A hybrid vehicle power domain control system and control method

Through the combination of power data acquisition and prediction module, scheduling learning and pattern recognition module, RNN and DQN algorithms and genetic algorithms, the problem of insufficient flexibility and response capabilities of the power domain control method of hybrid vehicles is solved, and dynamic driving is achieved Load optimization and energy management in the environment improve energy utilization efficiency and driving experience.

CN118753268BActive Publication Date: 2024-12-03LONGHUI (INNER MONGOLIA) INTELLIGENT TECHNOLOGY CO LTD

Patent Information

Application Number
CN202411050116.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-12-03
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

The existing hybrid vehicle power domain control methods lack flexibility and real-time response capabilities, making it difficult to achieve real-time optimized load allocation in dynamic driving environments, resulting in low energy utilization efficiency and low mode switching efficiency, and energy waste.

Method used

Power data acquisition module, dynamic load prediction module, scheduling learning module and pattern recognition module are adopted to predict load changes in real time through RNN and DQN algorithms and adjust engine and motor load distribution, combined with genetic algorithms to optimize global energy distribution, realize multi-mode switching and Energy management strategies are automatically adjusted.

Benefits of technology

It improves the energy management efficiency of hybrid cars in dynamic driving environments, dynamically adapts to load changes, optimizes the load distribution of engines and motors, reduces energy waste, and improves the overall energy utilization efficiency and driving experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118753268B_ABST
    Figure CN118753268B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of automobile power control technology, and in particular to a hybrid electric vehicle power domain control system and a control method, comprising a power data acquisition module, a dynamic load prediction module, a scheduling learning module, a pattern recognition module, and a multi-mode switching module; the present invention uses real-time sensor data and historical driving data to establish a dynamic load prediction model to predict load changes in a short period of time in the future; uses a recursive neural network to predict load requirements under different working conditions in real time; dynamically adjusts the load distribution of the engine and the motor; outputs an optimal scheduling strategy through reinforcement learning to balance the workload of each subsystem and optimize the overall energy efficiency; and provides multi-mode recognition and switching at the same time: uses a machine learning algorithm to identify the current driving mode in real time, and automatically switches the energy management strategy according to the mode change; optimizes the mode switching process, improves energy utilization efficiency, and reduces energy waste.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of automobile power control, and in particular to a hybrid electric vehicle power domain control system and a control method. Background Art

[0002] The power system of a hybrid vehicle includes multiple subsystems such as the engine, electric motor and battery.

[0003] Chinese patent publication number: CN112918461A discloses a parallel hybrid electric vehicle power domain control system, including an engine control module, a motor control module, a battery management module, a transmission control module, and an energy management module; by integrating the hybrid controller of the parallel hybrid electric vehicle with the transmission controller, the number of vehicle controllers is reduced, thereby improving the real-time performance of the vehicle power control system signal interaction and shortening the communication time; formulating corresponding economical shifting strategies for different hybrid electric vehicle working modes can significantly reduce the fuel consumption of the parallel hybrid electric vehicle.

[0004] For the power system of hybrid vehicles, the current power scheduling methods mainly include fixed rule scheduling and feedback control: fixed rule scheduling power distribution is difficult to achieve real-time optimization in the face of complex and dynamically changing driving environments. Under different load conditions, the preset rules are difficult to cope with emergencies, resulting in low energy utilization efficiency; and some feedback control strategies such as PID control have a slow response speed to real-time load changes, which will cause system response lag and affect driving experience and energy efficiency;

[0005] It can be seen that the above scheduling scheme lacks flexibility and real-time response capability for the power domain control mode of hybrid vehicles, and it is difficult to achieve real-time optimal distribution of loads in a dynamic driving environment; therefore, a hybrid vehicle power domain control system and control method are urgently needed to solve such problems. Summary of the invention

[0006] To this end, the present invention provides a hybrid electric vehicle power domain control system and control method to overcome the problems that the power domain control method of hybrid electric vehicles in the prior art lacks flexibility and real-time response capability, making it difficult to achieve real-time optimal distribution of loads in a dynamic driving environment; and it is impossible to achieve global optimal energy management between different driving modes, resulting in low mode switching efficiency and energy waste.

