WO2008128416A1

WO2008128416A1 - Energy management for hybrid electric vehicles

Info

Publication number: WO2008128416A1
Application number: PCT/CN2008/000505
Authority: WO
Inventors: Yangsheng Xu; Guoqing Xu; Zhancheng Wang; Hong Tong; Chenggang Xia; Zhiru Liu; Weimin Li; Huihuan Qian; Tiande Mo; Tin Lun Lam; Bufu Huang; Weizhong Ye; Ka Keung Caramon Lee; Xinyu Wu; Wing Kwong Chung; Zhi Zhong
Original assignee: The Chinese University Of Hong Kong
Priority date: 2007-04-19
Filing date: 2008-03-14
Publication date: 2008-10-30

Abstract

Disclosed are an energy management system and a method for hybrid electric vehicle. The system comprises: a load forecast system and a vehicle management system. The load forecast system is configured to receive input information and carry out a learning on the input information to obtain a future load level. The vehicle management system is configured to receive the future load level from the load forecasting system, determine an optimal future power output based on the future load level, a state of charge of a battery and a vehicle speed, and coordinate operation of an engine, a generator and the battery of the hybrid electric vehicle based on the optimal future output power. The energy management strategy of the present invention realizes an optimal energy management for a hybrid electric vehicle, only using information provided by the vehicle itself. Moreover, the present invention balances the goals of fuel economy while always meeting demand power and maintaining the functionality of the battery system.

Description

ENERGY MANAGEMENT FOR HYBRID ELECTRIC VEHICLES

TECHNICAL FIELD OF THE INVENTION

[0001 ] The present invention relates to the field of energy management of vehicles, and more particularly to systems and methods of energy management of Hybrid Electric Vehicles (HEVs) in view of future load levels.

BACKGROUND OF THE INVENTION

[0002] At present, three kinds of power strategies for a hybrid electric vehicle are provided to satisfy power requirements with the cooperation of an electric power source and a fuel power source in order to solve the fundamental problem regarding the optimal split of the power of the engine/generator system and the battery of a HEV at a particular demand power.

[ 0003 ] The first approach is rule-based algorithm, such as the known "energy follow" and "thermostatic control", hi the energy follow strategy, the engine is always in a state of "ON" but the energy output of the engine is changeable based on energy requirements. The thermostatic control strategy obtains the power from the generator and the engine. The engine is turned on or turned off based on the State of Charge (SOC) of the battery. This control strategy works fast and reliably, but it is very hard to realize an optimal control.

[0004] The second approach is based on a static optimization method. The optimization solutions figure out the proper split between the electric power and fuel power by using steady-state efficiency maps, and realize HEV optimization based on the operation with optimal parameters. However, how to determine the efficiency maps and optimal parameters are the challenges of this kind of approach.

[0005 ] The third approach is a real-time optimization. Several algorithms have been proposed, including fuzzy logic controller, energy-flow analysis, Dynamic Programming, etc. The control strategy with the real-time optimization calculates an optimal torque based on feature parameters of a vehicle, and decides an actual torque output by modifying the optimal torque based on the real-time road situation and SOC. Dynamic optimization parameters can be changed based on the real-time operation state and energy requirements, then a real-time optimal solution can be concluded for power controlling. This approach is more accurate under transient conditions than the first two methods, but the inherent disadvantage of heavy computational cost limits its application.

[0006] As stated above, the existing approaches have some limitations that they use only the current vehicle state for making decision in the power distribution. Little consideration is given to the future load. As a result, the existing approaches fail to consider the effects of variations on emissions and fuel-consumption over the future load to which the vehicle may be subjected.

SUMMARY OF THE INVENTION

[0007] To solve the aforesaid limitations, the present invention provides a system and method for management energy which integrates an on-board load forecasting system, thereby realizing an optimal energy management for a hybrid electric vehicle (HEV).

