Intelligent network connection electric vehicle speed optimization method considering side vehicle cut-in scene

Abstract

The invention relates to the technical field of intelligent networked electric vehicles, in particular to an intelligent networked electric vehicle speed optimization method considering a side vehicle cut-in scene, which comprises the following steps: 1) obtaining historical tracks of each side vehicle and surrounding vehicles corresponding to the side vehicle; 2) designing a hierarchical control strategy of side vehicle lane changing track prediction and economic vehicle speed optimization; 3) obtaining space-time interaction characteristics; 4) decoding the space-time interaction features, and predicting the future lane changing trajectory of the side vehicle; 5) constructing a multi-constraint optimization problem of vehicle speed optimization in a side vehicle cut-in scene; 6) predicting a lane changing intention and a lane changing track of the side vehicle and a motion conflict scene under the cut-in scene; and 7) generating an economic vehicle speed curve of the vehicle by adopting a QP method. According to the method, the lane changing track of the adjacent side vehicle can be predicted more accurately, the optimal solution can be found quickly, the future lane changing track of the adjacent side vehicle can be processed in real time, and the vehicle speed can be optimized in real time; the energy consumption of the BEV in the driving process is effectively reduced, the purpose of energy-saving driving is achieved, and the overall energy efficiency of the vehicle is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent connected electric vehicles, and in particular to a speed optimization method for an intelligent connected electric vehicle taking into account a scenario where a vehicle cuts in from a nearby vehicle. Background Art

[0002] Electric Connected and Automated Vehicles (ECAVs) offer significant advantages in reducing energy consumption compared to traditional fuel vehicles and hybrid electric vehicles (HEVs). This is primarily due to their efficient electric drive systems and regenerative braking, which further improves energy efficiency. Currently, the penetration rate of connected and autonomous vehicles (CAVs) is low, and the future traffic environment will likely be a mixed environment consisting of both human-driven vehicles (HDVs) and CAVs. Because HDVs' complex driving behaviors are often difficult to predict, it is crucial to fully consider their uncertainties when implementing energy-efficient CAV driving control.

[0003] The control strategies of CAVs that consider the random effects of HDV motion can be divided into two categories: passive response and active control. The traditional method adopts a passive response strategy, that is, after the HDV completes the cut-in behavior, the CAV responds to the HDV's uncertain driving behavior through real-time economic speed optimization. This passive response control strategy has the inherent defect of delayed response and is prone to large speed fluctuations, which affects the vehicle's energy economy and comfort, and even endangers driving safety. Active control strategies adaptively control the longitudinal speed of the vehicle by predicting the longitudinal speed changes of the preceding vehicle. However, there is a lack of prediction for the sudden cut-in scenario of the vehicle next to the adjacent lane. The vehicle will experience large speed fluctuations to avoid collision with the cutting-in vehicle, making it difficult to predict the impact of the HDV's lateral motion disturbance on the vehicle's speed.

[0004] At the level of vehicle speed optimization algorithms, existing methods can be mainly divided into three technical routes: rule-based, optimization-based, and machine learning-based methods. Among them, the rule-based method is limited by the rigid constraints of the preset logic, and it is difficult to achieve deep integration and optimization of external traffic information. Although optimization-based methods (such as dynamic programming DP, etc.) can effectively integrate multi-source traffic information for global optimization, the excessive computational load caused by the complexity of their algorithms restricts the application of real-time control. The machine learning-based method model is highly dependent on the quality and coverage of the training data, and will also fall into the "model complexity-computational efficiency" trade-off dilemma. It is worth noting that the quadratic programming (QP) method, while maintaining the information fusion capability of the optimization algorithm, can quickly obtain the optimal solution under limited computing resources by virtue of its convex optimization characteristics and efficient solution efficiency, providing a feasible solution for real-time economic vehicle speed optimization. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a speed optimization method for an intelligent connected electric vehicle considering the scenario of a vehicle cutting in from a nearby vehicle.

