNN [75].

Hu [76] and Bai [77] et al. adopted PNN in DCR. In Bai’s paper, the compound bisected method was used to divide working conditions with a 120 s cycle, and five characteristic parameters were selected. The identification results indicated that PNN had the ability to identify the category of test-driving conditions accurately. Compared with methods for DCR introduced above, PNN is less frequently used in DCR.

3.4. DCR Based on Fuzzy Reasoning

The fuzzy reasoning method is implemented with rules that evolved from controllers’ experience but without mathematical models. This strategy is of benefit due to its simplicity and ease of implementation, with execution issues as follows: (1) the setting of membership functions, (2) the definition of fuzzy inference rules. Due to various characteristics of identified driving conditions, different membership functions and fuzzy rules are selected for different simulations.

To increase its ability to manage unclearness, improved fuzzy reasoning with fuzzy knowledge-based systems was carried out [78]. The membership degree was represented by an interval as an alternative of a fuzzy set to remove the drawback of the algorithm [79]. A new fuzzy reasoning method was proposed that employed an explicit feedback mechanism to improve the robustness of fuzzy reasoning [80].

The fuzzy reasoning method was used by Yang [81]. In the literature, the language set and membership function of maximum speed, average speed, maximum acceleration and maximum deceleration were established with MATLAB/Simulink fuzzy rule control box. Next, fuzzy inference rules were made for the simulation based on accumulated experience. Finally, the fuzzy identification results were fuzzed properly to obtain the membership functions of the output parameters. Three identified results could be acquired, namely, the congestion driving condition of the city, the smooth driving condition of the city, and the suburban/high-speed condition. The emulation verified the effectiveness of the designed fuzzy recognizer. Likewise, Zhou [82] set four input variables and one output variable for the fuzzy DCR. The input variables and the output variables were four characteristic parameters and the recognition results, respectively. Additionally, the effectiveness of the fuzzy recognizer was also verified.

Furthermore, the fuzzy reasoning method was also applied to drivers’ driving style recognition [83,84] and the vehicular-collision tendency estimate [85].

3.5. DCR Based on Cluster Analysis

Given the shortcomings of neural networks and fuzzy recognition algorithms, a DCR method based on a cluster analysis is proposed. Driving conditions could be recognized by calculating characteristic parameters of driving conditions to be identified as well as the distance between conditions and clustering centers of all driving conditions. Driving conditions belong to the category nearest to them.

In addition to the partition-based cluster analysis above, the hierarchies-based cluster analysis was proposed by Karypis et al. [86], which does not require the predefined number of groups. To find groups of arbitrary shape and eliminate noise effectively, the density-based cluster analysis was introduced [87]. To improve computational efficiency, the grid-based cluster analysis was discussed by Wang et al. [88].

The K-means algorithm is most commonly used in the area of driving conditions clustering. Lv et al. [89,90] verified the accuracy of the method. They selected three parameters of average vehicle speed, average absolute acceleration, and idle speed time ratio as characteristic parameters. The accuracy of the method was determined by comparing the results of random conditions to be identified with themselves. However, the K-means algorithm were flawed, without clear distribution boundaries and handling samples with small differences. In view of this situation, **e et al. [91] applied density based spatial clustering of applications with the noise (DBSCAN) algorithm as the cluster algorithm that can reduce the influence of isolated points on the accuracy of cluster-center coordinates estimation. The clustering performance, according to the DBSCAN clustering result, was better than the result based on the K-means algorithm.

3.6. DCR Based on SVM

SVM is a classifier developed from a generalized portrait algorithm in pattern recognitions. Compared with neural networks and other methods, SVM has its unique advantages in pattern recognition due to outstanding learning performance and good capabilities in generalization. Therefore, the method has extensive application prospects in the fields of time prediction, classification, density estimation and so on [92]. An optimized hyperplane for classification has been developed to maximize the area between it and the adjacent samples [93]. Figure 8 provides the sample graph of SVM.

SVM has been developed for two classes of linearly separable sample data. However, there are still nonlinear and indivisible data that are in a nonideal state in practical application. Therefore, a kernel function was designed and applied to SVM. Considering the differences in generalization and learning ability of different kernel functions, the mixed kernel function was proposed [94].

Bu [92] applied SVM to DCR. In contrast to the control strategy without driving-condition recognition, it provided considerable improvement in fuel economy. To improve the application efficiency, Zhang et al. [95] proposed an extended SVM with an embedded feature-selection ability, allowing the driving-condition discrimination system to achieve higher recognition efficiency and robustness. However, the current forecast methods predict the speed mostly based on traffic conditions while ignoring the impact of the driver–vehicle–road system on the actual speed profile. To compensate for the above deficiencies, a model was constructed by Li et al. to effectively incorporate the factors of the driver-vehicle-road traffic system for a faithful speed representation, such as driving habits (driver), vehicle state (vehicle), geographic information (road) and traffic flow (traffic) [96]. The optimized genetic algorithm-SVM (GA-SVM) vehicle-speed prediction model was established in accordance with the optimized GA-SVM prediction algorithm at different times. Moreover, the validity and rationality of the prediction algorithm were confirmed by the test results. To improve the accuracy of classification and recognition, Zhang [97] applied the PSO algorithm to optimize the core parameters of SVM. Compared with the BP neural network, the proposed algorithm achieved a better recognition and classification effect.

3.7. Other Identification Methods of Working Conditions

In addition to the methods mentioned above, various methods have been applied to DCR. For instance, Lu [98] established a driving-state recognition model based on the naive Bayesian classifier. Hailong et al. [99] proposed a method of driving state recognition based on the PSO algorithm. Ba [35] used the random forest algorithm to achieve real-time performance, laying a foundation for the implementation of the subsequent coupled energy management strategy established in the paper.

Despite the wider usage, the application of such classification algorithms in DCR is yet to be well compared. Due to different driving conditions, weights of neural networks, number of neurons and layers and confidence-intervals overlap, it would be difficult to make a fair appraisal of DCR algorithms’ performance. Nonetheless, all the foregoing DCR approaches are summarized in

Table 1. Main DCR methods characteristics summary.

Methods	Main Advantages	Main Disadvantages
LVQNN	High learning, adaptive and fault-tolerant ability Fast computation	Space waste and iteration time increase (Some neurons cannot be enabled) Accuracy: initial weights selected by experience
BPNN	High learning and fault-tolerant ability	Local minimization Slow convergence speed
PNN	Simple learning process Less training computation time High accuracy and fault-tolerant ability Escape from local optima	Accuracy: the spread value selected by experience The requirement of massive memory for operation
Fuzzy Reasoning	Simple learning process and strong robustness	Accuracy: the selection of membership degree mostly depends on experience
Cluster Analysis	Free from the constraint and interference of prior knowledge Clear division rules and fast convergence speed	Sensitive to abnormal data K value to be determined
SVM	Strong learning and generalization ability Wide comprehension Strong universality and robustness	Difficult implementation for large training pattern set and multi-classification problems

according to their principles, with their main characteristics.