[ 00083 An aspect of the present invention is to provide an energy management system for a HEV , comprising: a load forecast system configured to carry out a learning on a speed sequence, an accelerator pedal sequence and a brake pedal sequence received from the HEV to obtain a future load level associated with the HEV; and a vehicle management system configured to receive the future load level from the load forecasting system, information associated with a state of charge (SOC) from a battery of the HEV, and a vehicle speed from the HEV; wherein the vehicle management system is further configured to determine an optimal future power output based on the received future load level, the information associated with the SOC of the battery and the vehicle speed, so as to coordinate operation of an engine, a generator and the battery of the HEV.

[ 0009 ] Another aspect of the present invention is to provide a method for managing energy of a HEV, comprising: retrieving a speed sequence, an accelerator pedal sequence and a brake pedal sequence from the HEV ; carrying out a learning on the retrieved sequences to obtain a future load level associated with the HEV; retrieving information associated with a state of charge (SOC) of a battery of the HEV, and a vehicle speed of the HEV; and determining an optimal future power output based on the obtained future load level, the retrieved information associated with the SOC and the vehicle speed, so as to coordinate operation of an engine, a generator and the battery of the HEV.

[0010] In a preferred embodiment of the invention, the input information comprises: a speed sequence, an accelerator pedal sequence and a brake pedal sequence.

[0011 ] The method and the system of the present invention can realize an optimal energy management for a HEV, only using information provided by the vehicle itself. The method and the system provided by the present invention can balance the goals of fuel economy while meeting demand power and maintaining the functionality of the battery system.

BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and other advantages of the present invention will become more apparent by describing the preferred embodiments in detail with reference to the accompanying drawings, in which:

[0013] Fig. 1 schematically illustrates an energy management system which integrates a load forecasting system with a vehicle management system (VMS) of a HEV according to the present invention;

[0014] Fig. 2a illustrates a membership function for input variables; [ 0015 ] Fig. 2b illustrates a membership function for output variables; [0016] Fig. 3 is a table showing fuzzy rules between the input variables and a future driving load level;

[ 0017 ] Fig. 4 illustrates a process for obtaining the training samples; and [0018] Fig. 5 shows a flowchart of an example method of energy management for a HEV.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] As shown in Fig.l, an energy management system 30 includes a load forecasting system 1 adapted to estimate a future load level 22 and output the estimated level to a VMS 2. The VMS 2 is configured to communicate with an energy control unit (ECU) 3, a generator control unit (GCU) 5, and a battery control unit (BCU) 7, which control the operation of an engine 4, a generator 6 and a battery 8, respectively.

[0020] According to the present invention, the load forecasting system 1 is configured to receive, but not limited to, a speed sequence, an accelerator pedal sequence and a brake pedal sequence. The speed sequence records a serial of speeds the vehicle experienced in the past time period. The accelerator pedal sequence and the brake pedal sequence record information of an accelerator pedal (not shown) and a brake pedal (not shown), respectively.

[0021 ] Using those inputs, the load forecasting system 1 is able to estimate a future load level 22 by adopting various known fuzzy logics and neural networks and outputs the estimated future load level 22. The estimation of a future load level 22 by adopting known fuzzy logics and neural networks according to the invention will be explained in detail later.

[0022] The VMS 2 is connected to the ECU 3, the GCU 5 and the BCU 7 via a CAN bus. The ECU 3, GCU 5 and BCU 7 are connected to the engine 4, the generator 6 and the battery 8 of the HEV, respectively, so that the VMS can coordinate the operations of the engine 4, the generator 6 and the battery 8 through the ECU 3, the GCU 5 and the BCU 7. For example, the VMS needs to control the amount of fuel supply to the engine 4 by controlling the throttle position in the engine 4. The VMS needs to control the voltage and power output of the generator 6. The VMS also needs to control the charging/discharging of the battery 8 and read the SOC of the battery 8 as well. To control the engine 4, the VMS sends the expected information to ECU 3 via CAN bus and then the ECU 3 adjusts amount of fuel supply to the engine 4 and the throttle position in the engine 4 to the expected value. To control the voltage and power output of the generator 6, the VMS sends the expected value to GCU 5 and the GCU 5 will control the output voltage and power of the generator 6. When any error occurs, all units 3, 5 and 7 will share the error information (SOS commands) and take a corresponding action to shut down the corresponding device. The generator 6 is mechanically connected to the engine 4 and electrically connected to the battery 8. When the vehicle uses fuel energy, the engine 4 works to make the generator 6 drive the motor (not shown) of the HEV. When the vehicle uses electric energy, the battery 8 drives the motor directly. When the battery 8 needs to be charged, the generator 6 transforms electric energy into chemical energy.