[0006] The technical problem to be solved by the present invention is achieved by adopting the following technical solutions:

[0007] A method for optimizing the speed of an intelligent connected electric vehicle considering a side vehicle cut-in scenario includes the following steps:

[0008] 1) Using onboard sensors, the system collects the speed, acceleration, and position information of surrounding vehicles, thereby obtaining the historical trajectory of each adjacent vehicle and its corresponding surrounding vehicles;

[0009] 2) Design a hierarchical control strategy for lane-changing trajectory prediction and economic speed optimization for adjacent vehicles;

[0010] 3) The Transformer encoder module encodes the historical trajectories of adjacent vehicles and vehicles around them, and uses the attention mechanism to capture the interaction relationship between vehicles to obtain spatiotemporal interaction features;

[0011] 4) The Transformer decoder decodes the spatiotemporal interaction features and predicts the future lane-changing trajectory of the adjacent vehicle through linear transformation and softmax layer;

[0012] 5) Using motor torque as the optimization variable and minimizing motor energy consumption as the objective function, and considering both vehicle dynamics and road environment constraints, a multi-constraint optimization problem for vehicle speed optimization in a vehicle-on-road scenario is constructed.

[0013] 6) Using the hierarchical control strategy designed in step 2), the lane-changing intention and trajectory of the adjacent vehicle are predicted based on the spatiotemporal interaction characteristics, and the motion conflict scenario in the cut-in scenario is predicted;

[0014] 7) Based on the hierarchical control strategy designed in step 2), and according to the conflict scenario predicted in step 6), the economic speed curve of this vehicle is generated using the QP method.

[0015] As a further improvement of the present invention, the hierarchical control strategy in step 2) includes an upper layer and a lower layer, the upper layer is a lane-changing trajectory prediction layer for adjacent vehicles, and the lower layer is an economical vehicle speed optimization layer.

[0016] As a further improvement of the present invention, the specific process of step 3) is as follows:

[0017] 3-1) For the historical trajectories of the adjacent vehicle and the vehicles around it, denoted as G, their positions are encoded as follows:

[0018] G={E, 0|E=(x0, y0), 0=(Δx i , Δy i )}

[0019] P=Position(G)

[0020] Among them, (x0, y0) represents the position information of the adjacent vehicle; (Δx i , Δy i ) represents the relative distance between the adjacent vehicle i and the own vehicle on the horizontal and vertical axes;

[0021] The historical trajectories of the adjacent vehicles and the vehicles around them are calculated by the position encoding formula and input into the trajectory information P of the adjacent vehicles and the vehicles around them required by the next spatial interaction layer. The formula is as follows:

[0022]

[0023] Among them, pos represents the position index in the sequence; d m Represents the dimension of the data; i represents the i-th dimension among all dimensions;

[0024] 3-2) The data obtained in step 3-1) is first used through the spatial interaction layer to extract information about the impact of the surrounding vehicles on the adjacent vehicle's trajectory. This is done by quantifying the differential impact of the adjacent vehicles on the adjacent vehicle's trajectory based on the changes in the relative distance between the adjacent vehicle and the surrounding vehicles. The specific formula is as follows:

[0025]

[0026] Where Δx i , Δy i Indicates the relative distance between vehicle i and the adjacent vehicle on the horizontal and vertical axes; the relative distance change ΔD i The expression is as follows:

[0027] ΔD i =D i，t-D i，t+1

[0028] Among them, D i，t Denotes the relative distance at time t, D i，t+1 represents the relative distance at time t+1;

[0029] Based on the change in the relative distance between the surrounding vehicles and the adjacent vehicles, the weight factor α of the trajectory influence is dynamically allocated i , as follows:

[0030]

[0031] Among them, position i is the real vehicle time p i =1, when position i is a virtual vehicle p i =0; the total weight factor A i Perform Hadamard multiplication with the trajectory information P of the adjacent vehicle and the surrounding vehicles to obtain the vehicle trajectory information P containing the vehicle spatial interaction information 1 , as follows:

[0032] A i ={α0,α i}

[0033] P 1 =P⊙A i

[0034] Among them, α0 represents the weight factor of the vehicle. After the trajectory information of the adjacent vehicle and the vehicles around the adjacent vehicle is processed through the spatial interaction layer, each can obtain its own output result P 1 ;

