Research and Comparative Analysis of Energy Management Strategies for Hybrid Electric Vehicles: A Review

Abstract

Hybrid electric vehicles have received more and more attention owing to energy saving and environmental protection. Optimized energy-management strategies are critical to improve vehicle energy efficiency and reduce the emissions of hybrid electric vehicles. This study summarized the research status of energy-management strategies for hybrid electric vehicles and analyzed the energy allocation and modeling methods of hybrid power systems. The principles, advantages, and limitations of rule-based and optimized and learning-based energy-management strategies were compared. It is found that the optimized energy-management strategies can improve fuel economy by approximately 6% compared with the rule-based energy-management strategies. The learning-based energy-management strategies can reduce fuel consumption by about 5.2~17%. This study can provide a theoretical basis and practical guidance for the efficient design and optimization of hybrid electric vehicle energy-management systems, which can promote the development and application of related technologies.

1. Introduction

With the increase in global greenhouse gases emissions and the intensification of the energy crisis, environmental protection and sustainable development have become the focus of governments and society [1]. As one of the major hydrocarbon emission sources, transportation has an increasingly urgent need for energy conservation and emission reductions [2]. Against this backdrop, hybrid electric vehicles (HEVs) have emerged as a critical solution. HEVs are vehicles powered by multiple drive systems in combination, with the most common type being the gasoline–electric hybrid. They can leverage the high energy density of fossil fuels and the renewability of electricity [3], which are the key driver for green transportation.

Series, parallel, and series-parallel HEVs are the typically feature core architectures, which are defined by distinct mechanical configurations. The engine is mechanically decoupled from wheels in series HEVs to drive only a generator [4]. In parallel HEVs, the traction power can be supplied by the internal combustion engine alone, the electric motor alone, or both acting together [5]. Series–parallel HEVs integrate mechanical components to dynamically switch between series and parallel operation modes. Each architecture exhibits unique energy flow dynamics and loss patterns. In series HEVs, energy flows electrically (engine→generator→battery/motor) with losses from generator inefficiency and battery cycles. For parallel HEVs, energy flows mechanically from the engine, electrically from the motor, or both. Losses arise from engine operation outside its optimal efficiency zone or inefficient power splitting between sources. Series–parallel HEVs enable flexible energy paths but face complex losses from coordinating engine-motor operations and battery management. Each architecture’s energy dynamics demand tailored energy-management strategies (EMSs) to minimize waste and optimize performance.

The EMS optimizes the power allocation from multiple sources to meet the driving demands. The core objectives of EMSs include minimizing fuel consumption, reducing harmful emissions, and extending the battery lifespan by managing the state of charge (SOC). Balancing these objectives requires dynamic prioritization to optimize HEV performance, sustainability, and component longevity. Effectiveness is typically evaluated against baselines like traditional rule-based strategies with metrics of fuel economy, equivalent hydrogen consumption, and computational latency. These metrics are quantifiable. In the field of HEV EMSs, the rule-based, optimized, and learning-based strategies are the research hotspots [6]. Each strategy is grounded in distinct theoretical frameworks and practical requirements.

The rule-based strategies are characterized by a series of fixed rules extracted from experienced designers or optimized rules [7]. Deterministic rules play a crucial role due to their simplicity, reliability, and wide-spread acceptance. Although deterministic rule-based strategies dominate in this field, our study mainly focuses on the fuzzy rules. As an important branch of the rule-based strategy, the fuzzy rules innovatively introduce the concept of fuzzy logic, which allows the input variables and the output variables to have fuzziness and uncertainty. This characteristic makes the fuzzy rules more flexible in dealing with the complex conditions and dynamic changes.

The use of fuzzy rule-based strategies can effectively reduce fuel consumption and emissions. Yang et al. proposed a fuzzy logic rule-based EMS, which fully considered the real-time storage and release of the flywheel’s kinetic energy. The strategy effectively reduced the equivalent fuel consumption and exhaust emissions [8]. This demonstrated the direct benefits of fuzzy rules in optimizing energy-management systems. Building on this, some researchers have made improvements and expansions to fuzzy rules from different perspectives. Mazouzi et al. used the genetic algorithm (GA) to optimize the fuzzy logic parameters, which significantly improved the fitness function and highlighted the high sensitivity of fuzzy logic systems to optimize the parameters [9]. Bo et al. utilized an adaptive network-based fuzzy inference system (ANFIS) to construct the optimal control strategy and objective function. Online training on the fuzzy rules and the parameters was conducted via the Q-learning algorithm. The gradient descent methods were conducted, which significantly reduced the training time [10]. Luca et al. proposed a mutant fuzzy controller, which can output a membership function based on the degradation changes of the fuel cell to extend the fuel cell’s lifespan [11]. Vignesh et al. proposed an EMS based on the ANFIS, which integrated the intelligent control logic and used a combination of three standard driving cycles for training. It can significantly improve fuel efficiency [12].

The optimized strategies are based on mathematical models, which can achieve the optimal solution for energy distribution through the global or local optimization to improve the system efficiency. The dynamic programming (DP) methods and model predictive control (MPC) are widely applied [13].

DP is a method of global optimization. It obtains the theoretically global optimal solution by conducting an exhaustive search of all possible energy distribution paths. This characteristic makes DP a powerful tool for finding the best solutions in complex optimization problems. Using DP in vehicle control systems can effectively reduce fuel consumption by optimizing driving strategies. For instance, Xu et al. utilized the knowledge extracted from DP to effectively improve the control performance of the EMS based on the soft actor-critic (SAC) algorithm, which reduced energy consumption [14]. Tang et al. proposed an adaptive DP approach that utilized the heuristic DP and dual heuristic programming. The approach reduced fuel consumption and ensured the real-time planning speed, as well as good tracking performance [15]. Liu et al. introduced the reference information of DP during the training process of reinforcement learning (RL) and constructed a reward function related to the optimal trajectory to implement imitation learning, which significantly improved fuel economy [16]. Moreover, during the RL process, training the agent with the optimal solution of the DP can alleviate the cold start problem caused by random exploration [17].

In contrast to the global optimization nature of DP, MPC exhibits the characteristic of local optimization. It is based on the vehicle powertrain model, which predicts the vehicle’s driving conditions over a future period and optimizes the energy distribution strategy according to the prediction results. Zhou et al. proposed an EMS that fused the whale optimization algorithm, DP, sequential quadratic programming (SQP), and the back propagation neural network (BPNN). In the MPC controller of the EMS, the multi-collision method was used to discretize the control variables of the optimal control problem during the prediction period, which successfully reduced costs and improved the system performance [18]. Han et al. proposed a predictive EMS that took the electric motor (EM) thermal control into account. The study put forward an MPC framework based on the pontryagin’s minimum principle (PMP). This framework effectively improved the fuel economy and ensured that the EM temperature remained below the limit to feature high computational efficiency [19].