[0007] To achieve the above objectives, on the one hand, the present invention provides a hybrid electric vehicle power domain control system, comprising:

[0008] Power data acquisition module, used to obtain multi-dimensional power data during vehicle driving;

[0009] Dynamic load prediction module, which inputs sensor data into the RNN model in real time during vehicle driving to predict load demand under different working conditions;

[0010] A scheduling learning module to dynamically adjust the load distribution of the engine and the motor;

[0011] A mode recognition module for identifying the current driving mode;

[0012] The multi-mode switching module automatically switches to the corresponding energy management strategy based on the real-time identification of the driving mode.

[0013] On the other hand, the present invention also provides a hybrid electric vehicle power domain control method, comprising:

[0014] Step S1, vehicle power data collection, deploy sensors to obtain multi-dimensional power data of the vehicle during driving; integrate the vehicle's historical driving data as basic data for prediction and optimization; clean the collected data and extract key features; the multi-dimensional power data includes: speed, acceleration, battery power, and engine status;

[0015] Step S2, dynamic load prediction, uses the RNN model to predict the load changes in the short term in the future; inputs sensor data into the RNN model in real time during vehicle driving to predict the load demand under different working conditions;

[0016] Step S3, reinforcement learning scheduling, defining the state, action space and objective function, using the deep Q network DQN to train the scheduling strategy, and deploying it to the vehicle control system to dynamically adjust the load distribution of the engine and the motor;

[0017] Step S4, driving mode recognition, extracting driving mode related features from real-time sensor data; using the extracted features to train a pattern recognition model; using the recognition model to recognize the current driving mode in real time; related features include speed, road type, and traffic conditions;

[0018] Step S5, multi-mode optimization and switching, automatically switching to the corresponding energy management strategy according to the real-time identified driving mode; in each driving mode, using a genetic algorithm to optimize the global energy distribution;

[0019] Furthermore, the step of performing load prediction by using a recurrent neural network RNN ​​in step S2 includes:

[0020] Collect sensor data and historical driving data of the vehicle during driving, including but not limited to:

[0021] Vehicle speed V t , acceleration A t 、Engine load L eng,t , Motor load Lmot,t , Battery Capacity SOC t 、Ambient temperature T env,t , clean and normalize the data, and generate a time series data set;

[0022] Building the RNN model:

[0023] Input feature vector definition: Define the input feature vector X with a time step of t t :X t =[V t ,A t ,L eng,t ,L mot,t ,SOC t ,T env,t ];

[0024] Output target definition: Define the load forecast value for the next time step t+1 Among them, h t is the hidden state of RNN;

[0025] The model structure adopts the long short-term memory LSTM network structure, including:

[0026] Forget Gate: where f t is the forget gate vector, σ is the Sigmoid activation function, W f represents the weight matrix of the forget gate, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous time step;

[0027] Input Gate: where i t is the input gate vector, is the candidate cell state vector, W i , W C is the weight matrix of the input gate and cell state, b i 、b C is the bias vector for the input gate and cell state;

[0028] Cell status update: Among them C t is the cell state at the current time step, C t-1 is the cell state at the previous time step;

[0029] Output Gate: Among them t is the output gate vector, W o is the weight matrix of the output gate, b orepresents the bias vector of the output gate;

[0030] Hide status update:h t =o t tanh(C t ), where h t is the hidden state of the current time step;

[0031] Use mean square error (MSE) as the loss function to calculate the error between the predicted value and the actual load. Here n is the number of samples, Indicates the predicted load value, L t is the real load value;

[0032] Use the back-propagation algorithm to train the LSTM network through time BPTT;

[0033] Deploy the trained LSTM model to the vehicle control system and input sensor data in real time for load prediction;

[0034] During vehicle driving, the LSTM model is used to predict the load changes in the short term in the future in real time; load prediction is performed through a recurrent neural network (LSTM), and a dynamic load prediction model is established using real-time sensor data and historical driving data, effectively improving the energy management efficiency of hybrid vehicles under different working conditions, achieving higher fuel economy and driving experience;

[0035] Further, in step S3, data including the current speed, acceleration, load prediction value, and battery power of the vehicle are defined as a state space;

[0036] The data including the power distribution between the engine and the electric motor is defined as the action space;

[0037] Furthermore, in step S3, a reward function is constructed based on energy efficiency and driving experience, the goal is to maximize the total reward, and a deep Q network DQN is used to train the scheduling strategy to dynamically adjust the load distribution of the engine and the motor;

