An intelligent energy management method for hybrid electric vehicles based on slope prediction

Abstract

The present invention provides a hybrid vehicle (HEV) intelligent energy management method based on slope prediction, comprising the following steps: establishing models for each energy source and the entire vehicle; predicting the total power demand using a long-short-term memory (LSTM) network based on the vehicle's current slope and speed; employing a fuzzy-controlled adaptive low-pass filter and an equivalent power minimization strategy to perform hierarchical power processing; and constructing a HEV energy management system based on slope prediction. Learning objectives include the total equivalent hydrogen consumption of each energy source, the lifespan of the fuel cell, and the system-on-chip (SoC) of the energy storage system under predicted slope. An efficient hybrid experience replay technique is designed to update the network in the intelligent algorithm. This method allows for online optimization and adjustment of HEV energy management based on road slope information. By using real-time road slope prediction, the optimal strategy can be updated in complex and random driving environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of hybrid vehicle energy management, and in particular relates to an intelligent energy management method for a hybrid vehicle based on slope prediction. Background Art

[0002] To address the air pollution and oil resource depletion caused by traditional gasoline and diesel vehicles, vehicle electrification has become a promising alternative. Fuel cell hybrid vehicles (FCEVs) stand out for their pollution-free, zero-emission, and long driving range. However, energy management technology for these new energy hybrid vehicles is still immature, and most rely on simple energy management systems with fixed operating conditions and limited environmental requirements. This lacks a comprehensive and integrated approach to optimizing energy distribution and improving fuel economy. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent energy management method for hybrid electric vehicles based on slope prediction. Taking into account road slope information, changes in motor load power demand, and the SoC of the energy storage system, the method can minimize fuel consumption, extend the service life of the energy storage system and fuel cell, and achieve online optimization and adjustment of active energy management of fuel cell hybrid electric vehicles.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is: a hybrid electric vehicle intelligent energy management method based on slope prediction, comprising the following steps:

[0005] Step S1: Establishing the energy source models and the whole vehicle model in the hybrid vehicle;

[0006] Step S2: Based on the vehicle's current slope and speed information, a long short-term memory network is used to predict the total power demand. The power is processed in layers using a fuzzy control adaptive low-pass filter and an equivalent power minimization strategy to construct a hybrid vehicle energy management system based on slope prediction.

[0007] Step S3: Taking the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope as learning objectives, an efficient hybrid experience replay technology is designed to update the network, and real-time energy management of hybrid vehicles is performed through an improved dual-delay deep deterministic policy gradient intelligent algorithm.

[0008] Furthermore, the method of establishing the energy source models and the whole vehicle model in the hybrid vehicle in step S1 includes:

[0009] 1) Establish a hydrogen-oxygen fuel cell model:

[0010] V FC =Ncell ×(E cell -V act.loss -V ohm.loss )

[0011]

[0012] Where V FC is the output voltage of the fuel cell stack, N cell is the number of cells on the fuel cell stack, E cell is the electromotive force, V act.loss is the loss when the battery is activated, V ohm.loss is the ohmic loss of the battery, η FC For the efficiency of the fuel cell, is the theoretical power related to hydrogen flow consumption in the fuel cell, P FC is the power of the fuel cell;

[0013] 2) Establish a lithium battery model:

[0014] V BAT =V(SoC BAT ) BAT.oc -βI BAT R(SoC BAT )

[0015]

[0016]

[0017] Where V BAT is the lithium battery voltage, V(SoC BAT ) BAT .o c and R(SoC BAT ) are respectively when the lithium battery SoC is SoC BAT The open circuit voltage and internal resistance of the lithium battery, β = ± 1 is the lithium battery charge and discharge state selection coefficient (negative when charging, positive when discharging), I BAT is the lithium battery current, P BAT is the power of lithium battery, SoC ini It is the initial SoC of lithium battery, C nom is the rated capacity of the lithium battery;

[0018] 3) Establish supercapacitor model:

[0019] V UC =SoC UC ·(V UC.max -V UC.min )+V UC.min

[0020]

[0021] Where V UC is the supercapacitor voltage, SoC UC is the state of charge of the supercapacitor, V UC.max and V UC.min are the maximum and minimum output voltages of the supercapacitor, R UC and I UC is the equivalent internal resistance and current of the supercapacitor;