The optimized strategies apply some algorithms to enhance the scientific nature and the efficiency of energy distribution. The particle swarm optimization (PSO) algorithm abstracted the potential solutions to the problem as particles with each particle representing a solution, which conducted an iterative optimization in a multi-dimensional search space. The hierarchical EMS proposed by Tang et al. achieved an energy consumption reduction while ensuring the safe and efficient passage of vehicles in uncertain traffic scenarios through scenario-adaptive speed planning, offline optimization of the ANFIS, and the PSO algorithm [20]. The Lyapunov algorithm analyzed the change trend of the system state by constructing an appropriate Lyapunov function, which judged the stability of the system. Li et al. optimized the EMS with the Lyapunov algorithm, achieving better performance compared with the adaptive equivalent consumption minimization strategy (AECMS) [21]. The golden section search (GSS) can effectively find the optimal solution within a specific interval and perform optimization within the defined search interval. Yang et al. proposed a collaborative EMS based on the variable optimization domain, which adopted a sub-cycle division method based on the historical driving data and was optimized by the GSS algorithm, outperforming the conventional AECMS [22].

The learning-based strategies leverage artificial intelligence technologies to achieve adaptive control in a data-driven manner. The strategy can keenly sense and adapt to the complex and changeable operating environments. The cutting-edge technologies, such as RL and neural networks, are the research focus, which are applied to the strategy optimization under complex working conditions.

The learning-based strategies can effectively reduce fuel consumption, improve convergence speed, and mitigate battery degradation. Zou et al. applied the prioritized replay strategy in the deep Q-network (DQN), which adopted the normalized advantage function, priority rule, and the Markov chain working condition prediction method. The method conferred fast convergence speed and accurate working condition prediction to the model [23]. The EMS proposed by Chen et al. embedded the fuzzy logic control (FLC) as an expert demonstration into the learning process of the twin delayed deep deterministic policy gradient (TD3) [24]. It was a good balance between the operational efficiency improvement and maintenance of the SOC of the battery. The EMS proposed by Qi et al. used a hierarchical structure in deep Q-learning (DQL). Compared with non-hierarchical DQL, the training speed was significantly increased and the fuel consumption was remarkably reduced [25]. Lv et al. modified the reward function with the weight coefficients obtained through inverse RL, which fed the updated reward function into the forward RL task. The SOC change range was maintained in the region where the battery featured the high efficiency and low internal resistance [26]. Huang et al. utilized the TD3 algorithm and integrated the operating cost and energy source aging terms as the primary components into the multi-objective reward function. Compared with the EMS that ignores the energy aging factor, this approach led to significant improvements in reducing operating costs and the attenuation of the fuel cell under the suburbs city highway driving cycle (SCHDC), the urban dynamometer driving schedule (UDDS), and the new European driving cycle (NEDC) [27]. Hua et al. proposed a handshake strategy by introducing the correlation ratio in the multi-agent deep reinforcement learning (MADRL) framework. This allowed for two learning agents using the deep deterministic policy gradient (DDPG) algorithm to collaborate. Compared with traditional rule-based EMSs, this strategy significantly reduced energy consumption [28]. Qin et al. considered the thermal characteristics of the electric drive system and adjusted the weight coefficients with DDPG, which controlled the battery and the motor temperatures within a safe range and improved the overall performance of the vehicle [29]. In the neural network domain, Madhanakkumar et al. employed the hybrid waterwheel plant algorithm and dual stream spectrum deconvolution neural network, achieving notable fuel savings [30]. Chen et al. utilized the bidirectional recurrent neural network and the BPNN and trained the model with the optimal solution of DP, which can significantly improve fuel economy [31]. Nayak et al. utilized the three-layer back propagation artificial neural network, which effectively reduced the overall energy cost when the initial SOC level was 0.95 [32].

In addition, learning-based strategies have been improved from different perspectives, which demonstrated the stronger adaptability and the optimization capabilities under complex and changeable operating conditions. Huang et al. introduced deep reinforcement learning (DRL) into the MPC framework, which alleviated the inherent drawbacks of poor generalization and insufficient adaptability in DRL. The DRL significantly enhanced the robustness of economic driving decisions in unknown scenarios [33]. Chen et al. developed a multi-experience replay buffer with the improved TD3 to enhance the quality and efficiency of agent exploration [34]. The adaptive cruise control-EMS (ACC-EMS) based on the hierarchical RL proposed by Zhang et al. successfully solved the local optimum problem of the standard RL-based ACC-EMS through self-learning, which achieved an interaction to improve the training speed and stability [35]. Chen et al. adopted the improved SAC algorithm and replaced the Gaussian policy by the infinite support with the Beta policy with the finite support, which effectively reduced the cost [36]. The EMS proposed by Guo et al. was based on the FIS, which used the fuzzy baseline function to approximate the EMS strategy function and learned the strategy parameters through the policy gradient RL [37]. It had fast and stable convergence, which can adapt to environmental changes.

In HEVs, the optimized energy management can improve the vehicle’s energy efficiency and reduce fuel consumption and operating costs, which can decrease harmful gas emissions and contribute to environmental protection. Compared with previous studies, this study innovatively conducted a comparison of the energy-saving effects that different types of EMSs can achieve. It systematically examined the research status of EMSs for HEVs. Then, it conducted an in-depth comparative analysis of various strategies and provided a clear direction for future research. These efforts established a theoretical foundation for the optimization of energy management systems. They also offered practical guidance for the efficient design of HEV energy management systems. This study’s findings can directly support the development and application of related technologies. These contributions will help accelerate the widespread adoption of HEVs and advance the global goals of sustainable transportation.

2. Energy Configuration Methods of Hybrid Power Systems

Among the global energy transition, the configuration of hybrid power systems and the key components have emerged as the critical determinants.

2.1. Typical Hybrid Energy Configuration of the HEV

Hybrid power systems balance the power performance and the fuel economy via rational energy configuration and efficient use, which integrate the benefits of multiple energy sources. As illustrated in Figure 1, the three primary system architectures of the series, parallel, and series–parallel dominate the current landscape of the HEV [38]. Based on the distinct energy flow patterns and the efficiency characteristics, the designs are pivotal for developing effective EMSs to enhance the overall system performance.

