Fuel cell automobile energy management strategy fused with traffic perception

Abstract

The invention provides a fuel cell automobile energy management strategy fused with traffic perception, which constructs a future speed prediction model based on a Transform network, can effectively represent a dynamic interaction relationship between a main automobile and a plurality of surrounding automobiles, and has higher accuracy and stability compared with a traditional automobile speed prediction mode based on historical data or single automobile information; a deep integration model structure is also introduced for speed prediction, the uncertainty of a prediction result can be quantified, and good prediction credibility judgment and output adaptive adjustment can be realized in combination with a dynamic threshold mechanism, so that the robustness of the method in a complex traffic scene can be remarkably improved; and a reinforcement learning energy management strategy is constructed based on the predicted future speed sequence, so that the future traffic change can be perceived in advance and quickly coped, and the method has obvious advantages in the aspect of fuel economy.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy management of intelligent connected hybrid vehicles, and specifically relates to an energy management strategy for fuel cell vehicles integrating traffic perception. Background Art

[0002] For hybrid powertrains, energy management strategies, as the core link that coordinates the output power of each energy unit, determine the energy consumption and economic level of the entire vehicle. With the development of related technologies, traditional energy management methods based solely on rules or optimization are gradually unable to meet current needs. Recently emerging energy management strategies based on reinforcement learning have been widely deployed due to their advantages such as model independence, strong self-learning capabilities, and realistic environmental representation. However, this approach relies too much on current and historical operating condition data and lacks a good prediction mechanism for short-term future information. As a result, its real-time performance is still insufficient, limiting its energy-saving performance. Summary of the Invention

[0003] In view of this, and in response to the technical problems existing in this field, the present invention provides a fuel cell vehicle energy management strategy integrating traffic perception, which specifically includes the following steps:

[0004] Step 1: Obtain traffic perception information through on-board sensors and vehicle-to-everything (V2X) communication equipment, collect the speed and relative distance of at least six adjacent vehicles around the main vehicle (left front, front, right front, left rear, rear, right rear), and construct a traffic state input vector in real time;

[0005] Step 2: Build a speed prediction model based on the Transformer algorithm, which includes several sub-models. The traffic state input vector obtained in Step 1 is used as the model input, and the predicted future speed sequence of the main vehicle is used as the model output. Each sub-model is trained independently. A deep ensemble method is used to calculate the mean and variance of the future speed sequence predictions obtained by each sub-model, and the uncertainty quantification index of the model prediction is determined from this.

[0006] Step 3: Set corresponding thresholds for the uncertainty quantification indicators and compare the two to determine whether the current prediction results are credible. If the results are credible, the prediction results of the speed prediction model are directly output. If the results are not credible, exponential smoothing prediction is performed to output credible prediction results.

[0007] Step 4: Define the system state space based on the future speed sequence and the vehicle energy state, define the action space based on the fuel cell system power, and design and define the corresponding reward function taking into account the energy consumption of the fuel cell and power battery. On this basis, build an energy management optimizer based on deep reinforcement learning. It can select the corresponding fuel cell power control action based on the predicted future speed sequence and the current vehicle energy state, obtain the corresponding reward, and enter the next state;

[0008] Step 5: Train the energy management optimizer based on reinforcement learning on historical operating condition data, and use the corresponding action-reward evaluation mechanism to update the optimizer network parameters so that the output fuel cell control instructions tend to be optimal;

[0009] Step 6: Deploy the trained energy management optimizer for online application.

[0010] Furthermore, the traffic state input vector x constructed in step 1 t The specific form is as follows:

[0011] x t =[v0,v1,…,v6,d1,…,d6]

[0012] Where, v0 is the current speed of the main vehicle; v i (i=1,…,6) is the current speed of the i-th adjacent vehicle; d i (i=1,…,6) is the relative distance between the i-th adjacent vehicle and the main vehicle.