[0022] 4) Establish a hybrid vehicle system model:

[0023]

[0024] P demand =η FC P FC +η BAT P BAT +η UC P UC

[0025] Where, P demand is the load power demand, C D is the vehicle aerodynamic drag coefficient, A s is the frontal area, v is the vehicle speed, Δ is the air density, μ is the road rolling resistance coefficient, m and θ are the vehicle mass and road slope respectively, g is the acceleration of gravity, δ is the vehicle rotation mass conversion coefficient, η FC ,η BAT and η UC are the efficiency of the DC / DC converter connected to the fuel cell, lithium battery and supercapacitor, P FC 、P BAT and P UC are the output power of fuel cells, lithium batteries and supercapacitors respectively. When lithium batteries and supercapacitors are discharged, P BAT and P UC is positive; when the lithium battery and supercapacitor are charging, P BAT and P UC is negative.

[0026] Furthermore, in step S2, the total power demand is predicted using a long short-term memory network based on the current slope and speed information of the vehicle. The power is processed in layers using a fuzzy control adaptive low-pass filter and an equivalent consumption minimization strategy. The method for constructing a hybrid vehicle energy management system based on slope prediction is as follows:

[0027] 1) Take the current accelerator pedal change amplitude and road slope information as input variables, make short-term predictions through the long short-term memory network, and obtain the corresponding total load demand power P based on the established vehicle system model. demand ;

[0028] 2) Obtain the weighted SoC of the energy storage system (lithium battery and supercapacitor). The calculation principle of supercapacitor SoC and lithium battery SoC is the same, both using the integration method;

[0029] 3) The total required power P demand The weighted SoC of the energy storage system is input as input variables to the designed fuzzy inference system. After fuzzification, fuzzy inference and defuzzification, the adjustable frequency f is obtained. s ;

[0030] 4) Adjust the frequency f s and the total required power P demand Input to the low-pass filter, the required power P demand Perform frequency separation, and the output is high frequency power P demand.h and mid- and low-frequency power P demand.ml , where high-frequency power is borne by supercapacitors, and medium- and low-frequency power is shared by fuel cells and lithium batteries;

[0031] 5) The medium and low frequency power P obtained by decoupling the fuzzy controlled adaptive low-pass filter demand.ml Based on the idea of ​​the equivalent consumption minimum strategy, the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope are taken as learning objectives to design a multi-objective optimization problem:

[0032] minC total (t) = λ FC C FC (t)+λ BAT C BAT (t)+λ UC C UC (t)

[0033] minΔP FC =n·(|P FC(t) -P FC(t-1) |-10%P FC_max )

[0034]

[0035] Where minC total (t) is the total instantaneous minimum hydrogen consumption, which is mainly composed of the direct hydrogen consumption of the fuel cell C FC (t), equivalent hydrogen consumption of lithium battery C BAT(t) and supercapacitor equivalent hydrogen consumption C UC (t) composition, λ FC The penalty factor to ensure that the fuel cell operates in the high efficiency range, λ BAT and λ UC is the equivalent factor of lithium battery and supercapacitor, ΔSoC BAT SoCSoC is the reference value obtained by dynamic programming strategy under the current SoC of lithium battery and predicted slope ref The deviation between FC Indicates the situation of large power fluctuation of fuel cell, n represents the number of times the output power of fuel cell fluctuates by more than 10% of its rated power, P FC(t) and P FC(t-1) are the output power of FC at the current moment and the previous moment, P FC_max is the maximum output power of FC, i.e. rated power;

[0036] 6) Use the multi-objective optimization problem in the above steps as the reward function in the improved deep reinforcement learning algorithm. Since the goal of reinforcement learning is to maximize the expected return, the reward function R is taken as:

[0037] R=-[C total +αΔP FC +β(ΔSoC BAT ) 2 ]

[0038] In the formula, α and β are the total (t), ΔP FC and (DSoC) 2 have adjustment coefficients of the same order of magnitude.