2.1.1. Series Hybrid Power System

In series power system design, the integrating fuel cells and electric energy enhance the vehicle power system flexibility and efficiency. As shown in Figure 1a, the vehicles utilize the fuel and the electricity as dual energy sources. Unlike conventional internal combustion engines (ICEs), the engine in series HEVs does not directly power the drive wheels. Instead, it converts the generated electricity via the DC/DC converter to the power system, which supplies the EM for propulsion [39]. The engine can adapt its operation to real-time power demands to ensure optimal efficiency, which lacks a direct mechanical link to the wheels. The proposed approach boosts the engine performance, which can minimize energy losses and improve the overall vehicle economy.

Despite the relatively mature architecture design of the series HEVs, there is still room for improvement. Under most operating conditions, the EM needs to generate 50% of the total power required to drive the vehicle, which means that series HEVs need to be equipped with capacious and expensive batteries [42]. Moreover, the inherent multiple energy conversion stages are coupled with battery voltage and current constraints, which diminish the overall system efficiency [43].

2.1.2. Parallel Hybrid Power System

In the parallel hybrid power system given in Figure 1b, the engine and EM share a common shaft to enable joint or independent propulsion. The power coupling of this system is achieved through the dry clutch, and the EM functions as either the motor or the generator based on the power demands. Depending on its different operating modes, parallel HEVs have multiple driving modes, including engine, EM, and hybrid driving modes. These modes can be flexibly switched according to the different situations to improve the vehicle’s fuel economy effectively. In addition, the six-speed automated mechanical transmission (AMT) can match the torque-to-speed ratios between the driveshaft and the wheels, which can ensure efficient power delivery [40].

The parallel power system does not require capacious batteries resulting in a limited pure-electric driving range and greater reliance on an engine for energy. However, the multi-component architecture increases the design complexity, manufacturing difficulty, and production costs. Due to complex power coupling and switching, coordinating engine EM operation across driving modes also presents challenges.

2.1.3. Series–Parallel Hybrid Power System

In the series–parallel hybrid power system of Figure 1c, the planetary gear set efficiently connects the ICE, traction motor, generator, and driveshaft. By virtue of its high reduction ratio and the linear arrangement of the input and output shafts [44], the planetary gear enables efficient power coupling and transmission among the system components [41]. This architecture allows the ICE and motor to drive the vehicle independently, which is similar to parallel hybrid power systems. The ICE also drives the generator to supply power for motor assistance, which is similar to the series hybrid power system. As a result, series–parallel HEVs combine the strengths of series and parallel configurations, which can balance the power output, fuel consumption, and driving smoothness.

The series–parallel system incurs efficiency losses during the energy conversion process. When the electrical energy is converted into mechanical energy, a certain amount of energy loss in each component of the system reduces the overall performance below ideal levels. In addition, the frequent mode switching further challenges the coordinated control to cause short-term efficiency drops. Consequently, the power conversion efficiency may be lower than that expected in real-world driving under the high load or the complex traffic environments, which affects the fuel economy and the environmental friendliness.

Table 1. The advantages and limitations of different hybrid power systems.

Types	Advantage	Limitations
Series hybrid power system	Real-time power demand, optimal efficiency, energy losses reduction, the economy improvement	High-capacity battery, battery voltage and current constraint, overall system efficiency reduction
Parallel hybrid power system	Multiple driving modes to be flexibly switched, fuel economy improvement, small capacity battery	Short pure-electric driving range relies on an engine for energy, overall structure complexity, complicated control of the power coupling and switching process
Series–parallel hybrid power system	Power output, fuel consumption, and driving smoothness balance	Efficiency losses during the energy conversion process, frequent mode switching

is the respective advantages and limitations of the different hybrid power systems. The configurations critically inform the EMS design and guide the energy distribution and mode-switching control. To meet the current social demands for energy conservation and emission reductions, as well as the increasingly stringent requirements of users for vehicle performance, the research should address system inefficiencies, optimize energy utilization, and advance sustainable transportation.

2.2. Key Components of the Power System

The power system components are essential for the energy flow analysis and the performance characteristics of hybrid power systems. As the core energy storage unit, the power battery needs to have a high energy density to support the pure-electric driving range and should meet the cycle-life requirements for frequent charging and discharging. The power battery performance directly affects the system energy efficiency. The drive motor serves as the energy conversion interface, which can efficiently transform the electricity into the mechanical power during the operation and recover the energy via regenerative braking. The coordinated operation of the power battery and the drive motor determines the dynamic response and overall energy consumption performance of the hybrid power system.

2.2.1. Power Battery Energy Expression of the HEV

Under the worldwide harmonized light vehicles test cycle (WLTC) conditions, the error between the results of the model in series–parallel HEVs and the vehicle experiment results was less than 0.02. The battery model can reflect the changing state of the battery with relatively high accuracy [45]. According to Kirchhoff’s laws, the power, current, and charge state of a power battery can be calculated using Equation (1) [29]:

$$P_e=U_{oc}{I}_{bat}-R_{in}{I}_{bat}^2$$

(1)

where P_e is the power of the battery, W. U_oc is the open-circuit voltage, V. I_bat is the charging and discharging current of the battery, A. R_in is internal resistance of the battery, Ω.

The commonly used methods for SOC estimation are the Coulomb counting method and Kalman filter (KF) method, which are applicable to all battery systems. The Coulomb counting method calculates the SOC value by integrating the battery current over the usage time. The SOC can be calculated with Equation (2) [45] and Equation (3), respectively [46]:

$$SOC=-{\displaystyle \frac{U_{oc}-\sqrt{U_{oc}^2-4R_{in}{P}_e}}{2R_{in}{C}_{eq}}}$$

(2)

where C_eq is equivalent capacitance, F.

$$SOC(t)=-{\displaystyle \frac{{\displaystyle {\int }_{t_0}^{t}I\left(t\right)\eta dt}}{Q_N}}+SOC(t_0)$$

(3)

where SOC(t₀) is the initial SOC. I(t) is the current at time t, A. η is the Coulomb coefficient. Q_N is the rated capacity of the battery, As. t is the time, s.

The accuracy of this method mainly depends on the precise measurement of the battery current and the accurate estimation of the initial SOC. Therefore, if there are deviations in the current measurement or the initial SOC is inaccurate, errors will accumulate over time, which require additional SOC calibration [47].