[0035] 3-3) Then enter the time interaction layer, integrating the temporal interaction features of the neighboring vehicle's own historical trajectory, as follows:

[0036]

[0037] V m，n =β m，n V m，n

[0038]

[0039] in, Indicates t m The trajectory information that has not passed the time interaction layer at time W Q , W K , W V Represents three weight matrices; Q represents the query vector, K represents the key vector, V represents the value vector, Q m , K m , V m Represented as tm Three vectors at the moment; β is the similarity factor, β m.n Indicates time t n Trajectory information at time t m Similarity factor of trajectory information; represents the scaling factor; V m，n Indicates time t m The value vector of time t n Trajectory information influence; P m Indicates time t m The trajectory information of is affected by the trajectory information of all moments;

[0040] 3-4) The data undergoes residual calculation and normalization, then passes through the feedforward layer and performs residual calculation and normalization again, as follows:

[0041] P′=P st +P

[0042]

[0043] Where μ is the mean of P′, σ is the standard deviation of P′, γ and δ are adjustable parameters during training, and ò is a constant. The feedforward layer contains two fully connected layers, one for residual calculation and normalization. The fully connected layer is used to learn the complex relationship between input features and extract higher-level abstract features, as follows:

[0044]

[0045] Among them, ω1 and are the weight matrix and bias vector of the first linear transformation, and finally the trajectory information P of the adjacent vehicle and the surrounding vehicles containing spatiotemporal interaction features is obtained. f .

[0046] As a further improvement of the present invention, the specific process of step 4) is as follows:

[0047] 4-1) First, position decoding is performed, and then input into the mask layer for residual calculation and normalization, as follows:

[0048] F′=Masked(F)+F

[0049]

[0050] Among them, F is the data after position encoding, and Masked(F) is the data after mask processing;

[0051] 4-2) In the multi-head attention layer of the decoding layer, the encoder output P f The operation is performed with the decoder's mask layer output F″ as follows:

[0052]

[0053] Q f =f”W Q , K p =p f W K , V P =P f W V

[0054] f sa =Self-Attention(Q f , K p , V p )

[0055] F sa =contact(f sa )

[0056] Among them, p f and f″ respectively represent P f The data obtained by splitting F and F″ into multiple heads, head represents the number of heads, Self-Attention() represents the calculation of the self-attention mechanism, and contact() represents splicing the data fragments together;

[0057] 4-3) At the end of the decoding layer, another feedforward layer calculation, residual calculation and normalization are performed as follows:

[0058]

[0059] F′ sa =FFN(F sa )+F sa

[0060]

[0061] Where μ is F sa ', δ is the mean of F sa 'Standard deviation, ω1 and are the weight matrix and bias vector of the first linear transformation, F sa ” is the final output of the decoding layer;

[0062] 4-4) Finally, a linear transformation is performed on the output of the decoding layer, and the vehicle's future lane-changing trajectory is obtained from the maximum probability output layer, as follows:

[0063] F l =WF″ sa +b

[0064]

[0065] F f =select(F l |MAX(P F ))

[0066] Among them, W is the weight matrix, b is the bias term, F l is the vector obtained by linear transformation, f n F l Data at any time, N is the prediction step length, F f The future lane-changing trajectory of the vehicle.

[0067] As a further improvement of the present invention, the motor energy consumption calculation formula in step 5) is as follows:

[0068]

[0069] Among them, E m represents the energy consumption of the profile motor in discrete time Δt, T m To express the output torque of the profile motor, n m Indicates the speed of the profile motor, η m Expressed as the efficiency of the profile motor, T m With n m Obtained by looking up the table.

[0070] As a further improvement of the present invention, the objective function of the speed planning in step 5) is:

[0071]

[0072] Among them, Q is a positive definite matrix, which represents the quadratic relationship between energy consumption and torque.