[0039] Furthermore, in step S3, the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope are used as learning objectives, and an efficient hybrid experience replay technology is designed to update the network. The method for real-time energy management of hybrid electric vehicles using an improved double-delay deep deterministic policy gradient intelligent algorithm is as follows:

[0040] f) Using five different historical road slope data to simulate various driving scenarios to train the neural network, and using dynamic programming to solve multiple times to obtain the optimal experience samples under the corresponding working conditions And store it in the optimal experience pool OER for assisting deep reinforcement algorithm learning;

[0041] g) Collect slope data in real time during vehicle driving and combine it with the experience samples (s t ,a t ,r t ,s t+1) is stored in the real-time experience pool IER;

[0042] h) During the network training phase, a priority sampling method based on a ranking mechanism is used for the experience in the IER, while a conventional random uniform sampling method is used for the experience in the OER due to its inherent optimality. Through the hybrid experience replay technology, the network parameters are updated efficiently;

[0043] i) Taking the total equivalent hydrogen consumption of each energy source in step S2 as the minimum, the number of high power fluctuations of the fuel cell as the minimum, and the SoC of the energy storage system under the predicted slope as the optimization goal, the SoC of the lithium battery and supercapacitor, the total required power P demand The speed v of the working condition, the road slope, the lithium battery reference SoC obtained by dynamic programming, and the power of the supercapacitor are used as states. The improved double-delay deep deterministic policy gradient TD3 algorithm is used to optimize the neural network parameters and obtain the corresponding action output, which is the output power ratio of the fuel cell and lithium battery at the corresponding moment;

[0044] j) Using the output power ratio of the previous step, calculate online to obtain the current fuel cell power P FC And the power of lithium battery P BAT

[0045] According to the above steps, active energy management of hybrid vehicles can be achieved, and the corresponding fuel cell power P can be obtained in real time. FC , lithium battery power P BAT and supercapacitor power P UC .

[0046] The beneficial effects of adopting this technical solution are:

[0047] 1. The present invention uses a long short-term memory network to predict the total power demand in real time based on the current slope and vehicle speed information. Based on the changes in the total power demand and the weighted SoC of the energy storage system, an adaptive low-pass filter based on fuzzy control is used to perform frequency decoupling on the total power demand. High-frequency power is borne by supercapacitors, while medium and low-frequency power is shared by fuel cells and lithium batteries. This avoids the impact of peak power on fuel cells and lithium batteries, and improves the dynamic performance of the vehicle.

[0048] 2. This invention adopts the concept of minimum equivalent consumption strategy, takes the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under predicted slope as learning objectives, designs a multi-objective optimization function model, realizes the coordinated and active adjustment of the fuel cell hybrid vehicle energy management system, and improves the vehicle's economy.

[0049] 3. The present invention adopts an improved dual-delay deep deterministic policy gradient intelligent algorithm and effectively updates the network by designing an efficient hybrid experience replay technology, which can improve the adaptability and optimization of the energy management system when the vehicle faces complex and random driving environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a structural diagram of the hybrid vehicle intelligent energy management system based on slope prediction according to the present invention;

[0051] Figure 2 This is a structural diagram of the hybrid vehicle power system of the present invention;

[0052] Figure 3 Improved deep reinforcement learning intelligent algorithm block diagram for the present invention;

[0053] Figure 4 This is the fuzzy rule table of the present invention. DETAILED DESCRIPTION

[0054] The technical solution proposed by the present invention is further described and illustrated below with reference to the accompanying drawings:

[0055] The structure diagram of the hybrid vehicle intelligent energy management system based on slope prediction of the present invention is as follows: Figure 1 As shown: The object of study in this invention is a hybrid vehicle with proton exchange membrane fuel cell, lithium battery and super capacitor as energy source. The power system structure of the hybrid vehicle is as shown in FIG. Figure 2 As shown, the fuel cell serves as the main power source and is connected to the DC bus via a unidirectional DC / DC converter. The hydrogen involved in the chemical reaction comes from a hydrogen cylinder. The supercapacitor, as a power pulse device, is connected to the DC bus via a bidirectional DC / DC converter to recover or generate transient power during sudden acceleration and deceleration of the vehicle. The lithium battery serves as an energy buffer and is connected to the DC bus via a bidirectional DC / DC converter to assist the fuel cell in better meeting the power requirements of the vehicle, reducing hydrogen consumption and improving the dynamic performance of the vehicle. Finally, the electrical energy on the DC bus is converted by a DC / AC inverter and then transmitted to the three-phase motor to drive the vehicle. The present invention is an intelligent energy management method for hybrid electric vehicles based on slope prediction, comprising the following steps:

[0056] Step S1: Establishing the energy source models and the whole vehicle model in the hybrid vehicle;

[0057] Step S2: Based on the vehicle's current slope and speed information, a long short-term memory network is used to predict the total power demand. The power is processed in layers using a fuzzy control adaptive low-pass filter and an equivalent power minimization strategy to construct a hybrid vehicle energy management system based on slope prediction.

[0058] Step S3: Taking the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under predicted slope as learning objectives, an efficient hybrid experience replay technology is designed to update the network. The improved double-delay deep deterministic policy gradient intelligent algorithm is used to perform real-time energy management of the hybrid vehicle.

[0059] The method for establishing the energy source models and the whole vehicle model in the hybrid vehicle in step S1 includes:

[0060] 1) Establish a hydrogen-oxygen fuel cell model:

[0061] V FC =N cell ×(E cell -V act.loss -V ohm.loss )

[0062]

[0063] Where V FC is the output voltage of the fuel cell stack, N cell is the number of cells on the fuel cell stack, E cell is the electromotive force, V act.loss is the loss when the battery is activated, V ohm.loss is the ohmic loss of the battery, η FC For the efficiency of the fuel cell, is the theoretical power related to hydrogen flow consumption in the fuel cell, P FC is the power of the fuel cell;

[0064] 2) Establish a lithium battery model:

[0065] V BAT =V(SoC BAT ) BAT.oc -βI BAT R(SoC BAT )

[0066]

[0067]

[0068] Where V BAT is the lithium battery voltage, V(SoC BAT ) BAT.oc and R(SoC BAT ) are respectively when the lithium battery SoC is SoC BAT The open circuit voltage and internal resistance of the lithium battery, β = ± 1 is the lithium battery charge and discharge state selection coefficient (negative when charging, positive when discharging), I BAT is the lithium battery current, P BATis the power of lithium battery, SoC ini It is the initial SoC of lithium battery, C nom is the rated capacity of the lithium battery;

[0069] 3) Establish supercapacitor model:

[0070] V UC =SoC UC ·(V UC.max -V UC.min )+V UC.min

[0071]

[0072] Where V UC is the supercapacitor voltage, SoC UC is the state of charge of the supercapacitor, V UC.max and V UC.min are the maximum and minimum output voltages of the supercapacitor, R UC and I UC is the equivalent internal resistance and current of the supercapacitor;

[0073] 4) Establish a hybrid vehicle system model:

[0074]

[0075] P demand =η FC P FC +η BAT P BAT +η UC P UC

[0076] Where, P demand is the load power demand, C D is the vehicle aerodynamic drag coefficient, A s is the frontal area, v is the vehicle speed, Δ is the air density, μ is the road rolling resistance coefficient, m and θ are the vehicle mass and road slope respectively, g is the acceleration of gravity, δ is the vehicle rotation mass conversion coefficient, η FC ,η BAT and η UC are the efficiency of the DC / DC converter connected to the fuel cell, lithium battery and supercapacitor, P FC 、P BAT and P UC are the output power of fuel cells, lithium batteries and supercapacitors respectively. When lithium batteries and supercapacitors are discharged, P BAT and P UC is positive; when the lithium battery and supercapacitor are charging, P BAT and P UCis negative;

[0077] In step S2, the total power demand is predicted using a long short-term memory network based on the vehicle's current slope and speed information. A fuzzy-controlled adaptive low-pass filter and an equivalent consumption minimization strategy are used to perform hierarchical processing on the power. A method for constructing an intelligent energy management system for hybrid electric vehicles based on slope prediction is as follows:

[0078] 1) Take the current accelerator pedal change amplitude and road slope information as input variables, make short-term predictions through the long short-term memory network, and obtain the corresponding total load demand power P based on the established vehicle system model. demand ;