[0023] The VMS 2 receives a future load level 22 from the output of the load forecasting system 1, a vehicle current speed 20 from a speed sensor (not shown in Fig. 1), and the SOC 24 from the BCU 7. The VMS 2 can determine an optimal future output power of the engine/generator mechanism and the output voltage and power of the battery in a future time slot on the basis of the mentioned inputs: the SOC 24, the future load level 22 and the current vehicle speed 20 by using a learning machine, thereby coordinating the generator, the engine and the battery.

[ 0024 ] It is understood by those skilled in the art, the ECU 3, the GCU 5 and the BCU 7 can be physically combined in any combination or can stand as separate units. They are described as separated units here for the ease of description.

[0025] Based on the aforesaid system 30, the present invention provides a method of energy management for a HEV. An embodiment is given to exemplify the estimation of the load forecasting level in a future time slot.

[ 0026 ] Firstly, the present invention proposes to use fuzzy logic to obtain a load level. It is known that the driving load is proportional to the vehicle net wheel torque, but the relationship between the driving load and the acceleration/speed of a vehicle is complex and nonlinear. As a matter of fact, there are infinite kinds of future load sequences, and forecasting the entire all kinds of future load sequences will cost enormous computation. Considering the trade-off between the computation cost and the HEV power control strategy requirements, it is suggested to preset five levels of driving loads: very low (VL), low (L), medium (M), high (H) and very high (VH). It is recommended to use a fuzzy logic approach to determine an actual driving load level.

[0027] Fig.2a and Fig. 2b illustrate membership functions of the fuzzy sets for input variables and output variables, respectively. It is learnt that an average acceleration and an average speed are chosen as input parameters of the fuzzy logic approach and both of them are divided into five levels. The average speed is obtained from the speed sequence, and the average acceleration is obtained from the combined information of the accelerator pedal sequence and the brake pedal sequence provided by the accelerator pedal (not shown) and the brake pedal (not shown), respectively.

[0028] Fig. 3 is a table to represent fuzzy rules in a linguistic form. The rules are formulated based on the experienced gained. To reduce the computation burden of defuzzification, it is recommended to select a centroid method as the defuzzifier. Namely, after gaining a single output value, the output should be rounded to one of the five load levels.

[ 0029 ] For example, if the membership of the ith speed is μs_it and the membership of the corresponding acceleration is μA_t, the firing strength μ, of the premise of is calculated based on min operator as the following equation (1). μi=min(μsi, μAj) (equation 1)

[0030] After gaining all firing strength μ, for all 25 rules of the table in Fig. 3, the corresponding driving load level will be obtained by a weighted average according to the equation (2),

25 / 25

L = round (∑ //,.«,. / ∑ μ_t , ) (equation 2)

(=1 / 1=1 wherein 04 is a value of consequent parts on ith rule.

[ 0031 ] As mentioned above, the purpose of the invention is to forecast a future load level so that the value of the load level can be used in the HEV power management stratery to reduce the fuel consumption and emissions. According to an embodiment of the present invention, the load forecasting system 1 proceeds to forecast a future load level on the basis of a Cascade Neural Network with Node-Decoupled Extended Kalman Filtering (CNN-NDEKF) due to its effectiveness in the classification of complex and nonlinearly separable target classes. This neural network was known in the paper entitled "Human control stratery: abstraction, verification and replication" (authors: M. Nechyba and Y. S. Xu , published in IEEE Control System Magazine, October 1997), all contents of which are referred to by this application. According to the paper, a CNN-NDEKF classifies its input vector into one of 5 target load levels through a two stage process.