The KF method is a dynamic estimation method that consists of two steps of prediction and measurement. Based on the state-space model, it integrates the real-time measured data, such as current and voltage, to estimate the SOC through the prediction-update iterative process. The state equation and measurement equation are given in Equation (4) and Equation (5), respectively [48]:

$$X_k=f(X_{k-1})=X_{k-1}-{\displaystyle \frac{{\eta }_{bat}{I}_k\Delta t}{Q_N}}+v_{k-1}$$

(4)

where X_k is the estimated SOC at time step k. f(·) is the system state function. η_bat is the charging and discharging efficiency of the battery. I_k is the battery current at the time step k, A. Δt is the sampling interval, s. v_k−1 is the noise item, which is added to the state equation.

$$Z_k=G(I_k,T_k,V_{k-1},X_k)+{\omega }_{k-1}$$

(5)

where Z_k is the voltage at time step k, V. G(·) is the system measurement function. T_k is the battery temperature at time step k, °C. V_k−1 is the battery voltage at time step k−1, V. ω_k−1 is the noise items, which is added to the measurement equation, V.

The KF first predicts the output of the SOC and the covariance error. Then, the Kalman gain and the state value are calculated, which corrects the covariance error and eliminates the influence of noise in the internal parameter estimation [49]. However, KF is the most suitable model for linear systems. The extended KF and the unscented KF were developed to estimate the SOC more accurately because many systems are nonlinear in nature.

2.2.2. Drive Motor Parameter of the HEV

The most used motor in HEVs is the permanent magnet synchronous motor (PMSM). The PMSM is a three-phase AC motor. Its electromagnetic torque is generated by the interaction between the stator current and the rotor magnetic field. The motor controls the power distribution of two power sources through the motor windings with the dual inverters, which can serve in the vehicle’s energy management [50]. The voltage, current, and torque can be calculated with Equation (6), Equation (7), and Equation (8), respectively [50]:

$$\left\{\begin{array}{l}{u}_d=R_s{i}_d+L_d{\displaystyle \frac{d{i}_d}{dt}}-{\omega }_s{L}_q{i}_q\\ u_q=R_s{i}_q+L_q{\displaystyle \frac{d{i}_q}{dt}}+{\omega }_s(L_d{i}_d+{\Psi }_f)\end{array}\right.$$

(6)

where u_d and u_q are the voltages of the d-axis and q-axis, respectively, V. i_d and i_q are the currents of the d-axis and q-axis, respectively, A. R_s is the stator resistance of the motor, Ω. L_d and L_q are the inductances of the d-axis and q-axis, respectively, H. ω_s is the electrical angular velocity of the motor, rad/s. ψ_f is the flux linkage of the permanent magnet, Wb.

$$\left\{\begin{array}{l}{i}_d=i_{vd}+(L_d{\displaystyle \frac{d{i}_{vd}}{dt}}-{\omega }_s{L}_q{i}_{vq})/R_c\\ i_q=i_{vq}+\left[L_q{\displaystyle \frac{d{i}_{vq}}{dt}}+{\omega }_s(L_d{i}_{vd}+{\psi }_f)\right]/R_c\end{array}\right.$$

(7)

where i_vd is the effective component of i_d, A. i_vq is the effective component of i_q, A. R_c is the equivalent core loss resistance of the motor, Ω.

$$T_e=p{i}_{vq}\left[{\Psi }_f+(L_d-L_q)i_{vd}\right]$$

(8)

where T_e is the electromagnetic torque, N·m. p is the number of pole pairs of the motor.

PMSMs have advantages of a high-power density, high torque density, compact structure, and high efficiency [51,52]. The models correct the variation of inductance and flux linkage by improving the parameters of the nonlinear function [53]. The advanced control strategies of the adaptive control [54] and robust control [55,56] are employed in terms of control algorithms to enable the motor control system to adapt to real-time changes in motor parameters, which can ensure that the motor always operates in an efficient and stable state.

Collectively, this comprehensive analysis of hybrid power system configurations and key components provides both a theoretical basis and practical guidance for the efficient design and optimization of HEV energy-management systems. It serves as a solid foundation for promoting the development and application of related technologies, ultimately contributing to more sustainable and efficient hybrid electric vehicle solutions.

3. Classification and Comparison of EMSs

In the actual operation of HEVs, the EMS is pivotal for maximizing the efficiency and minimizing the energy consumption. As shown in Figure 2, three primary EMS categories dominating this field are the rule-based strategies, optimized strategies, and learning-based strategies. Rooted in the fundamental theories of vehicle dynamics and energy conversion, the strategies integrate the multi-dimensional data, including real-time driving cycle information and the SOC, which can dynamically regulate and optimize the energy flow among power system components. This approach effectively curbs fuel usage, which can enhance the overall vehicle performance and promote efficient energy utilization. While the following sections review various EMSs and their reported performance, it is critical to contextualize these findings. The reliability of quantitative metrics is inherently tied to their validation platforms (SIL (Software-in-the-Loop), HIL (Hardware-in-the-Loop), processor-in-the-loop testing, testbench, or final vehicle), which are explicitly documented for each study in the subsequent analysis. This transparency aims to facilitate a nuanced understanding of each method’s applicability across different experimental contexts.

3.1. Rule-Based Strategy of EMSs

As simple and highly implementable EMSs, the rule-based strategies typically adopt the “if-then” structure, which is widely utilized in real-time energy management. This approach primarily comprises two subtypes of the deterministic rule-based and fuzzy logic rule-based strategies [57]. Deterministic rule-based strategies set thresholds based on the predefined conditions, such as the vehicle speed, torque, power demand, and battery state. The conditions determine the energy distribution and the operational modes across various driving conditions, which can improve the vehicle performance and optimize fuel efficiency. These strategies are further categorized into the thermostat, power follower, and state machine strategies. The thermostat (ON/OFF) strategy employs the generator and ICE to generate electricity, which can keep the battery SOC within preset limits [58]. The power follower strategy can dynamically adjust the power output of the ICE and generator according to the vehicle power demands, which can ensure the efficient system operation. The state machine strategy partitions the energy system into multiple distinct states, and each is linked to specific operating modes and control rules. For example, Li et al. leveraged this strategy to distribute the required power according to the state changes. As a result, the vehicle had safe operating conditions during idle periods to achieve better energy efficiency [59].

The traditional deterministic rule-based strategies face limitations in complex scenarios in practice. Takrouri et al. addressed this by introducing the improved power follower strategy with an exponential smoothing function with a single adjustable parameter [60]. Compared with the conventional power follower strategies, this strategy reduced the battery root-mean-square current by 19%, which can mitigate the fluctuations to extend the battery lifespan and enhance efficiency (HIL). Wang et al. extracted the control rules from the recognized optimal algorithms to optimize the control parameters offline and refine online, which achieved the near-optimal fuel economy and SOC balance [61]. FLC is the control algorithm based on the fuzzy set theory, the linguistic variables, and logic reasoning [57]. FLC can accurately transform the linguistic control built on the expert knowledge into the automated control as an advanced control strategy. The FLC operation process mainly comprises three stages. The first process is the fuzzification. The precise input values are converted into corresponding fuzzy sets via predefined membership functions. The second process is the rule application stage. The rules exacted from extensive expert knowledge are applied to determine the fuzzy output level. The final process is the defuzzification. The core task of this stage is to accurately convert the fuzzy output level determined in the previous stage into the precise value through the specific algorithm [62]. The FIS integrated the fuzzification interface, the rule base, the inference engine, and the defuzzification interface (Figure 3). FLC can efficiently and robustly handle the uncertainties and fuzziness in the system, which can significantly enhance the flexibility and efficiency of the control process.