[0079] 2) Obtain the weighted SoC of the energy storage system (lithium battery and supercapacitor). The calculation principle of supercapacitor SoC and lithium battery SoC is the same, both using the integration method;

[0080] 3) The total required power P demand The weighted SoC of the energy storage system is input as input variables to the designed fuzzy inference system. After fuzzification, fuzzy inference and defuzzification, the adjustable frequency f is obtained. s ;

[0081] 4) Adjust the frequency f s and the total power demand P demand Input to the low-pass filter, the required power P demand Perform frequency separation, and the output is high frequency power P demand.h and mid- and low-frequency power P demand.ml , where high-frequency power is borne by supercapacitors, and medium- and low-frequency power is shared by fuel cells and lithium batteries;

[0082] 5) The medium and low frequency power P obtained by decoupling the fuzzy controlled adaptive low-pass filter demand.ml Based on the idea of ​​the equivalent consumption minimum strategy, the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope are taken as learning objectives to design a multi-objective optimization problem:

[0083] minC total (t) = λ FC C FC (t)+λ BAT C BAT (t)+λ UC C UC (t)

[0084] minΔP FC =n·(|P FC(t) -P FC(t-1) |-10%PFC_max )

[0085]

[0086] Where minC total (t) is the total instantaneous minimum hydrogen consumption, which is mainly composed of the direct hydrogen consumption of the fuel cell C FC (t), equivalent hydrogen consumption of lithium battery C BAT (t) and supercapacitor equivalent hydrogen consumption C UC (t) composition, λ FC The penalty factor to ensure that the fuel cell operates in the high efficiency range, λ BAT and λ UC is the equivalent factor of lithium battery and supercapacitor, ΔSoC BAT SoCSoC is the reference value obtained by dynamic programming strategy under the current SoC of lithium battery and predicted slope ref The deviation between FC Indicates the situation of large power fluctuation of fuel cell, n represents the number of times the output power of fuel cell fluctuates by more than 10% of its rated power, P FC(t) and P FC(t-1) are the output power of FC at the current moment and the previous moment, P FC_max is the maximum output power of FC, i.e. rated power;

[0087] 6) Use the multi-objective optimization problem in the above steps as the reward function in the improved deep reinforcement learning algorithm. Since the goal of reinforcement learning is to maximize the expected return, the reward function R is taken as:

[0088] R=-[C total +αΔP FC +β(ΔSoC BAT ) 2 ]

[0089] In the formula, α and β are the total (t), ΔP FC and (DSoC) 2 have adjustment coefficients of the same order of magnitude;

[0090] like Figure 3 As shown, in step S3, the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope are used as learning objectives. An efficient hybrid experience replay technology is designed to update the network. The method for real-time energy management of hybrid vehicles using the improved double-delay deep deterministic policy gradient intelligent algorithm is as follows:

[0091] k) Use 5 different historical road slope data to simulate various driving scenarios to train the neural network, and use dynamic programming to solve multiple times to obtain the optimal experience sample under the corresponding working conditions And store it in the optimal experience pool OER for assisting deep reinforcement algorithm learning;

[0092] l) Collect slope data in real time during vehicle driving and use it to interact with the agent to obtain experience samples (s t ,a t ,r t ,s t+1 ) is stored in the real-time experience pool IER;

[0093] m) During the network training phase, a priority sampling method based on a ranking mechanism is used for the experiences in the IER, while a conventional random uniform sampling method is used for the experiences in the OER due to their inherent optimality. Through the hybrid experience replay technology, the network parameters are updated efficiently;

[0094] n) Taking the minimum total equivalent hydrogen consumption of each energy source in step S2, the minimum number of high power fluctuations of the fuel cell, and the maintenance of the SoC of the energy storage system under the predicted slope as the optimization goals, the SoC of the lithium battery and supercapacitor, the total required power P demand The speed v of the working condition, the road slope, the lithium battery reference SoC obtained by dynamic programming, and the power of the supercapacitor are used as states. The improved double-delay deep deterministic policy gradient TD3 algorithm is used to optimize the neural network parameters and obtain the corresponding action output, which is the output power ratio of the fuel cell and lithium battery at the corresponding moment;

[0095] o) Using the output power ratio of the previous step, calculate online to obtain the current fuel cell power P FC And the power of lithium battery P BAT ,

[0096] According to the above steps, active energy management of hybrid vehicles can be achieved, and the corresponding fuel cell power P can be obtained in real time. FC , lithium battery power P BAT and supercapacitor power P UC .