[0073] As a further improvement of the present invention, the constraint conditions in step 5) are:

[0074] n m =30ω m / π

[0075] η m =f(T m , n m )

[0076] T min ≤T m ≤T max

[0077] (ΔT m ) min <ΔT m <(ΔT m ) max

[0078] 0≤v≤ν lim

[0079] a lim1 ≤a≤a lim2

[0080]

[0081] Δy m ≥y safe

[0082] Δx m ≥x safe

[0083] Where, ΔT m is the torque increment of the profile motor, (ΔT m ) min and (ΔT m ) max is the torque increment amplitude of the profile motor, v lim is the road speed limit, x m and y m Before the adjacent vehicle changes to the lane of the vehicle, m The horizontal and vertical coordinates of the time, Δx m and Δy m is the lateral and longitudinal distance between the adjacent vehicle and the vehicle after lane change, x 0，m and y 0，m For the car at t m The horizontal and vertical coordinates of the time, x safe and y safe are the lateral and longitudinal safety thresholds between vehicles.

[0084] As a further improvement of the present invention, the specific process of step 7) is as follows:

[0085] The objective function and constraints in step 5) are organized into a standard QP form, that is, the objective function is converted into a quadratic form, and the constraints in step 5) are used as constraints, as follows:

[0086]

[0087] Among them, c is a linear coefficient vector, which represents the linear relationship between energy consumption and torque. Then, the MATLAB QP solver quadprog is used to solve the optimized motor torque, thereby obtaining the speed optimization curve.

[0088] The beneficial effects of the present invention are:

[0089] The proposed Transformer-based deep learning method predicts lane-changing trajectories of adjacent vehicles. The Transformer's self-attention mechanism can capture the long-range influence of adjacent vehicle trajectories on their historical state and surrounding environment. Furthermore, the multi-head attention mechanism can simultaneously process the trajectories of multiple vehicles, capturing the interactions between them and achieving more accurate lane-changing trajectory prediction for adjacent vehicles.

[0090] Optimize the vehicle speed curve based on QP. QP can handle constraints such as motor torque and its increment constraints, acceleration constraints, road speed limits, and surrounding vehicle trajectories. It has high computational efficiency and can quickly find the optimal solution. It can also process the future lane-changing trajectories of adjacent vehicles in real time and optimize the vehicle speed in real time.

[0091] The proposed speed curve optimization based on the prediction of the future lane-changing trajectory of adjacent vehicles enables the BEV to obtain the driving intention of adjacent vehicles in advance by predicting their lane-changing trajectories in real time, and optimize its own speed curve accordingly, effectively reducing the energy consumption of the BEV during driving, achieving the goal of energy-saving driving, and thus improving the overall energy efficiency of the vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0093] Figure 1 Flowchart of the present invention;

[0094] Figure 2 Schematic diagram of the hierarchical control strategy for lane-changing trajectory prediction and economical speed optimization of adjacent vehicles;

[0095] Figure 3 Flowchart for predicting the future lane-changing trajectory sequence of adjacent vehicles;

[0096] Figure 4 Schematic diagram of vehicle speed optimization. DETAILED DESCRIPTION

[0097] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0098] like Figure 1 As shown, a speed optimization method for an intelligent connected electric vehicle considering a side vehicle cutting-in scenario includes the following steps:

[0099] 1) Through the on-board sensor equipment, the speed, acceleration, and position information of surrounding vehicles are collected to obtain the historical trajectory of each adjacent vehicle and the surrounding vehicles corresponding to the adjacent vehicle.

[0100] 2) Design a hierarchical control strategy for lane-changing trajectory prediction of adjacent vehicles and economical speed optimization. The upper layer is the lane-changing trajectory prediction layer of adjacent vehicles, and the lower layer is the economical speed optimization layer, such as Figure 2 shown.

[0101] 3) If Figure 3 The Transformer encoder module encodes the historical trajectories of adjacent vehicles and vehicles around them, and uses the attention mechanism to capture the interaction relationship between vehicles to obtain spatiotemporal interaction features.