Recent studies have advanced the FLC-based EMS applications. Yang et al. optimized strategy parameters with the adaptive fireworks algorithm and developed the rule-based shift strategy and the Takagi–Sugeno controller for torque distribution. Compared with the classical fuzzy torque distribution strategy, the approach improved HEV fuel economy by approximately 11% under the Chinese urban cycle with 14% in real-world driving (HIL) [40]. Deng et al. integrated FLC with the state machine, which leveraged the vehicle power demands, the SOC, and the battery power limit to control the fuel cell operations. Under the WLTC, hydrogen consumption was reduced by approximately 9.8% compared with the power follower control (PFC), and the SOC of the battery was effectively stabilized (SIL) [63]. Rezk et al. proposed EMS switching between fuzzy logic and proportional integral (PI) control strategies based on the thresholds, achieving approximately 3.4% and 9.1% performance boosts over PI and FLC alone, respectively (SIL) [64].

There are differences between the deterministic and fuzzy logic rule-based strategies, which both have the characteristics of high computational efficiency, simple control, and strong robustness [57]. The deterministic rules can quickly respond to changes in the vehicle state, making them suitable for the application scenarios with high real-time requirements. The fuzzy logic rules are more suitable for handling complex and nonlinear systems, which could make effective decisions under uncertain conditions and enhance the flexibility of control. The overall performance can be enhanced by combining the approaches. However, the traditional rule-based EMSs rely heavily on expert experience [65]. They cannot dynamically optimize energy distribution, which makes it difficult to achieve optimal performance [66]. Moreover, the fuzzy logic-based power distribution may additionally introduce the optimization parameters causing deviations from the global optimal solution [45]. Nowadays, research has introduced improvements, such as modified power follower strategies and integrations with fuzzy logic controllers, which have enhanced fuel economy and other performance metrics. Moving forward, efforts should focus on combining the strengths of rule-based EMSs with advanced optimization techniques to reduce dependence on expert knowledge and improve adaptability. This will drive the development of more efficient and intelligent rule-based EMSs for HEV technology and sustainable transportation.

3.2. Optimized Strategy of EMSs

Optimized EMSs aim to adjust the control variables based on the numerical calculation results by minimizing a predetermined cost function under feasible constraints [67]. The methods are mainly divided into two categories: global and the local optimization. Global-optimized EMSs apply the static data to optimize entire control sequences for specific operating conditions [68] with the DP, PMP, and GA. However, the reliance on prior global knowledge restricts the online implementation. Consequently, the local optimization often serves as a practical approximation to the global optima [68]. Local-optimized EMSs focus on the specifically regional or sub systemic searches, which solve the standalone subproblems or are integrated into the global frameworks. Common local methods include the Equivalent Consumption Minimization Strategy (ECMS) and MPC.

3.2.1. Global-Optimized EMSs

DP is grounded in Bellman’s optimality principle, which decomposes complex multi-stage problems into a series of simpler single-stage problems [69]. It identifies the global optimal solution by recursively solving subproblems. The EMSs optimized by the DP algorithm can significantly enhance fuel economy. The EMSs that combined DP and decision rules improved fuel economy by approximately 6% compared with the rule-based EMSs (SIL) [70]. Since DP can obtain the global optimal solution, it was often regarded as a benchmark for other EMSs [71,72], which was limited to offline use owing to the requirement for predefined speed profiles [6]. Moreover, the computational complexity of the DP algorithm increased sharply when it dealt with large-scale problems [73]. Additionally, the computation time of DP grew exponentially with the change in the problem dimension, which led to the low computational efficiency [74].

PMP transforms the global optimization problems into the instantaneous Hamiltonian optimization problems via the variational calculus of DP, which is an optimal control theory-based approach [58]. The EMS proposed by Sun et al. identified driving characteristics by collecting the output power of the motor, which used the suboptimal costate to implement the PMP algorithm in real-time during driving. It cut the hydrogen consumption by approximately 29% compared to rule-based EMSs, with an average online computation time of about 3.6 ms per step and sampling intervals of less than 1 s (SIL and final vehicle) [75]. The PMP strategies can outpace DP-based EMSs with a computation time reduction of up to 89%, which greatly reduced the computational complexity and made it suitable for real-time application scenarios (SILs) [76]. Quan et al. proposed the PMP-EMS to dynamically update costate variables according to the real-time driving conditions, which slashed the equivalent hydrogen consumption by approximately 14% and 9.2% compared with the traditional PFC strategies and static PMP-EMSs with fuel cell degradation rate reductions of 93% and 8.7% (SIL) [77]. The PMP combined with the real-time driving conditions can achieve optimal control with both instantaneous and global characteristics [78]. It should be noted that the PMP-based EMSs cannot guarantee the optimality of its solution in the absence of information about future driving conditions [79].

As a heuristic global optimization method, GA mimics the mechanisms of natural selection and genetics, which seeks the optimal solution through an iterative evolutionary process. The core process of GA involves initializing a population within the feasible solution space to generate the new populations with genetic operations, which includes the selection, the crossover, and the mutation to evaluate the termination criteria for population convergence [58]. Ma et al. used GA to adjust the weights and constraints of MPC strategies to achieve an optimal flexible power allocation process [80]. Wang et al. optimized the formulas of the fuzzy membership functions with GA to narrow the gap with a DP of 2.7%, which boosted the energy economy by 2.6%, 2.4%, and 3.3% at different temperatures (10 °C, 25 °C, 40 °C) (SIL) [65]. Nassar et al. used a variant of the GA algorithm, multi-objective GA, optimized EMS offline, and stored optimal maps to enable the EMS to provide a continuous improvement in fuel economy across multiple driving cycles [81]. Li et al. modified an Otto-cycle engine to an Atkinson-cycle engine, which was optimized with the non-dominated sorting GA-II. The cumulative fuel consumption was reduced by approximately 4.6% under the NEDC condition (Testbench and SIL) [82]. GA had a strong global search ability to effectively avoid getting trapped in local optimal solutions, which made it suitable for handling nonlinear, multi-objective, and high-dimensional optimization problems. However, GA had issues of high computational requirements [58], slow convergence speed, and sensitivity to parameter settings.