Claims

1. A hybrid electric vehicle intelligent energy management method based on slope prediction, comprising the following steps: Step S1: Establishing the energy source models and the whole vehicle model in the hybrid vehicle; Step S2: Based on the vehicle's current slope and speed information, a long short-term memory network is used to predict the total power demand. The power is processed in layers using a fuzzy control adaptive low-pass filter and an equivalent power minimization strategy to construct a hybrid vehicle energy management system based on slope prediction. Step S3: Taking the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under predicted slope as learning objectives, an efficient hybrid experience replay technology is designed to update the network. The improved double-delay deep deterministic policy gradient intelligent algorithm is used to perform real-time energy management of the hybrid vehicle. In step S2, the total power demand is obtained by predicting the vehicle's current slope and speed using a long short-term memory network. A fuzzy-controlled adaptive low-pass filter and an equivalent consumption minimization strategy are used to perform power layering processing. A method for constructing a hybrid vehicle energy management system based on slope prediction is as follows: 1) Taking the current accelerator pedal change amplitude and road slope information as input variables, a short-term prediction is performed through the long short-term memory network, and based on the established vehicle system model, the corresponding total load demand power P is obtained. demand ; 2) Obtain the weighted SoC of the energy storage system, which includes lithium batteries and supercapacitors. The supercapacitor SoC calculation principle is the same as the lithium battery SoC calculation principle, both using the integration method; 3) The total required power P demand The weighted SoC of the energy storage system is input as input variables to the designed fuzzy inference system. After fuzzification, fuzzy inference and defuzzification, the adjustable frequency f is obtained. s ; 4) Adjust the adjustable frequency f s and the total power demand P demand Input to the low-pass filter, the required power P demand Perform frequency separation, and the output is high frequency power P demand.h and mid- and low-frequency power P demand.ml , where high-frequency power is borne by supercapacitors, and medium- and low-frequency power is shared by fuel cells and lithium batteries; 5) The medium and low frequency power P obtained by decoupling the fuzzy controlled adaptive low-pass filter demand.ml Based on the idea of ​​the equivalent consumption minimum strategy, the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope are taken as learning objectives to design a multi-objective optimization problem: minC total (t)=λ FC C FC (t)+λ BAT C BAT (t)+λ UC C UC (t) minΔP FC n·(|P FC(t) -P FC(t-1) |-10%P FC_max ) Where minC total (t) is the total instantaneous minimum hydrogen consumption, which is mainly composed of the direct hydrogen consumption of the fuel cell C FC (t), equivalent hydrogen consumption of lithium battery C BAT (t) and supercapacitor equivalent hydrogen consumption C UC (t) composition, λ FC The penalty factor to ensure that the fuel cell operates in the high efficiency range, λ BAT and λ UC is the equivalent factor of lithium battery and supercapacitor, ΔSoC BAT The current SoC of the lithium battery and the reference value SoC obtained by the dynamic programming strategy under the predicted slope ref The deviation between FC Indicates the situation of large power fluctuation of fuel cell, n represents the number of times the output power of fuel cell fluctuates by more than 10% of its rated power, P FC(t) and P FC(t-1) are the output power of FC at the current moment and the previous moment, P FC_max is the maximum output power of FC, i.e. rated power; 6) The multi-objective optimization problem in the above steps is used as the reward function in the improved deep reinforcement learning algorithm. Since the goal of reinforcement learning is to maximize the expected return, the reward function R is taken as: R=-[C total (t)+αΔP FC +β(ΔSoC BAT ) 2 ] In the formula, α and β are the total (t), ΔP FC and (ΔSoC BAT ) 2 have adjustment coefficients of the same order of magnitude; In step S3, the total equivalent hydrogen consumption of each energy source, the life of the fuel cell, and the SoC of the energy storage system under the predicted slope are used as learning objectives. An efficient hybrid experience replay technology is designed to update the network. The method for real-time energy management of hybrid electric vehicles using an improved double-delay deep deterministic policy gradient intelligent algorithm is as follows: a) Using five different historical road slope data to simulate various driving scenarios to train the neural network, and using dynamic programming to solve multiple times to obtain the optimal experience samples under the corresponding working conditions And store it in the optimal experience pool OER for assisting deep reinforcement algorithm learning; b) Collect slope data in real time during vehicle driving and use it to interact with the agent to obtain experience samples (s t ,a t ,r t ,s t+1 ) is stored in the real-time experience pool IER; c) During the network training phase, a priority sampling method based on a ranking mechanism is used for the experience in the IER, while a conventional random uniform sampling method is used for the experience in the OER due to its inherent optimality. Through the hybrid experience replay technology, the network parameters are updated efficiently; d) Taking the total equivalent hydrogen consumption of each energy source in step S2 as the minimum, the number of high power fluctuations of the fuel cell as the minimum, and the maintenance of the SoC of the energy storage system under the predicted slope as the optimization goal, the SoC of the lithium battery and supercapacitor, the total required power P demand The speed v of the working condition, the road slope, the lithium battery reference SoC obtained by dynamic programming, and the power of the supercapacitor are used as states. The improved double-delay deep deterministic policy gradient TD3 algorithm is used to optimize the neural network parameters and obtain the corresponding action output, which is the output power ratio of the fuel cell and lithium battery at the corresponding moment; e) Using the output power ratio of the previous step, calculate online to obtain the current fuel cell power P FC And the power of lithium battery P BAT According to the above steps, active energy management of hybrid vehicles can be achieved, and the corresponding fuel cell power P can be obtained in real time. FC , lithium battery power P BAT and supercapacitor power P UC .