[0032] To train the CNN network for a load forecasting classification problem, the statistics of three typical drive cycles, i.e. US06 (average speed: 77.2mh/h), NEDC (average speed: 33.3mh/h) and Manhattan (average speed: 10.98mh/h) will be calculated to generate the training database according to the invention. At this time, the load forecasting system 1 acts as a classifier. These three cycles represent a wide variety of conditions because no two cycles seem to provide similar features. They are significantly different in acceleration rates and average speeds. Fig. 4 shows a corresponding flow diagram to prepare training data. Here, the procedure for one driving cycle is described as below for the purpose of exemplification.

[0033 ] Step 1: initializing the segment lengths of Tp and Tf where Tp represents a past time slot which is used to forecast a future load level, while Ty represents an objective future time slot.

[0034] Step 2: randomly choosing a start point wherein the segment length from start-up of the vehicle to this start point is denoted as Ts, and the sum of Ts, Tp and Tf should be equal to or less than a preset driving cycle time length and Zs is shorter than Tp.

[0035 ] Step 3: applying fast fourier transform (FFT) to the signal associated with the speeds and accelerations during the past time slot Tp to select characteristic features and generate a CNN-NDEKF input vector P.

[0036] Step 4: calculating the average speed and acceleration for the future time slot Tf and applying the fuzzy logic approach to the signal associated with the speed and acceleration during the past time slot Tf to calculate the load level as the neural network output^, so as to generate a training sample {P, y).

[0037] Step 5: returning to step 2 until there are 100 training samples. [0038] In this procedure, the signal associated with the speeds and accelerations during the past time slot Tp cannot be used to train directly, because the dimension of the past time slot Tp is usually so large that it requires numerous instances to determine the result. Herein, FFT is used on said signal to select characteristic features and reduce the dimension. The first n coefficients of the FFT transform are chosen to form the input vector P.

[0039] In this procedure, the selection of Tp and Tf is very important because these two parameters will significantly influence the performance of training. Generally speaking, a larger Tp means more past information will be considered so that the corresponding result will be better while the computation cost will be much more. Similarly, a smaller 7/^"implies that the classifier will work more frequently but cost more computation, hi addition, the past driving cycle cannot be used to train directly because the dimension of the cycle is so large that it requires numerous instances to determine the result. To address this issue, the invention uses the fast fourier transform (FFT) to select characteristic features and reduce the dimension. The first n coefficients of FFT are chosen to form a new input vector. Further, in CNN-NDEKF, MaxHidden is a very important parameter to the classification result. A larger MaxHidden might correspond to higher testing accuracy as well as more times of iteration, and however, over fitting will occur. In this embodiment, MaxHidden is set as 5.

[0040] Since three driving cycles are considered in the embodiment and 100 training samples are obtained for each of the three driving cycles. Total 300 training samples will be obtained. It is understood that the selected three cycles and obtained training database are of illustrative purpose, and they should not render any limitation to this invention.

[0041 ] After training samples/data are available, they will be used to train the neural network until the training error converge to zero or the weight matrix Wl and W2 converge. Up to now, a stable neural network is established to be used for the load forecasting system 1.

[0042] When an online application is realized, a serial of real time input variables, such as a speed sequence, an accelerator pedal sequence and a brake pedal sequence, are input into the established neural network to calculate the future load level 22. [ 0043 ] The above illustrates the generation of the future load level by using fuzzy logic and neural network. Hereafter, the VMS will use the future load level 22 to realize the coordination of the operations of the engine, the generator and the battery.

[0044] VMS 2 is configured to receive the future load level 22 from the forecasting load system 1 , the vehicle current speed 20 and the SOC 24 to accordingly determine the power output from engine/generator system and from the battery for a future time slot so as to satisfy power requirement of the electric power source and the fuel power source and coordinate operations of the engine/generator system and the battery. Synchronously, the corresponding fuel consumption and emissions should be as low as possible.