3.2.2. Local Optimized EMSs

ECMS is the local optimization method based on the PMP, which calculates the equivalent factor (EF) to equate the process of electrical energy consumption and fuel consumption. It achieves the optimal power distribution between the engine and motor in hybrid systems by optimizing the EF. The optimization effect of the ECMS depends on the accuracy of the EF. Hu et al. introduced a data-driven driving condition recognition model [83]. Sun et al. added the correction factor that can adjust the EF according to the motion characteristics of each driving segment [84]. Vignesh et al. combined ECMSs with intelligent technologies and proposed a robust EF correction method for the SOC under various uncertain driving conditions to realize the adaptive adjustment of the equivalent factor [85].

The innovative methods of the ECMS in reducing fuel consumption have been deeply explored. At present, ECMSs can effectively reduce fuel consumption to approximately 5.3 L per 100 km (SIL) [86]. Zhang et al. fused the advanced driver assistance system and maps, who proposed a data-driven predictive equivalent consumption minimization strategy (D-PECMS) with a multi-variable long short-term memory (LSTM) network. The method reduced the computational burden and slashed fuel consumption by approximately 13% versus rule-based control strategies (HIL) [87]. The energy management problem of HEV can be transformed into an optimal path search with the breadth-first search algorithm to achieve global optimization. Hao et al. took the breadth-first search results as the reference to propose an AECMS, which used the PSO algorithm to adjust the EF in real-time. Compared with the rule-based strategies, the fuel consumption can be reduced by approximately 10–15% and 8–10% under the federated test procedure (FTP75) and the worldwide harmonized light vehicles test procedure (WLTP) cycles (SIL) [88]. Gao et al. proposed the scenario-oriented AECMS that predicted the future speeds with a hybrid neural network model based on the Nanjing oriented driving cycle. It effectively restricted the SOC within a narrow fluctuation range of 0.12–0.33%. The equivalent hydrogen fuel consumption was reduced to approximately 7.1 g/km (SIL) [89]. Figure 4a,b are the framework diagrams of D-PECMS and AECMS, respectively. Despite the online viability and the low complexity [90,91], ECMS still has difficulty in maintaining the optimal performance of the charge in real-time applications [42] to balance the driving performance with computational efficiency [58].

The MPC targets efficient energy management and power distribution like the ECMS. MPC strategies forecast future driving with the system models and prediction algorithms, which leverage the forecasts to inform the power allocation and the control decisions. Xue et al. organically combined the MPC with ECMS by means of driving mode recognition technology, which retained the advantage of the MPC in obtaining optimization results through rolling optimization to update and predict control parameters in the next time domain [68]. The hybrid approach achieved fuel economy, adaptability, and near-global optimality with high computational efficiency [92]. Cui et al. proposed a multi-objective hierarchical EMS that merged resistance network-triggered DP for motion planning with MPC for the convex torque optimization. The framework demonstrated the excellent robustness against the random fluctuations of traffic flow by applying the alternating direction method of multipliers [93]. The strong robustness of the MPC made it suitable for the control of uncertain and nonlinear dynamic systems [94]. However, the performance highly depends on the accuracy of the prediction model. It may lead to unsatisfactory control effects if the prediction error is significant. The constant speed prediction method can mitigate the inaccuracies caused by the changes in traffic fluctuations or driving habits, which can significantly improve the prediction accuracy [68]. Nevertheless, the MPC has the high computational complexity and requires substantial computational resources.

In conclusion, optimized EMSs for HEVs have their respective merits and drawbacks. Global optimization methods are capable of achieving high-quality optimization. However, they face challenges such as high computational demands, reliance on prior knowledge, or a lack of optimality guarantees without future information. On the other hand, local optimization methods offer online viability and lower complexity to some extent. Nevertheless, these methods struggle with maintaining optimal performance or heavily depend on the prediction accuracy. These strategies have advanced HEV energy management, but continuous research is needed to address their limitations. Combining the strengths of global and local optimization approaches could lead to the development of more efficient, robust, and adaptable EMSs for the evolving needs of HEV technology.

3.3. Learning-Based Strategy

Learning-based strategies deeply integrate cutting-edge technologies, such as artificial intelligence and RL, which optimize energy management through iterative learning and experience accumulation driven by data. The intelligent approaches have shown remarkable effectiveness in controlling large nonlinear systems as given in Figure 5 [95]. As a transformative technology in planning and optimization, RL empowered agents to learn model-free and efficiently by interacting with their environment [58]. Rooted in the Markov decision process, RL conducted the learning and decision-making. The model of the Markov decision process involved the construction of state spaces, action spaces, and reward functions. This framework has significant advantages in solving sequential decision-making problems under uncertain conditions, which can ensure global optimality within the entire planning scope [96]. The main algorithms involved in RL are policy gradient and Q-learning. Policy gradient directly tuned policy parameters by estimating gradients and updating in the ascent direction to maximize cumulative rewards. Conversely, Q-learning was a model-free algorithm based on the value function. By iteratively updating the Q-value function with the Bellman equation, the agent learns in different states to obtain the maximum long-term cumulative reward to obtain the optimal policy. Figure 5a is the framework of the EMS based on Q-learning. Figure 5b–d are the framework of the EMSs based on DQN, DDPG, and SAC, respectively. The neural networks underpin energy system prediction, optimization, and control. Each neural network receives multiple input signals comprising the interconnected neurons, which are processed through weighted summation and nonlinear activation functions to generate an output signal. The networks capture complex input–output mappings, enabling energy system modeling and prediction by adjusting the connection weights between neurons.

Non-RL neural networks mainly include the LSTM, the gated recurrent units (GRUs), and the recurrent neural networks (RNNs). Figure 6a–c are the structures of the LSTM, GRU, and RNN [97,98]. DRL combines RL with the neural networks to effectively handle the complex scenes with high-dimensional sensory inputs [58]. Key DRL algorithms include DQN, DDPG, and SAC. DQN integrates deep learning with Q-learning, which uses the Q-network to approximate Q-values and improve the training stability via experience replay and target networks. It is used to solve RL problems in discrete action spaces, and its neural network structure is listed in Figure 6d [99]. Wang et al. applied the parameterized DQN to reduce the lithium-ion battery aging by approximately 11% while achieving approximately 99.5% of DP global optimality (HIL) [100]. DDPG combines the deep neural networks with the deterministic policies, which is suitable for continuous action spaces. In DDPG, the actor network outputs deterministic actions and the critic network evaluates the action values. The neural network structure is given in Figure 6e [101]. SAC is grounded in the maximum entropy RL to maximize the cumulative reward, which performs well in continuous space. Studies confirmed the DRL-based EMSs’ superior fuel economy.