[0038] Furthermore, in step S3, the steps of using the deep Q network DQN algorithm to train the scheduling strategy include:

[0039] Define the environment and state, state space S: define the current state of the vehicle, expressed as: Where V t is the vehicle speed, A t is the acceleration, To predict load, SOC t is the battery capacity, T env,t is the ambient temperature, L eng,t is the current engine load, L mot,tis the current motor load;

[0040] Action space A: defined as the power distribution between the engine and the motor, expressed as: A t =[P eng,t ,P mot,t ], where P eng,t is the engine power, P mot,t is the motor power;

[0041] Instant Rewards t To optimize energy efficiency and driving experience, with the goal of minimizing fuel consumption and emissions: t =-(α1·FC t +α2·EM t -β·DE t ), where FC t Indicates fuel consumption, EM t Denotes emissions, DE t It represents the driving experience, expressed as the acceleration smoothness, α1, α2, β all represent the corresponding weight coefficients, and t is the time step;

[0042] Furthermore, in step S3, the deep Q network DQN training scheduling strategy is used as follows:

[0043] Q network structure: Use a deep neural network to estimate the state action value function, with the input being the state S t , the output is the Q value Q(S t ,A t ;θ): Here θ is the Q network parameter, θ′ represents the target Q network parameter, and γ is the discount factor;

[0044] Stores experience fragments of the agent's interaction with the environment (S t ,A t ,r t ,S t+1 ) to the experience replay pool, randomly extract small batches of samples for training, and break the sample correlation;

[0045] Calculate the target Q value y using a delayed updated target network t , stabilize the training process: Here θ - is the target Q network parameter, Q(S t+1 ,A t+1 θ - ) represents the Q value of the target Q network output, that is, the state S at the next moment t+1 Next take action A t+1 The maximum Q value of

[0046] Furthermore, the training process in step S3 using the deep Q network DQN training scheduling strategy is:

[0047] a. Initialization: Initialize the Q network parameters θ and the target Q network parameters θ′, set as copies of θ;

[0048] b. Collect experience: perform actions in the environment and collect (S t ,A t ,r t ,S t+1 ) and stored in the experience replay pool D;

[0049] c. Extract small batch samples: randomly extract small batch samples B from the experience replay pool;

[0050] d. Calculate the target Q value: For each sample, calculate the target Q value y t ,

[0051] e. Update Q network: minimize the loss function L(θ) and update the Q network parameters θ: Here, E represents the expected value, (S t ,A t ,r t ,S t+1 )~D represent small batch samples drawn from the experience replay pool D, each sample contains the current state S t 、Current Action A t , Instant Rewards t and the next state S t+1 , Q(S t ,A t ; θ) is the Q value of the current Q network output, indicating that in the current state S t Next take action A t Q value;

[0052] f. Periodically update the target network: copy the Q network parameters θ to the target network parameters θ′ at fixed steps;

[0053] g. Repeat the training process of steps af until the loss function converges or reaches the predetermined number of training rounds;

[0054] By using the deep Q network (DQN) algorithm to train the scheduling strategy, the load distribution of the engine and the motor can be dynamically adjusted to maximize energy utilization efficiency;

[0055] Furthermore, the global energy allocation optimization method is:

[0056] Initialize the population: generate an initial energy allocation strategy set;

[0057] Fitness evaluation: Calculate the fitness value of each strategy based on energy efficiency and driving experience;

[0058] Selection, crossover, and mutation: Generate new strategy combinations and iterate until convergence;

[0059] Achieve global optimal energy management between different driving modes, optimize the mode switching process, improve energy utilization efficiency, and reduce energy waste;

[0060] Furthermore, the global energy allocation optimization method in step S5 is:

[0061] 1. Generate an initial energy allocation strategy set. Each individual represents an energy allocation strategy. Suppose the population size is N, and each individual x i represents the power distribution strategy between the engine and the motor, x i =[P eng,i ,P mot,i ], P eng,i is the engine power, P mot,i is the motor power;

[0062] Randomly generate individuals in the population so that each individual satisfies the power constraint, P eng,i +P mot,i =P total , P total represents the total power demand;