2. The hybrid electric vehicle intelligent energy management method based on slope prediction according to claim 1, characterized in that: The method for establishing the energy source models and the whole vehicle model in the hybrid vehicle in step S1 includes: 1) Establish a hydrogen-oxygen fuel cell model: V FC =N cell ×(E cell -V act.loss -V ohm.loss ) Where V FC is the output voltage of the fuel cell stack, N cell is the number of cells on the fuel cell stack, E cell is the electromotive force, V act.loss is the loss when the battery is activated, V ohm.loss is the ohmic loss of the battery, η FC For the efficiency of the fuel cell, is the theoretical power related to hydrogen flow consumption in the fuel cell, P FC is the power of the fuel cell; 2) Establish lithium battery model: V BAT =V(SoC BAT ) BAT.oc -βI BAT R(SoC BAT ) Where V BAT is the lithium battery voltage, V(SoC BAT ) BAT.oc and R(SoC BAT ) are respectively when the lithium battery SoC is SoC BAT The open circuit voltage and internal resistance of the lithium battery, β = ± 1 is the lithium battery charge and discharge state selection coefficient, where it is negative when charging and positive when discharging, I BAT is the lithium battery current, P BAT is the power of lithium battery, SoC BAT.ini It is the initial SoC of lithium battery, C nom is the rated capacity of the lithium battery; 3) Establish supercapacitor model: V UC =SoC UC ·(V UC.max -V UC.min )+V UC.min Where V UC is the supercapacitor voltage, SoC UC is the state of charge of the supercapacitor, V UC.max and V UC.min are the maximum and minimum output voltages of the supercapacitor, R UC and I UC is the equivalent internal resistance and current of the supercapacitor; 4) Establish a hybrid vehicle system model: P demand =the FC P FC +n BAT P BAT +n UC P UC Where, P demand is the load power demand, C D is the vehicle aerodynamic drag coefficient, A s is the frontal area, v is the vehicle speed, Δ is the air density, μ is the road rolling resistance coefficient, m and θ are the vehicle mass and road slope respectively, g is the acceleration of gravity, δ is the vehicle rotation mass conversion coefficient, η FC ,η BAT and η UC are the efficiency of the DC / DC converter connected to the fuel cell, lithium battery and supercapacitor, P FC 、P BAT and P UC are the output power of fuel cells, lithium batteries and supercapacitors respectively. When lithium batteries and supercapacitors are discharged, P BAT and P UC is positive; when the lithium battery and supercapacitor are charging, P BAT and P UC is negative.

Images