Table 2. The performance of different learning-based strategies based on fuel consumption.

Strategy	Fuel Consumption
Radial basis function neural network + DQL + improved SQP + dynamic process coordination control algorithm (SIL) [71]	↓approximately 8.6% compared with the rule-based strategy
Prioritized experience replay (PER) + TD3 + deep transfer RL (SIL) [72]	↓approximately 6.1% in the composite cycle and approximately 6% in China typical city bus driving cycle compared with EMSs
Knowledge assistance + DDPG + DQN (SIL) [102]	↓approximately 7–7.3% (offline mode) and approximately 5.2–5.7% (online mode) compared with the rule-based strategy
Multi-objective optimization of reward function + SAC (SIL) [103]	approximately 4.2 L per 100 km under urban–suburban-highway driving conditions
SAC (SIL) [104]	↓approximately 8.3% and 3.2% compared with the rule-based strategy and ECMS
Multi-agent RL (SIL) [105]	↓approximately 16% and 13% compared with the hierarchical MPC and single-agent DRL strategy
Driving cycle information + double DQN + BPNN (HIL) [106]	↓approximately 17% and 12% compared with the strategy adopting the charge depleting/charge sustaining (CD/CS) rule and proportional AECMS
Self-supervised learning model + DQL (SIL) [107]	↓approximately 6.1% and 4.8% compared with the DQN and DDPG algorithms
SAC + PER + emphasizing recent experience (ERE) + Muchausen RL (SIL) [108]	↓approximately 4.6% and 2.5% compared with the EMSs based on DDPG and TD3

Note: ↓ indicates that the data in this column shows a downward trend.

,

Table 3. The performance of different learning-based strategies based on hydrogen fuel consumption.

Strategy	Hydrogen Fuel Consumption
PMP + fuzzy C-means clustering + RBFNN + Hamiltonian function (SIL) [78]	↓approximately 8.9% compared with the rule-based EMSs, ↓approximately 3.5% compared with the optimal hydrogen consumption of the offline PMP
Unsupervised clustering strategy + DQN (SIL) [109]	↓approximately 48% compared with the rule-based EMSs
BPNN + predicted ECMS + DQN (HIL) [110]	↓approximately 56% (UDDS) and 27% (extra urban driving cycle) compared with the CD/CS strategy, ↓approximately 5.3% (UDDS) compared with the SQP strategy
Double DDPG + previous action guidance mechanism (SIL) [111]	↓approximately 30% (SCHDC), 26% (UDDS) and 28% (NEDC) compared with the rule-based EMSs

Note: ↓ indicates that the data in this column shows a downward trend.

and

Table 4. The performance of different learning-based strategies based on fuel economy.

Strategy	Fuel Economy
Multi-objective reward function + driving condition recognition + DDPG (SIL) [41]	average improvement of 11% compared without adaptive driving condition recognition
Adaptive online RL + hybrid control unit (SIL) [99]	an improvement of approximately 10.5% compared with the rule-based strategy
Adaptive fuzzy filter + adaptive greedy noise + heuristic experience replay method + DDPG (SIL simulation calibrated with testbench data) [101]	average improvement of 14% compared with EMSs based on the DQN
SAC + PER + ERE + Muchausen RL (SIL) [108]	↑approximately 4.7%, 3.4% and 3.2% compared with the EMSs based on DP, DDPG and TD3
DDQL + modified PER + adaptive optimization method named AMS Grad (SIL simulation calibrated with testbench data) [112]	↑approximately 1.1% and 2.5% compared with the original DDQL and DQL
Adaptive fuzzy filte +TD3 + LSTM + DP (SIL) [113]	↑approximately 33% compared with the traditional EMSs based on TD3

Note: ↑ indicates that the data in this column shows an upward trend.

show the fuel consumption, hydrogen consumption, and fuel economy across diverse learning-based strategies.

The solutions have been actively explored to address the prolonged training time of DRL. Li et al. designed novel reward terms to prompt agents to autonomously identify the optimal SOC range by randomly combining diverse load curves for DQL model training. It achieved a 96.5% reduction in the training time with a 55.4% decrease in the computation time compared with the Q-learning-based strategies (SIL) [114]. Li et al. also proposed the EMS based on DDPG and designed a new reward function that considered the battery aging model to conduct the training within a vehicle-to-cloud framework. This approach reduced the number of training rounds by 98.5% and 20.0% relative to Q-learning and DQL-based strategies, respectively (processor-in-the-loop testing) [115]. Other researchers have targeted the convergence speed.

Table 5. The performance of different learning-based strategies on convergence rate.

Strategy	Convergence rate
SAC + DP + Matlab parallel computing toolbox (SIL) [14]	↑approximately 206% compared with the EMSs based on DDPG
SAC + Muchausen RL + PER + DP (SIL) [17]	↑approximately 31% compared with SAC
PER + TD3 (SIL) [72]	↑approximately 46% compared with TD3
SAC + PER + ERE + Muchausen RL (SIL) [108]	↑approximately 94%, 88% and 91% compared with the EMSs based on DDPG, TD3 and other DRL
DDPG + regenerative braking control strategy (SIL) [116]	12 episodes to reach the convergence state (WLTC)
SAC + power limit (SIL) [117]	↑approximately 50%, 9.1% and 88% compared with the algorithm based on DDPG, proximal policy optimization and SAC

Note: ↑ indicates that the data in this column shows an upward trend.

compares the convergence performance of various learning-based strategies. While the results demonstrated in Table 2, Table 3, Table 4 and Table 5 show performance improvements, it is essential to emphasize that the reliability of these metrics is closely linked to their respective validation platforms. All such platforms are explicitly denoted in the corresponding studies to ensure transparency. Nevertheless, hybrid algorithms often demand extensive data support [58]. Moreover, the complexity of RL-based EMSs has been continuously increasing as RL algorithms evolve and deep neural networks become more complex, which affect the computational workload and time consumption during the strategy application phase [96].

In conclusion, learning-based strategies have revolutionized HEV energy management by leveraging artificial intelligence and RL. These strategies excel in handling complex nonlinear systems and sequential decision-making under uncertainty, significantly enhancing fuel economy and system performance. While innovative solutions have been developed to mitigate long training times and slow convergence, challenges remain. The increasing complexity of hybrid algorithms and their high data requirements pose obstacles to practical implementation. Continued research should focus on simplifying algorithms, reducing data dependence, and improving the computational efficiency to fully realize the potential of learning-based strategies in HEV energy management.