[0063] 2. Define the fitness function to evaluate the quality of each individual's energy allocation strategy. The fitness function F(x i ): Where E eng,i is the engine energy output, E mot,i is the motor energy output, DE i Indicates driving experience evaluation, FC i is the fuel consumption, EM i is emission, η, β, α are weight coefficients, balancing energy efficiency, driving experience and emission;

[0064] 3. Use the roulette wheel selection method to select individuals as parents for crossover and mutation based on the fitness value. The selection probability is: Here p i That is, individual x i The probability of being selected, N represents the population size, and j represents the index of the jth individual;

[0065] 4. Use single-point crossover to combine two parent individuals to generate new individual offspring. The crossover formula is: Here x new,k is the newly generated individual, x parent,k ,x parent2,k For two parent individuals, Represents a single-point crossover operation;

[0066] 5. Use uniform mutation to randomly change part of the individual. The mutation formula is: x mut,k =x orig,k +δ, where x mut,k For the individual after mutation,

[0067] x orig,k represents the original individual, δ is a small random perturbation, satisfying x orig,k +δ≤P total ;

[0068] 6. Combine the new individuals generated by selection, crossover and mutation to form a new population and update the population: p new ={x new,1 ,x new,2 ,...,x new,N}, p new For new populations;

[0069] Repeat steps 2-6 until the maximum number of iterations is reached or the fitness value converges;

[0070] Select the individual with the highest fitness value from the final population as the optimal energy allocation strategy;

[0071] Through genetic algorithms, global energy distribution is optimized to find the optimal engine and motor power distribution strategy in complex driving environments; effectively improving the energy efficiency and driving experience of hybrid vehicles.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] The present invention uses real-time sensor data and historical driving data to establish a dynamic load prediction model to predict the load changes in the short term in the future; and uses a recursive neural network to predict the load demand under different working conditions in real time;

[0074] Adopt intelligent scheduling algorithm based on reinforcement learning to dynamically adjust the load distribution of engine and motor; output optimal scheduling strategy through reinforcement learning to balance the workload of each subsystem and optimize the overall energy efficiency; improve the system response speed, dynamically adapt to changing driving conditions and maximize energy utilization efficiency;

[0075] The present invention provides multi-mode recognition and switching: using machine learning algorithms to identify the current driving mode (city mode, highway mode, hybrid mode) in real time, and automatically switching the energy management strategy according to the mode change;

[0076] Adopting a global optimization algorithm based on genetic algorithm GA, the energy distribution of the engine and the electric motor is optimized under different driving modes to achieve global optimal energy utilization; achieving global optimal energy management between different driving modes; optimizing the mode switching process, improving energy utilization efficiency and reducing energy waste;

[0077] The present invention solves the problems that the power domain control method of hybrid vehicles in the prior art lacks flexibility and real-time response capability, making it difficult to achieve real-time optimal distribution of loads in a dynamic driving environment; and it is impossible to achieve global optimal energy management between different driving modes, resulting in low mode switching efficiency and energy waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A schematic diagram of a hybrid vehicle power domain control method according to the present invention;

[0079] Figure 2 It is a schematic diagram of the structure of the hybrid vehicle power domain control system of the present invention. DETAILED DESCRIPTION

[0080] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0081] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0082] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0083] Example 1

[0084] See also Figure 1-Figure 2,The present invention provides a hybrid electric vehicle power domain control system, comprising: a power data acquisition module, used to obtain multi-dimensional power data during the vehicle's driving process;

[0085] Dynamic load prediction module, which inputs sensor data into the RNN model in real time during vehicle driving to predict load demand under different working conditions;

[0086] A scheduling learning module to dynamically adjust the load distribution of the engine and the motor;

[0087] A mode recognition module for identifying the current driving mode;

[0088] The multi-mode switching module automatically switches to the corresponding energy management strategy based on the real-time identification of driving mode;

[0089] The present invention also provides a hybrid electric vehicle power domain control method, comprising:

[0090] Step S1. Vehicle power data collection: deploy sensors to obtain multi-dimensional power data during vehicle driving; integrate the vehicle's historical driving data as basic data for prediction and optimization; clean the collected data and extract key features; multi-dimensional power data includes: speed, acceleration, battery power, and engine status;

[0091] Step S2. Dynamic load prediction, using the RNN model to predict the load changes in the short term in the future; inputting sensor data into the RNN model in real time during vehicle driving to predict the load demand under different working conditions;

[0092] The load prediction step in step S2 using the recurrent neural network RNN ​​includes:

[0093] Collect sensor data and historical driving data of the vehicle during driving, including but not limited to:

[0094] Vehicle speed V t , acceleration A t 、Engine load L eng,t , Motor load L mot,t , Battery Capacity SOC t 、Ambient temperature T env,t , clean and normalize the data, and generate a time series data set;

[0095] Building the RNN model:

[0096] Input feature vector definition: Define the input feature vector X with a time step of t t :X t =[V t ,A t ,L eng,t,L mot,t ,SOC t ,T env,t ];

[0097] Output target definition: Define the load forecast value for the next time step t+1 Among them, ht is the hidden state of RNN;

[0098] The model structure adopts the long short-term memory LSTM network structure:

[0099] Forget Gate: where f t is the forget gate vector, σ is the Sigmoid activation function, W f represents the weight matrix of the forget gate, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous time step;

[0100] Input Gate: where i t is the input gate vector, is the candidate cell state vector, W i , W C is the weight matrix of the input gate and cell state, b i 、b C is the bias vector for the input gate and cell state;

[0101] Cell status update: Among them C t is the cell state at the current time step, C t-1 is the cell state at the previous time step;

[0102] Output Gate: Among them t is the output gate vector, W o is the weight matrix of the output gate, b o represents the bias vector of the output gate;

[0103] Hide status update:h t =o t tanh(C t ), where h t is the hidden state of the current time step;

[0104] Use mean square error (MSE) as the loss function to calculate the error between the predicted value and the actual load. Here n is the number of samples, Indicates the predicted load value, L t is the real load value;

[0105] Use the back-propagation algorithm to train the LSTM network through time BPTT;

[0106] Deploy the trained LSTM model to the vehicle control system and input sensor data in real time for load prediction;

[0107] Specifically, the long short-term memory (LSTM) model is used for dynamic load prediction, which enables the vehicle to adapt to different driving conditions in real time. By collecting and processing multi-dimensional vehicle power data, including speed, acceleration, battery power and ambient temperature, the load changes in the short term in the future are predicted; the vehicle can adjust the energy distribution strategy in advance, reduce the load prediction error, and improve the accuracy of energy management;

[0108] Step S3. Reinforcement learning scheduling, defining the state, action space and objective function, using the deep Q network DQN to train the scheduling strategy, and deploying it to the vehicle control system to dynamically adjust the load distribution of the engine and the motor;

[0109] In step S3, data including the current speed, acceleration, load prediction value, and battery power of the vehicle are defined as a state space;

[0110] The data including the power distribution between the engine and the electric motor is defined as the action space;

[0111] In step S3, a reward function is constructed based on energy efficiency and driving experience, the goal is to maximize the total reward, and a deep Q network DQN is used to train the scheduling strategy to dynamically adjust the load distribution of the engine and the motor;

[0112] In step S3, the steps of using the deep Q network DQN algorithm to train the scheduling strategy include:

[0113] Define the environment and state, state space S: define the current state of the vehicle, expressed as: Where V t is the vehicle speed, A t is the acceleration, To predict load, SOC t is the battery capacity, T env,t is the ambient temperature, L eng,t is the current engine load, L mot,t is the current motor load;

[0114] Action space A: defined as the power distribution between the engine and the motor, expressed as: A t =[P eng,t ,P mot,t ], where P eng,t is the engine power, P mot,t is the motor power;

[0115] Instant Rewards t To optimize energy efficiency and driving experience, with the goal of minimizing fuel consumption and emissions: t =-(α1·FC t +α2·EM t -β·DE t ), where FC t Indicates fuel consumption, EM t Denotes emissions, DE t It represents the driving experience, expressed as the acceleration smoothness, α1, α2, β all represent the corresponding weight coefficients, and t is the time step;

[0116] In step S3, the deep Q network DQN training scheduling strategy is used as follows:

[0117] Q network structure: Use a deep neural network to estimate the state action value function, with the input being the state S t , the output is the Q value Q(S t ,A t ;θ): Here θ is the Q network parameter, θ′ represents the target Q network parameter, and γ is the discount factor;

[0118] Stores experience fragments of the agent's interaction with the environment (S t ,A t ,r t ,S t+1 ) to the experience replay pool, and randomly select small batches of samples for training;