[ 0045 ] According to the present invention, the VMS 2 operates like a learning machine. The learning machine is built up off-line. It records various data previously, like the load level, SOC, current speed, as well as the corresponding power output from engine/generator system, the power output from the battery, and the vehicle fuel consumption of a future time slot. By comparing to those previous data, the VMS 2 can determine an optimal output power based on the vehicle current state like current speed, the current SOC, and the future load level in a future time slot. The VMS 2 is also able to update the optimal output power. For about 10% determination, the VMS 2 can randomly choose a point near to the optimal power output as the future output power of the engine so that the vehicle can be driven by the power. Meanwhile, the VMS records the corresponding fuel consumption. If the recorded fuel consumption is less than the old one, the VMS will update the old record by this new one. In this way, the VMS is able to adapt to the changing situation.

[ 0046 ] There are many types of learning machine that can be used for classification problem, such as Support Vector Machines (SVM), Radial Basis Function Networks, Nearest Neighbor Algorithm, Fisher Linear Discriminant and so on. In this embodiment, SVM is employed as an example of the learning machine due to its stronger theory interpretation and better generalization than the others. Moreover, compared with the other learning machines, SVM has three distinct characteristics. First, SVM estimates the classification by using a set of linear functions that are defined in a high-dimensional feature space. Second, SVM carries out the estimation according to a principle of minimization of risk. Third, SVM implements the principle which minimizes the risk function consisting of the empirical error and a regularized term. The learning machine is capable of keep in updating. The VMS will randomly choose a point near the optimal power output as the future power output. After the vehicle operates on this point, VMS records the corresponding fuel consumption. If this fuel consumption is less than the used one, VMS updates the record by a new one. By this updating ability, the VMS can adapt to the changing situation.

[0047] According to the present invention, an HEV may operate in three different modes which are defined as electric power only, fuel power only and power-assist (electric power plus fuel power). Then, the optimization of the control of an HEV can be simplified to a classification problem which could be addressed as a ternary pattern recognition using SVM as following:

• setting up operation database of different load sequences, initial SOCs and vehicle speeds with a HEV;

• applying Fast Fourier Transform (FFT) on load sequences to select certain number of characteristic features from the data resulted from the FFT, and generating new database based on the characteristic features, the initial SOCs and the vehicle speeds; and

• training with SVM using the new database to build a model and producing SVM classifier using the model.

[ 0048 ] In this procedure, the load sequences, initial SOCs and vehicle speeds set up in the operation database cannot be used to train directly, because the dimension of the database may be so large that it requires numerous instances to determine the result. Herein, FFT is used on said signal to select characteristic features and reduce the dimension. The first n coefficients of the FFT transform are chosen to form the new database together with the SOCs and vehicle speeds.

[0049] hi an illustrative embodiment according to the present application, to decide which operation mode to be used in a future time slot Tf, the initial battery SOC (SOCj_ni_t) ,the initial vehicle speed (S_ve) and the forecasting load sequence ( L ) are chosen as the input vector of the classifier. SOC_1nJt and S_ve can take on continuous values. For example, assuming that the initial SOCs are between 0.4 and 0.8 as well as the initial vehicle speeds are between Omps and 70mps, to solve the classification problem, SOC_init can be discretized as SOC_mit ≡ {0.40, 0.42, 0.44, ...0.76, 0.78, 0.80}and S_ve can be discretized as S_ve ^e {0, 10, 20, 30, 40, 50, 60, 70}. There are infinite kinds of future load sequences (L= (L₁, L₂, L₃, ..., Lt_f)). To guarantee the performance of the classifier, as many sequences as possible should be covered when preparing training data. In an illustrative embodiment according to the present application, a Hidden Markov Model (HMM) is used to randomly generate a future load sequence. The HMM estimates future load sequence based on the transition probabilities of past driving cycles and standard cycles. It can guarantee the forecasting load sequence to be similar with a real load sequence.