Table 6. Advantages and limitations of different types of EMSs.

The Type of EMSs	Advantages	Limitations
Ruled-based EMSs	(1) High computational efficiency, simple control, and strong robustness (2) Good rapid response and high real-time performance (3) Suitable for handling complex and nonlinear systems	(1) Relies on expert experience, unable to dynamically optimize energy allocation, and lacks optimality (2) Additional optimization parameters will be introduced with fuzzy logic for power allocation
Optimized EMSs	(1) DP can obtain the global optimal solution (2) PMP can reduce the complex calculation and achieve optimization control that combines instantaneity and globality (3) GA can avoid the problem of local optimal solutions (4) ECMS can be implemented online with low complexity and computational cost (5) MPC has strong robustness	(1) Hard to implement global optimization online (2) DP deals with large-scale problems with a sharp rise in the computation volume and exponential growth in the computation time with dimensions (3) PMP cannot ensure solution optimality without future driving information (4) GA has high computational demands and slow convergence and is sensitive to parameter settings (5) ECMS’s optimization depends on EF accuracy, struggles to maintain optimal charge, and faces driving performance and computational challenges (6) MPC’s performance highly relies on prediction model accuracy with high computational complexity
Learning-based EMSs	(1) RL can tackle the control challenges of large-scale nonlinear systems and achieve model-free and efficient learning (2) DRL can handle complex scenarios with high-dimensional sensory inputs and significantly improve fuel consumption and fuel economy	(1) Excessively long training time (2) Hybrid algorithms require a large amount of data for support (3) The increased complexity of the algorithms affects the computational volume and time consumption (4) Decision-making is uncertain in complex environments

summarizes the advantages and limitations of rule-based, optimized, and learning-based EMSs for HEVs, revealing the characteristics of each type. Rule-based EMSs are known for their computational efficiency and robustness but lack dynamic optimization capabilities. In contrast, optimized EMS methods, such as DP, can achieve global optimality but require significant computational resources, while PMP strikes a balance between optimization and complexity. Learning-based EMS approaches, such as RL and DRL, offer the potential for model-free learning and improved fuel economy. However, they face challenges related to long training times and high algorithmic complexity. This detailed comparative study provides guidance for EMS strategy selection and improvement. The EMS analysis underpins the development of more efficient, intelligent EMSs, driving innovation in HEVs and related fields to meet the evolving energy and performance demands [118].

4. Future Development Trends

Traditional EMSs increasingly struggle to meet the requirements of efficient, intelligent, and eco-friendly development. In the future, EMSs will make significant strides towards the in-depth integration of intelligence and real-time optimization, as well as multi-objective optimization. Through multi-dimensional in-depth integration and continuous optimization, EMSs will drive the efficient and sustainable operation of energy systems.

Future EMSs will closely integrate learning-based algorithms with optimized models to achieve dynamic adaptive control. The learning-based algorithms extract complex patterns from vast datasets, while the optimized models provide precise decision-making frameworks, which can interplay to propel EMSs’ intelligence to new heights. Ma et al. combined MPC with TD3 and used an improved SQP algorithm to enhance the real-time performance, which can minimize battery degradation and boost efficiency [119]. Edge computing will further amplify system responsiveness by reducing data latency to underpin the intelligent real-time integration. Zhang et al. developed a mobile edge computing framework, which decentralized the traditional on-board control units into cloud-integrated hierarchical asynchronous controllers [120]. This enabled load balance across the local nodes and near-zero communication delays.

Multifaceted objectives like fuel economy, emissions control, and drivability will define future EMS design, presenting a significant challenge in balancing these goals within dynamic real-world scenarios. Li et al. proposed a multi-objective EMS, which considered fuel economy to maintain the SOC and reduce battery degradation [121]. It used a multi-agent DDPG algorithm and treated the engine and the battery as cooperative agents to optimize power distribution and outperform traditional methods in multi-objective tasks. Cheng et al. proposed an adaptive EMS that integrated the rule-based control with multi-objective optimization [122], which incorporated a point-line strategy optimized by PSO. This approach reduced operational costs, extended the fuel cell lifespan, and enhanced the operating efficiency.

Among these research efforts, the multi-objective evolutionary algorithm based on decomposition (MOEAD) stands out as a highly innovative and promising approach. The MOEAD decomposes the multi-objective optimization problem into multiple subproblems, which can search for Pareto optimal solutions in a large solution space to achieve global optimization. This unique feature endows MOEAD with the ability to handle complex multi-objective problems more effectively compared to many traditional methods. Cui et al. applied MOEAD within a cyber–physical interaction framework, achieving remarkable results. Not only did it increase the computational speed while maintaining energy efficiency, but it also outperformed synchronous and hierarchical methods by a significant margin [123]. In future research, in-depth exploration of MOEAD is essential. First, its parameter settings should be optimized to adapt to different vehicle models and driving conditions. Then, it should be combined with emerging control strategies to enhance the efficiency of EMSs. Additionally, analyzing the potential of MOEAD in dealing with new multi-objective problems that may arise in the development of HEVs will also be an important research topic.

The convergence of intelligence and real-time optimization of the EMSs along with the widespread application of multi-objective strategies will herald a paradigm shift in energy management. Through continuous innovation, future EMSs will be pivotal in driving efficient energy use, environmental protection, and sustainable growth to steer the society toward a greener, smarter future.

5. Conclusions

The principles, advantages, and disadvantages of the rule-based, optimized, and learning-based EMSs and energy configuration methods (series, parallel, and series–parallel) for HEVs were introduced. The EMSs have played a significant role in fuel consumption reduction to improve fuel economy and vehicle performance in practical effectiveness. Compared with the rule-based strategies, the learning-based strategies can reduce fuel consumption by approximately 5.2%~17% with hydrogen consumption reductions by up to 56% and fuel economy improvements by up to 10.5%.

With the continuous development of intelligent and information technologies, the in-depth integration of intelligent and real-time optimization and multi-objective optimization strategies have become a highly promising development direction. The results of this study not only conformed this development trend but also further enriched it. This study aligned with global efforts to reduce carbon emissions in the automotive sector. By systematically analyzing EMSs for HEVs, the findings established a theoretical framework for optimizing system efficiency. Significant fuel savings and emission reductions can be achieved. These insights can directly inform the development of next-generation energy-management systems. The widespread adoption of such technologies will advance sustainable mobility and support decarbonization targets. Future research should focus on developing algorithms with high adaptability and robustness, emphasizing cost scalability to overcome barriers in real-world implementation.