[0013] Furthermore, the speed prediction model established in step 2 specifically uses the following multi-head attention mechanism to dynamically process traffic scene information:

[0014]

[0015] Q=XW Q K=XW K V=XW V

[0016] X=[x1,x2,...,x T ]

[0017] Where Q is the query vector matrix; K is the key vector matrix; V is the value vector matrix; W Q 、W K 、W V All are trainable parameters; X is the input sequence, specifically the traffic state input vector x of T time steps t composition; Used to measure the similarity between each time step; Used for scaling, avoiding dimension values ​​d kToo large leads to unstable gradients; the result obtained after performing softmax normalization is used as the weight, and the final representation of each time step is obtained from it.

[0018] Furthermore, in step 2, the deep integration method is used to calculate the mean and variance of the vehicle speed series predicted by the sub-model using the following formula:

[0019]

[0020] Where N is the number of sub-models; is the velocity prediction result of the nth sub-model at time t+h; μ h is the mean of the prediction results at time t+h; is the prediction variance, H is the prediction time step window length, and U is the uncertainty quantification indicator.

[0021] Furthermore, in step 3, the threshold δ of the uncertainty quantification index is specifically set t The form is as follows:

[0022]

[0023] in, is the historical uncertainty mean from tk to time t; β is the empirically set proportional coefficient used to adjust the prediction tolerance; k is the length of the historical sliding window;

[0024] The uncertainty quantification index is compared with the threshold δ t For comparison, if U≤δ t The current prediction result is determined to be credible, and the future speed sequence prediction result of the speed prediction model is directly output; if U>δ t The current prediction result is determined to be unreliable, and the following exponential smoothing is performed to output the future speed:

[0025]

[0026] in, is the future speed predicted by exponential smoothing; v t is the current speed of the main vehicle; α∈(0,1) is the smoothing coefficient, which is used to control the weight of the current speed and historical trend.

[0027] Furthermore, in step 4, the following state space is specifically defined:

[0028]

[0029] in, is the predicted future speed series; a t is the current acceleration; P fc,t is the power output of the current fuel cell system; ΔPfc,t =P fc,t -P fc,t-1 is the fuel cell power change rate; SOC t The current state of charge of the power battery;

[0030] Define the following action space:

[0031]

[0032] in, Control actions for energy management strategies; are the minimum and maximum limits of fuel cell power respectively;

[0033] And design the following reward function:

[0034]

[0035] Among them, r t is the reward value at the current moment; SOC t The current state of charge of the power battery; SOC ref is the reference target state of charge; ω1 and ω2 are the weight coefficients of each item in the reward function.

[0036] The fuel cell vehicle energy management strategy integrated with traffic perception provided by the present invention constructs a future speed prediction model based on the Transformer network, which can effectively characterize the dynamic interaction relationship between the main vehicle and multiple surrounding vehicles, and has higher accuracy and stability than the traditional vehicle speed prediction method based on historical data or single vehicle information; a deep integration model structure is also introduced for speed prediction, which can quantify the uncertainty of the prediction results. Combined with the dynamic threshold mechanism, it can achieve better prediction credibility judgment and output adaptive adjustment, thereby significantly improving the robustness of the method in complex traffic scenarios; based on the predicted future speed sequence, a reinforcement learning energy management strategy is constructed, which can achieve early perception and rapid response to future traffic changes, and also has obvious advantages in fuel economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the framework of the method provided by the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] The fuel cell vehicle energy management strategy integrated with traffic perception provided by the present invention is as follows: Figure 1 As shown, the specific steps include:

[0040] Step 1: Obtain traffic perception information through on-board sensors and vehicle-to-everything (V2X) communication equipment, collect the speed and relative distance of at least six adjacent vehicles around the main vehicle (left front, front, right front, left rear, rear, right rear), and construct a traffic state input vector in real time;

[0041] Step 2: Build a speed prediction model based on the Transformer algorithm, which includes several sub-models. The traffic state input vector obtained in Step 1 is used as the model input, and the predicted future speed sequence of the main vehicle is used as the model output. Each sub-model is trained independently. A deep ensemble method is used to calculate the mean and variance of the future speed sequence predictions obtained by each sub-model, and the uncertainty quantification index of the model prediction is determined from this.

[0042] Step 3: Set corresponding thresholds for the uncertainty quantification indicators and compare the two to determine whether the current prediction results are credible. If the results are credible, the prediction results of the speed prediction model are directly output. If the results are not credible, exponential smoothing prediction is performed to output credible prediction results.