[0102] The specific process is as follows:

[0103] 3-1) For the historical trajectories of the adjacent vehicle and the vehicles around it, denoted as G, their positions are encoded as follows:

[0104] G={E, 0|E=(x0, y0), 0=(Δx i , Δy i )}

[0105] P=Position(G)

[0106] Among them, (x0, y0) represents the position information of the adjacent vehicle; (Δx i , Δy i ) represents the relative distance between the adjacent vehicle i and the own vehicle on the horizontal and vertical axes;

[0107] The historical trajectories of the adjacent vehicles and the vehicles around them are calculated by the position encoding formula and input into the trajectory information P of the adjacent vehicles and the vehicles around them required by the next spatial interaction layer. The formula is as follows:

[0108]

[0109] Among them, pos represents the position index in the sequence; d m Represents the dimension of the data; i represents the i-th dimension among all dimensions;

[0110] 3-2) The data obtained in step 3-1) is first used through the spatial interaction layer to extract information about the impact of the surrounding vehicles on the adjacent vehicle's trajectory. This is done by quantifying the differential impact of the adjacent vehicles on the adjacent vehicle's trajectory based on the changes in the relative distance between the adjacent vehicle and the surrounding vehicles. The specific formula is as follows:

[0111]

[0112] Where Δx i , Δy i Indicates the relative distance between vehicle i and the adjacent vehicle on the horizontal and vertical axes; the relative distance change ΔD i The expression is as follows:

[0113] ΔD i =D i，t -D i，t+1

[0114] Among them, D i，t Denotes the relative distance at time t, D i，t+1 represents the relative distance at time t+1;

[0115] Based on the change in the relative distance between the surrounding vehicles and the adjacent vehicles, the weight factor α of the trajectory influence is dynamically allocated i , as follows:

[0116]

[0117] Among them, position i is the real vehicle time p i =1, when position i is a virtual vehicle p i =0; the total weight factor A i Perform Hadamard multiplication with the trajectory information P of the adjacent vehicle and the surrounding vehicles to obtain the vehicle trajectory information P containing the vehicle spatial interaction information 1 , as follows:

[0118] A i ={α0,α i}

[0119] P 1 =P⊙A i

[0120] Among them, α0 represents the weight factor of the vehicle (1), and the trajectory information of the adjacent vehicle and the vehicles around the adjacent vehicle are processed through the spatial interaction layer to obtain their own output results P 1 ;

[0121] 3-3) Then enter the time interaction layer, integrating the temporal interaction features of the neighboring vehicle's own historical trajectory, as follows:

[0122]

[0123] V m，n =β m，n V m，n

[0124]

[0125] in, Indicates t m The trajectory information that has not passed the time interaction layer at time W Q , W K , W V Represents three weight matrices; Q represents the query vector, K represents the key vector, and V represents the value vector (Q m, K m , V m Represented as t m Three vectors at the moment), Q m , K m , V m Represented as t m Three vectors at the moment; β is the similarity factor, β m.n Indicates time t n Trajectory information at time t m Similarity factor of trajectory information; Represents the scaling factor (controls the weight, usually 8); V m，n Indicates time t m The value vector of time t n Trajectory information influence; P m Indicates time t m The trajectory information of is affected by the trajectory information of all moments;

[0126] 3-4) The data undergoes residual calculation and normalization, then passes through the feedforward layer and performs residual calculation and normalization again, as follows:

[0127] P′=P st +P

[0128]

[0129] Where μ is the mean of P′, σ is the standard deviation of P′, γ and δ are adjustable parameters during training, and ò is a constant. The feedforward layer contains two fully connected layers, one for residual calculation and normalization. The fully connected layer is used to learn the complex relationship between input features and extract higher-level abstract features, as follows:

[0130]

[0131] Among them, ω1 and are the weight matrix and bias vector of the first linear transformation, and finally the trajectory information P of the adjacent vehicle and the surrounding vehicles containing spatiotemporal interaction features is obtained. f .

[0132] 4) The spatiotemporal interaction features are decoded through the Transformer decoder, and the future lane-changing trajectory of the adjacent vehicle is predicted through linear transformation and Softmax layer.