[ 0050 ] The procedure of preparing training data is described as below:

1) initializing SOC,_nit =0.4 and S_ve =0;

2) generating a load sequence L randomly based on HMM;

3) forming a state vector x={SOC_mit, S_ve, ^ }; running the HEV model to follow the demand power sequence with SOC_in,_t and S_ve based on electric power only mode (i.e., the vehicle being driven by the electric power with the engine is off ,usually during low speed operation), engine power only mode (i.e., the vehicle being driven by the engine with the battery is off ,usually during high speed operation) and power-assistant mode (i.e., the engine and the battery cooperating when the demand power is very big), respectively; recording the corresponding fuel consumption FC, final SOC SOCfi_nai ^and emissions HC, CO, NOx; and generating a new sample as T = { x , FC_ei_e, SOCfi_naι-eie, HC_ei_e, CO_e]_e, NOx_ei_e, FC_eng, oUCfjnai-eng, WC_eng5 CU_eng, JNUX_eng>

WCass* CU₃Ss₅ J^" UXass } >

4) Sv_e = S_ve + 10; if S_ve «Ξ 70, go to 3);

5) SOCnn = SOC_1n,, + 0.02; if SOC_ιnit ≤0.8, go to 3);

6) going to 2) until there are enough random power sequences; and

7) for every demand power sequence, finding out the trade-offs between fuel economy, emissions and SOC based on the method described.

[ 0051 ] Fig. 5 shows a flowchart of an embodiment of an energy management process for a HEV according to the present invention. According to the embodiment, the SOC is fixed to be from 40% and 80% so as to ensure that the battery always has an ability to collect regenerated energy while never too low to start the vehicle.

[0052] As shown in Fig. 5, when a vehicle is started, the load forecasting system 1 can calculate a future load level based on the vehicle's speed sequence, the accelerator pedal sequence and the brake pedal sequence (steps 101 and 102) and outputs the calculated future load level to the VMS 2.

[0053 ] VMS 2 receives the future load level from the system 1 and receives the vehicles' current speed data from a speed sensor and the current SOC data from the BCU 7. Then, the VMS 2 determines whether the current SOC satisfies the fixed value range of 0.4-0.8 in steps 103 and 104. If the SOC is equal to or less than 0.4, the VMS 2 commands the engine to charge the battery (step 107). If the SOC is equal to or larger than 0.8, the VMS 2 will turn off the engine and have the vehicle powered by the battery only (step 108).

[0054] If the current SOC of the battery is in the range of 0.4-0.8, the VMS 2 will determine an optimal operational point which actually reflects the optimal power output of the engine/generator system in a future time slot based on the received future load level, vehicle current speed and the current SOC in step 105 and sends out corresponding commands to ECU, GCU and BCU to coordinate the operations of the engine, the generator and the battery in step 106.

[0055] For example, the VMS 2 may sample the voltage and the charging current of the battery 8, and change the output voltage of generator 6. Thus, the output current of the generator 6 will follow the change of the voltage. It means that it is possible to control the output voltage and the corresponding output current of the generator 6 to desired values so as to make the generator work at an optimal output power.

[0056] Further, it is assumed that 90% of the output power of the engine can be converted to electric power by the generator. By the characteristic of the engine, the optimal throttle position can be determined so that the engine can work most efficiently to output the desired power. This throttle position will be kept by controlling the electric throttle motor of the engine. By using velocity characteristics based on a certain throttle position, the engine can automatically modify its work point to output the desired power. When the engine's output power equals to the generator's output power which is namely the optimal output power, the whole engine\generator\battery system converge to a balance point.

[0057] The present invention introduces an intelligent predictable control system into the HEV to obtain the future drive load level in accordance with the different further road situation and the recorded previous data. Accordingly, the VMS of an embodiment of the present invention analyzes the future load level, vehicle speed and SOC of the battery to determine the optimal output power from engine/generator system in a future time slot. The VMS then coordinates the engine, the battery and the generator by sending the corresponding commands to the ECU, the BCU and the GCU. The ECU, GCU and BCU will adjust the corresponding devices to desired points. The HEVs energy control strategy of the present invention balances the goals of fuel economy while always meeting demand power and maintain the functionality of the battery system.