[0043] Step 4: Define the system state space based on the future speed sequence and the vehicle energy state, define the action space based on the fuel cell system power, and design and define the corresponding reward function taking into account the energy consumption of the fuel cell and power battery. On this basis, build an energy management optimizer based on deep reinforcement learning. It can select the corresponding fuel cell power control action based on the predicted future speed sequence and the current vehicle energy state, obtain the corresponding reward, and enter the next state;

[0044] Step 5: Train the energy management optimizer based on reinforcement learning on historical operating condition data, and use the corresponding action-reward evaluation mechanism to update the optimizer network parameters so that the output fuel cell control instructions tend to be optimal;

[0045] Step 6: Deploy the trained energy management optimizer for online application.

[0046] In a preferred embodiment of the present invention, the traffic state input vector x constructed in step 1 is t The specific form is as follows:

[0047] x t =[v0,v1,…,v6,d1,…,d6]

[0048] Where, v0 is the current speed of the main vehicle; v i (i=1,…,6) is the current speed of the i-th adjacent vehicle; di (i=1,…,6) is the relative distance between the i-th adjacent vehicle and the main vehicle.

[0049] In a preferred embodiment of the present invention, the speed prediction model established in step 2 specifically utilizes the following multi-head attention mechanism to dynamically process traffic scene information:

[0050]

[0051] Q=XW Q K=XW K V=XW V

[0052] X=[x1,x2,...,x T ]

[0053] Where Q is the query vector matrix; K is the key vector matrix; V is the value vector matrix; W Q 、W K 、W V All are trainable parameters; X is the input sequence, specifically the traffic state input vector x of T time steps t composition; Used to measure the similarity between each time step; Used for scaling, avoiding dimension values ​​d k Too large leads to unstable gradients; the result obtained after performing softmax normalization is used as the weight, and the final representation of each time step is obtained from it.

[0054] In a preferred embodiment of the present invention, in step 2, the mean and variance of the vehicle speed sequence predicted by the sub-model are calculated using the deep integration method using the following formula:

[0055]

[0056]

[0057] Where N is the number of sub-models; is the velocity prediction result of the nth sub-model at time t+h; μ h is the mean of the prediction results at time t+h; is the prediction variance, H is the prediction time step window length, and U is the uncertainty quantification indicator.

[0058] In a preferred embodiment of the present invention, the threshold value δ of the uncertainty quantification index is specifically set in step 3. t The form is as follows:

[0059]

[0060] in, is the historical uncertainty mean from tk to time t; β is the empirically set proportional coefficient used to adjust the prediction tolerance; k is the length of the historical sliding window;

[0061] The uncertainty quantification index is compared with the threshold δ t For comparison, if U≤δ t The current prediction result is determined to be credible, and the future speed sequence prediction result of the speed prediction model is directly output; if U>δ t The current prediction result is determined to be unreliable, and the following exponential smoothing is performed to output the future speed:

[0062]

[0063] in, is the future speed predicted by exponential smoothing; v t is the current speed of the main vehicle; α∈(0,1) is the smoothing coefficient, which is used to control the weight of the current speed and historical trend.

[0064] In a preferred embodiment of the present invention, the following state space is specifically defined in step 4:

[0065]

[0066] in, is the predicted future speed series; a t is the current acceleration; P fc,t is the power output of the current fuel cell system; ΔP fc,t =P fc,t -P fc,t-1 is the fuel cell power change rate; SOC t The current state of charge of the power battery;

[0067] Define the following action space:

[0068]

[0069] in, Control actions for energy management strategies; are the minimum and maximum limits of fuel cell power respectively;

[0070] And design the following reward function:

[0071]

[0072] Among them, r t is the reward value at the current moment; SOC t The current state of charge of the power battery; SOC refis the reference target state of charge; ω1 and ω2 are the weight coefficients of each item in the reward function.