[0119] Calculate the target Q value y using a delayed updated target network t , stabilize the training process: Here θ - is the target Q network parameter, Q(S t+1 ,A t+1 θ - ) represents the Q value of the target Q network output, that is, the state S at the next moment t+1 Next take action A t+1 The maximum Q value of

[0120] The training process in step S3 using the deep Q network DQN training scheduling strategy is:

[0121] a. Initialization: Initialize the Q network parameters θ and the target Q network parameters θ′, set as copies of θ;

[0122] b. Collect experience: perform actions in the environment and collect (S t ,A t ,r t ,S t+1) and stored in the experience replay pool D;

[0123] c. Extract small batch samples: randomly extract small batch samples B from the experience replay pool;

[0124] d. Calculate the target Q value: For each sample, calculate the target Q value y t ,

[0125] e. Update Q network: minimize the loss function L(θ) and update the Q network parameters θ: Here, E represents the expected value, (S t ,A t ,r t ,S t+1 )~D represent small batch samples drawn from the experience replay pool D, each sample contains the current state S t 、Current Action A t , Instant Rewards t and the next state S t+1 , Q(S t ,A t ; θ) is the Q value of the current Q network output, indicating that in the current state S t Next take action A t Q value;

[0126] f. Periodically update the target network: copy the Q network parameters θ to the target network parameters θ′ at fixed steps;

[0127] g. Repeat the training process of steps af until the loss function converges or reaches the predetermined number of training rounds;

[0128] By using the deep Q network (DQN) algorithm to train the scheduling strategy, the load distribution of the engine and the motor can be dynamically adjusted to maximize energy utilization efficiency;

[0129] Specifically, through the deep Q network DQN algorithm, hybrid electric vehicles can achieve dynamic optimization of intelligent scheduling strategies, learn and update scheduling strategies from the experience replay pool, and maximize energy utilization efficiency; compared with fixed rules and PID control methods, the DQN algorithm reduces system response lag in dealing with complex nonlinear environments and real-time load changes, and improves overall energy efficiency;

[0130] Step S4. Driving mode recognition, extracting driving mode related features from real-time sensor data; using the extracted features to train a pattern recognition model; using the recognition model to recognize the current driving mode in real time; related features include speed, road type, and traffic conditions;

[0131] Optimize the energy management strategy. Hybrid vehicles have different energy requirements in different driving modes (such as urban mode, high-speed mode, hybrid mode, etc.). The current energy management strategy usually adopts decentralized local optimization, and each mode is optimized independently, ignoring the synergy between modes.

[0132] Specifically, it uses machine learning algorithms to identify driving modes, extract key features from real-time sensor data, identify the current driving mode (such as urban mode, high-speed mode, and mixed mode) in real time, and automatically switch to the corresponding energy management strategy according to the identified driving mode to achieve optimal energy allocation in each driving mode; improve the flexibility and response speed of the system, so that the vehicle can maintain the best operating state in different driving scenarios, reduce energy waste, and improve overall energy efficiency;

[0133] Step S5. Multi-mode optimization and switching: automatically switch to the corresponding energy management strategy according to the real-time identified driving mode; use genetic algorithm to optimize global energy distribution in each driving mode;

[0134] The global energy allocation optimization method is:

[0135] Initialize the population: generate an initial energy allocation strategy set;

[0136] Fitness evaluation: Calculate the fitness value of each strategy based on energy efficiency and driving experience;

[0137] Selection, crossover, and mutation: Generate new strategy combinations and iterate until convergence;

[0138] The global energy allocation optimization method in step S5 is:

[0139] 1. Generate an initial energy allocation strategy set. Each individual represents an energy allocation strategy. Suppose the population size is N, and each individual x i represents the power distribution strategy between the engine and the motor, x i =[P eng,i ,P mot,i ], P eng,i is the engine power, P mot,i is the motor power;

[0140] Randomly generate individuals in the population so that each individual satisfies the power constraint, P eng,i +P mot,i =P total , P total represents the total power demand;

[0141] 2. Define the fitness function to evaluate the quality of each individual's energy allocation strategy. The fitness function F(x i ): Where E eng,i is the engine energy output, E mot,i is the motor energy output, DE i Indicates driving experience evaluation, FC i is the fuel consumption, EM i is emission, η, β, α are weight coefficients, balancing energy efficiency, driving experience and emission;