[0133] Step 4) The specific process is as follows:

[0134] 4-1) First, position decoding is performed, and then input into the mask layer for residual calculation and normalization, as follows:

[0135] F′=Masked(F)+F

[0136]

[0137] Among them, F is the data after position encoding, and Masked(F) is the data after mask processing;

[0138] 4-2) In the multi-head attention layer of the decoding layer, the encoder output P f The operation is performed with the decoder's mask layer output F″ as follows:

[0139]

[0140] Q f =f”W Q , K p =p f W K , V P =P f W V

[0141] f sa =Self-Attention(Q f , K p , V p )

[0142] F sa =contact(f sa )

[0143] Among them, p f and f″ respectively represent P f The data obtained by splitting F and F″ into multiple heads, head represents the number of heads, Self-Attention() represents the calculation of the self-attention mechanism, and contact() represents splicing the data fragments together;

[0144] 4-3) At the end of the decoding layer, another feedforward layer calculation, residual calculation and normalization are performed as follows:

[0145]

[0146] F′ sa =FFN(F sa )+F sa

[0147]

[0148] Where μ is F sa ', δ is the mean of F sa ' standard deviation, ω1 and are the weight matrix and bias vector of the first linear transformation, Fsa ” is the final output of the decoding layer;

[0149] 4-4) Finally, a linear transformation is performed on the output of the decoding layer, and the vehicle's future lane-changing trajectory is obtained from the maximum probability output layer, as follows:

[0150] F l =WF″ sa +b

[0151]

[0152] F f =select(F l |MAX(P F ))

[0153] Among them, W is the weight matrix, b is the bias term, F l is the vector obtained by linear transformation, f n F l Data at any time, N is the prediction step length, F f The future lane-changing trajectory of the vehicle.

[0154] 5) Taking motor torque as the optimization variable and minimizing motor energy consumption as the objective function, and comprehensively considering vehicle dynamics and road environment constraints, a multi-constraint optimization problem for vehicle speed optimization in the scenario of cutting in with a nearby vehicle is constructed.

[0155] The motor energy consumption calculation formula in step 5) is as follows:

[0156]

[0157] Among them, E m represents the energy consumption of the profile motor in discrete time Δt, T m To express the output torque of the profile motor, n m Indicates the speed of the profile motor, η m Expressed as the efficiency of the profile motor, T m With n m Obtained by looking up the table.

[0158] The objective function of speed planning in step 5) is:

[0159]

[0160] Among them, Q is a positive definite matrix, which represents the quadratic relationship between energy consumption and torque.

[0161] Constraints in step 5):

[0162] n m =30ω m / π

[0163] η m =f(T m , n m )

[0164] T min ≤T m ≤T max

[0165] (ΔT m ) min <ΔT m <(ΔT m ) max

[0166] 0≤v≤v lim

[0167] a lim1 ≤a≤a lim2

[0168]

[0169] Δy m ≥y safe

[0170] Δx m ≥x safe

[0171] Where, ΔT m is the torque increment of the profile motor, (ΔT m ) min and (ΔT m ) max is the torque increment amplitude of the profile motor, ν lim is the road speed limit, x m and y m Before the adjacent vehicle changes to the lane of the vehicle, m The horizontal and vertical coordinates of the time, Δx m and Δy m is the lateral and longitudinal distance between the adjacent vehicle and the vehicle after lane change, x 0，m and y 0，m For the car at t m The horizontal and vertical coordinates of the time, x safe and y safe are the lateral and longitudinal safety thresholds between vehicles.

[0172] 6) Through the hierarchical control strategy designed in step 2), the lane-changing intention and lane-changing trajectory of the adjacent vehicle are predicted based on the spatiotemporal interaction characteristics, and the motion conflict scenario in the cut-in scenario is predicted.

[0173] 7) Based on the hierarchical control strategy designed in step 2), and according to the conflict scenario predicted in step 6), the economic speed curve of this vehicle is generated using the QP method.

[0174] Step 7) The specific process is:

[0175] The objective function and constraints in step 5) are organized into a standard QP form, that is, the objective function is converted into a quadratic form, and the constraints in step 5) are used as constraints, as follows:

[0176]

[0177] Among them, c is a linear coefficient vector, which represents the linear relationship between energy consumption and torque. Then, the MATLAB QP solver quadprog is used to solve the optimized motor torque, thereby obtaining the speed optimization curve.