[0073] It should be understood that the size of the serial numbers of the steps in the embodiment of the present invention does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0074] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A fuel cell vehicle energy management strategy integrating traffic perception, comprising the following steps: Step 1: Obtain traffic perception information through on-board sensors and vehicle network communication equipment, collect the speed and relative distance of at least six adjacent vehicles around the main vehicle in real time, and construct a traffic state input vector; Step 2: Build a speed prediction model based on the Transformer algorithm, which includes several sub-models. The traffic state input vector obtained in Step 1 is used as the model input, and the predicted future speed sequence of the main vehicle is used as the model output. Each sub-model is trained independently. A deep ensemble method is used to calculate the mean and variance of the future speed sequence predictions obtained by each sub-model, and the uncertainty quantification index of the model prediction is determined from this. Step 3: Set corresponding thresholds for the uncertainty quantification indicators and compare the two to determine whether the current prediction results are credible. If the results are credible, the prediction results of the speed prediction model are directly output. If the results are not credible, exponential smoothing prediction is performed to output credible prediction results. Step 4: Define the system state space based on the future speed sequence and the vehicle energy state, define the action space based on the fuel cell system power, and design and define the corresponding reward function taking into account the energy consumption of the fuel cell and power battery. On this basis, build an energy management optimizer based on deep reinforcement learning. It can select the corresponding fuel cell power control action based on the predicted future speed sequence and the current vehicle energy state, obtain the corresponding reward, and enter the next state; Step 5: Use historical operating conditions to train the energy management optimizer and use the corresponding action-reward evaluation mechanism to update the optimizer network parameters so that the output fuel cell control instructions tend to be optimal; Step 6: Deploy the trained energy management optimizer for online application.

2. The method according to claim 1, wherein: The traffic state input vector x constructed in step 1 t The specific form is as follows: x t =[v0,v1,…,v6,d1,…,d6] Where, v0 is the current speed of the main vehicle; v i (i=1,…,6) is the current speed of the i-th adjacent vehicle; d i (i=1,…,6) is the relative distance between the i-th adjacent vehicle and the main vehicle.

3. The method according to claim 1, wherein: The speed prediction model established in step 2 specifically uses the following multi-head attention mechanism to dynamically process traffic scene information: Q=XW Q K=XW K V=XW V X=[x1,x2,...,x T ] Where Q is the query vector matrix; K is the key vector matrix; V is the value vector matrix; W Q 、W K 、W V All are trainable parameters; X is the input sequence, specifically the traffic state input vector x of T time steps t composition; Used to measure the similarity between each time step; Used for scaling, avoiding dimension values ​​d k Too large leads to unstable gradients; the result obtained after performing softmax normalization is used as the weight, and the final representation of each time step is obtained from it.

4. The method according to claim 1, wherein: In step 2, the deep integration method is used to calculate the mean and variance of the speed series predicted by the sub-model using the following formula: Where N is the number of sub-models; is the velocity prediction result of the nth sub-model at time t+h; μ h is the mean of the prediction results at time t+h; is the prediction variance, H is the prediction time step window length, and U is the uncertainty quantification indicator.

5. The method according to claim 1, wherein: In step 3, the threshold δ of the uncertainty quantification index is set specifically t The form is as follows: in, is the historical uncertainty mean from tk to time t; β is the empirically set proportional coefficient used to adjust the prediction tolerance; k is the length of the historical sliding window; The uncertainty quantification index is compared with the threshold δ t For comparison, if U≤δ t The current prediction result is determined to be credible, and the future speed sequence prediction result of the speed prediction model is directly output; if U>δ t The current prediction result is determined to be unreliable, and the following exponential smoothing is performed to output the future speed: in, is the future speed predicted by exponential smoothing; v t is the current speed of the main vehicle; α∈(0,1) is the smoothing coefficient, which is used to control the weight of the current speed and historical trend.

6. The method according to claim 1, wherein: In step 4, the following state space is specifically defined: in, is the predicted future speed series; a t is the current acceleration; P fc,t is the power output of the current fuel cell system; αP fc,t =P fc,t -P fc,t-1 is the fuel cell power change rate; SOC t The current state of charge of the power battery; Define the following action space: in, Control actions for energy management strategies; are the minimum and maximum limits of fuel cell power respectively; And design the following reward function: Among them, r t is the reward value at the current moment; SOC t The current state of charge of the power battery; SOC ref is the reference target state of charge; ω1 and ω2 are the weight coefficients of each item in the reward function.

Images