[0058] The invention has been described in an illustrative manner, and it is to be understood that the terminology that has been used is intended to be in the nature of words of description rather than of limitation. Obviously, many modifications and variations of the present invention are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims the invention may be practiced otherwise than as specifically described.

Claims

1. An energy management system for a hybrid electric vehicle (HEV), comprising: a load forecast system configured to carry out a learning on a speed sequence, an accelerator pedal sequence and a brake pedal sequence received from the HEV to obtain a future load level associated with the HEV; and a vehicle management system configured to receive the future load level from the load forecasting system, information associated with a state of charge (SOC) from a battery of the HEV, and a vehicle speed from the HEV; wherein the vehicle management system is further configured to determine an optimal future power output based on the received future load level, the information associated with the SOC of the battery and the vehicle speed, so as to coordinate operation of an engine, a generator and the battery of the HEV.

2. The system according to claim 1, wherein the vehicle management system further comprises: an energy control unit, a generator control unit and a battery control unit configured to respectively control outputs from the engine , the generator and the battery for a future time slot.

3. The system according to claim 1, wherein the load forecast system carries out the learning on the received speed sequence, accelerator pedal sequence and brake pedal sequence to obtain the future load level by means of a fuzzy logic and a neural network embedded in the load forecast system.

4. The system according to claim 3, wherein the neural network comprises a Cascade Neural Network with Node-Decoupled Extended Kalman Filtering.

5. The system according to claim 1, wherein the vehicle management system functions as a learning machine to determine the optimal future power output based on the future load level, the information associated with the SOC of the battery and the vehicle speed.

6. The system according to claim 5, wherein the learning machine is selected from a group comprising a Support Vector Machine, a Radial Basis Function Network, a Nearest Neighbor Algorithm, and a Fisher Linear Discriminant.

7. A method for managing energy of a hybrid electric vehicle (HEV), comprising: retrieving a speed sequence, an accelerator pedal sequence and a brake pedal sequence from the HEV; carrying out a learning on the retrieved sequences to obtain a future load level associated with the HEV; retrieving information associated with a state of charge (SOC) of a battery of the HEV, and a vehicle speed of the HEV; and determining an optimal future power output based on the obtained future load level, the retrieved information associated with the SOC and the vehicle speed, so as to coordinate operation of an engine, a generator and the battery of the HEV.

8. The method according to claim 7, wherein determining the optimal future power output further comprises: turning on the engine and charging the battery if the SOC is equal to or smaller than a first predetermined value.

9. The method according to claim 8, wherein the first predetermined value is 0.4.

10. The method according to claim 8, wherein determining the optimal future power output further comprises: turning off the engine and powering the vehicle only by the battery if the SOC is equal to or larger than a second predetermined value.

11. The method according to claim 10, wherein the second predetermined value is 0.8.

12. The method according to claim 7, wherein determining the optimal future power output further comprises: determining the optimal power output of the engine in a future time slot based on the future load level, the information associated with the SOC of the battery and the vehicle speed if the SOC is larger than the first predetermined value and smaller than the second predetermined value; and coordinating the operation of the engine, the generator and the battery based on the determined optimal power output.

13. The method according to claim 7, wherein obtaining the future load level is performed by using a fuzzy logic and a neural network.

14. The method according to claim 13, wherein the neural network comprises a Cascade Neural Network with Node-Decoupled Extended Kalman Filtering.

15. The method according to claim 14, wherein obtaining the future load level further comprises a step of preparing training data by said Cascade Neural Network and further comprises preparing training data.

16. The method according to claim 7, wherein determining the optimal future power output is performed by using a learning machine which is built up off-line.

17. The method according to claim 16, wherein the learning machine comprises any one of Support Vector Machine, Radial Basis Function Network, Nearest Neighbor Algorithm, Fisher Linear Discriminant.