[0178] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and description merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A speed optimization method for an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario, characterized by: The following steps are involved: 1) Using onboard sensors, the system collects the speed, acceleration, and position information of surrounding vehicles, thereby obtaining the historical trajectory of each adjacent vehicle and its corresponding surrounding vehicles; 2) Design a hierarchical control strategy for lane-changing trajectory prediction and economic speed optimization for adjacent vehicles; 3) The Transformer encoder module encodes the historical trajectories of adjacent vehicles and vehicles around them, and uses the attention mechanism to capture the interaction relationship between vehicles to obtain spatiotemporal interaction features; 4) The Transformer decoder decodes the spatiotemporal interaction features and predicts the future lane-changing trajectory of the adjacent vehicle through linear transformation and softmax layer; 5) Using motor torque as the optimization variable and minimizing motor energy consumption as the objective function, and considering both vehicle dynamics and road environment constraints, a multi-constraint optimization problem for vehicle speed optimization in a vehicle-on-road scenario is constructed. 6) Using the hierarchical control strategy designed in step 2), the lane-changing intention and trajectory of the adjacent vehicle are predicted based on the spatiotemporal interaction characteristics, and the motion conflict scenario in the cut-in scenario is predicted; 7) Based on the hierarchical control strategy designed in step 2), and according to the conflict scenario predicted in step 6), the economic speed curve of this vehicle is generated using the QP method.

2. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: The hierarchical control strategy in step 2) includes two layers: the upper layer is the lane-changing trajectory prediction layer for adjacent vehicles, and the lower layer is the economical speed optimization layer.

3. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: Step 3) The specific process is as follows: 3-1) For the historical trajectories of the adjacent vehicle and the vehicles around it, denoted as G, their positions are encoded as follows: G={E,0∣E=(x0,y0),0=(Δx i ,Δy i )} P=Position(G) Among them, (x0, y0) represents the position information of the adjacent vehicle; (Δx i ,Δy i ) represents the relative distance between the adjacent vehicle i and the own vehicle on the horizontal and vertical axes; The historical trajectories of the adjacent vehicles and the vehicles around them are calculated by the position encoding formula and input into the trajectory information P of the adjacent vehicles and the vehicles around them required by the next spatial interaction layer. The formula is as follows: Among them, pos represents the position index in the sequence; d m Represents the dimension of the data; i represents the i-th dimension among all dimensions; 3-2) The data obtained in step 3-1) is first used through the spatial interaction layer to extract information about the impact of the surrounding vehicles on the adjacent vehicle's trajectory. This is done by quantifying the differential impact of the adjacent vehicles on the adjacent vehicle's trajectory based on the changes in the relative distance between the adjacent vehicle and the surrounding vehicles. The specific formula is as follows: Where Δx i , Δy i Indicates the relative distance between vehicle i and the adjacent vehicle on the horizontal and vertical axes; the relative distance change ΔD i The expression is as follows: ΔD i =D i,t -D i,t+1 Among them, D i,t Denotes the relative distance at time t, D i,t+1 represents the relative distance at time t+1; Based on the change in the relative distance between the surrounding vehicles and the adjacent vehicles, the weight factor α of the trajectory influence is dynamically allocated i , as follows: Among them, position i is the real vehicle time p i =1, when position i is a virtual vehicle p i =0; the total weight factor A i Perform Hadamard multiplication with the trajectory information P of the adjacent vehicle and the surrounding vehicles to obtain the vehicle trajectory information P containing the vehicle spatial interaction information 1 , as follows: A i ＝{α0,α i} P 1 =P⊙A i Among them, α0 represents the weight factor of the vehicle. After the trajectory information of the adjacent vehicle and the vehicles around the adjacent vehicle is processed through the spatial interaction layer, each can obtain its own output result P 1 ; 3-3) Then enter the time interaction layer, integrating the temporal interaction features of the neighboring vehicle's own historical trajectory, as follows: V m,n =β m,n V m,n in, Indicates t m The trajectory information that has not passed the time interaction layer at time W Q ,W K ,W V Represents three weight matrices; Q represents the query vector, K represents the key vector, V represents the value vector, Q m ,K m ,V m Represented as t m Three vectors at the moment; β is the similarity factor, β m.n Indicates time t n Trajectory information at time t m Similarity factor of trajectory information; represents the scaling factor; V m,n Indicates time t m The value vector of time t n Trajectory information influence; P m Indicates time t m The trajectory information of is affected by the trajectory information of all moments; 3-4) The data undergoes residual calculation and normalization, then passes through the feedforward layer and performs residual calculation and normalization again, as follows: P′=P st +P Where μ is the mean of P′, σ is the standard deviation of P′, γ and δ are adjustable parameters during training, and ò is a constant. The feedforward layer contains two fully connected layers, one for residual calculation and normalization. The fully connected layer is used to learn the complex relationship between input features and extract higher-level abstract features, as follows: Among them, ω1 and are the weight matrix and bias vector of the first linear transformation, and finally the trajectory information P of the adjacent vehicle and the surrounding vehicles containing spatiotemporal interaction features is obtained. f .

4. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: Step 4) The specific process is as follows: 4-1) First, position decoding is performed, and then input into the mask layer for residual calculation and normalization, as follows: F′=Masked(F)+F Among them, F is the data after position encoding, and Masked(F) is the data after mask processing; 4-2) In the multi-head attention layer of the decoding layer, the encoder output P f The operation is performed with the decoder's mask layer output F″ as follows: Q f =f”W Q ,K p =p f W K ,V P =P f W V f sa =Self-Attention(Q f ,K p ,V p ) F sa =contact(f sa ) Among them, p f and f″ respectively represent P f The data obtained by splitting F and F″ into multiple heads, head represents the number of heads, Self-Attention() represents the calculation of the self-attention mechanism, and contact() represents splicing the data fragments together; 4-3) At the end of the decoding layer, another feedforward layer calculation, residual calculation and normalization are performed as follows: F′ sa =FFN(F sa ) + F sa Where μ is F sa ', δ is the mean of F sa 'Standard deviation, ω1 and are the weight matrix and bias vector of the first linear transformation, F sa ” is the final output of the decoding layer; 4-4) Finally, a linear transformation is performed on the output of the decoding layer, and the vehicle's future lane-changing trajectory is obtained from the maximum probability output layer, as follows: F l =WF s ” a +b F f =select(F l |MAX(P F )) Among them, W is the weight matrix, b is the bias term, F l is the vector obtained by linear transformation, f n F l Data at any time, N is the prediction step length, F f The future lane-changing trajectory of the vehicle.

5. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: The motor energy consumption calculation formula in step 5) is as follows: Among them, E m represents the energy consumption of the profile motor in discrete time Δt, T m To express the output torque of the profile motor, n m Indicates the speed of the profile motor, η m Expressed as the efficiency of the profile motor, T m With n m Obtained by looking up the table.

6. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: The objective function of speed planning in step 5) is: Among them, Q is a positive definite matrix, which represents the quadratic relationship between energy consumption and torque.

7. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: Constraints in step 5): n m =30h m / p or m =f(T m ,n m ) T min ≤T m ≤T maX (ΔT m ) min <ΔT m <(ΔT m ) max 0≤v≤v lim a lim1 ≤a≤a lim2 Δy m ≥y safe Δx m ≥x safe Where, ΔT m is the torque increment of the profile motor, (ΔT m ) min and (ΔT m ) max is the torque increment amplitude of the profile motor, ν lim is the road speed limit, x m and y m Before the adjacent vehicle changes to the lane of the vehicle, m The horizontal and vertical coordinates of the time, Δx m and Δy m is the lateral and longitudinal distance between the adjacent vehicle and the vehicle after lane change, x 0,m and y 0,m For the car at t m The horizontal and vertical coordinates of the time, x safe and y safe are the lateral and longitudinal safety thresholds between vehicles.

8. The method for optimizing the speed of an intelligent connected electric vehicle considering a vehicle-on-the-side cut-in scenario according to claim 1, characterized in that: Step 7) The specific process is: The objective function and constraints in step 5) are organized into a standard QP form, that is, the objective function is converted into a quadratic form, and the constraints in step 5) are used as constraints, as follows: Among them, c is a linear coefficient vector, which represents the linear relationship between energy consumption and torque. Then, the MATLAB QP solver quadprog is used to solve the optimized motor torque, thereby obtaining the speed optimization curve.

Images