[0142] 3. Use the roulette wheel selection method to select individuals as parents for crossover and mutation based on the fitness value. The selection probability is: Here p i That is, individual x i The probability of being selected, N represents the population size, and j represents the index of the jth individual;

[0143] 4. Use single-point crossover to combine two parent individuals to generate new individual offspring. The crossover formula is: Here x new,k is the newly generated individual, x parent,k ,x parent2,k For two parent individuals, Represents a single-point crossover operation;

[0144] 5. Use uniform mutation to randomly change part of the individual. The mutation formula is: x mut,k =x orig,k +δ, where x mut,k For the individual after mutation,

[0145] x orig,k represents the original individual, δ is a small random perturbation, satisfying x orig,k +δ≤P total ;

[0146] 6. Combine the new individuals generated by selection, crossover and mutation to form a new population and update the population: p new ={x new,1 ,x new,2 ,...,x new,N}, p new For new populations;

[0147] Repeat steps 2-6 until the maximum number of iterations is reached or the fitness value converges;

[0148] Select the individual with the highest fitness value from the final population as the optimal energy allocation strategy;

[0149] Through genetic algorithms, global energy distribution optimization is performed to find the optimal engine and motor power distribution strategy in complex driving environments; effectively improving the energy efficiency and driving experience of hybrid vehicles;

[0150] Optimize the energy management strategy of the power system of hybrid vehicles. Hybrid vehicles have different energy requirements in different driving modes (such as urban mode, high-speed mode, hybrid mode, etc.), while the current energy management strategy usually adopts decentralized local optimization, and each mode is optimized independently, ignoring the synergy between modes;

[0151] Specifically, a genetic algorithm is used to optimize global energy allocation and generate an initial population. New energy allocation strategies are iteratively generated through steps such as fitness evaluation, selection, crossover, and mutation. Genetic algorithms perform well in dealing with global optimization problems and are able to find the optimal solution in a complex search space. By balancing factors such as energy efficiency, driving experience, and emissions, they ultimately select the individual with the highest fitness value as the optimal energy allocation strategy. Compared with local optimization methods, genetic algorithms can more effectively explore the global optimal solution and improve the comprehensive performance of the vehicle under different driving modes.

[0152] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A hybrid electric vehicle power domain control method, characterized in that: It includes a hybrid electric vehicle power domain control system, which includes a power data acquisition module, a dynamic load prediction module, a scheduling learning module, a pattern recognition module, and a multi-mode switching module, wherein: Power data acquisition module, used to obtain multi-dimensional power data during vehicle driving; Dynamic load prediction module, which inputs sensor data into the RNN model in real time during vehicle driving to predict load demand under different working conditions; A scheduling learning module to dynamically adjust the load distribution of the engine and the motor; A mode recognition module for identifying the current driving mode; The multi-mode switching module automatically switches to the corresponding energy management strategy based on the real-time identification of driving mode; The hybrid electric vehicle power domain control method comprises the following steps: Step S1. Vehicle power data collection: deploy sensors to obtain multi-dimensional power data of the vehicle during driving; integrate the vehicle's historical driving data; clean the collected data and extract key features; Step S2. Dynamic load prediction, using the RNN model to predict the load changes in the short term in the future; inputting sensor data into the RNN model in real time during vehicle driving to predict the load demand under different working conditions; Step S3. Reinforcement learning scheduling, defining the state, action space and objective function, using the deep Q network DQN to train the scheduling strategy, and deploying it to the vehicle control system to dynamically adjust the load distribution of the engine and the motor; Step S4. Driving mode recognition, extracting driving mode related features from real-time sensor data; using the extracted features to train a pattern recognition model; using the recognition model to recognize the current driving mode in real time; Step S5. Multi-mode optimization and switching: automatically switch to the corresponding energy management strategy according to the real-time identified driving mode; use genetic algorithm to optimize global energy distribution in each driving mode; The load prediction step in step S2 using the recurrent neural network RNN ​​includes: Collect sensor data and historical driving data of the vehicle during driving, including: Vehicle speed V t , acceleration A t 、Engine load L eng,t , Motor load L mot,t , Battery power SOC t 、Ambient temperature T env,t , clean and normalize the data, and generate a time series data set; Building the RNN model: Input feature vector definition: Define the input feature vector X with a time step of t t :X t =[V t ,A t ,L eng,t ,L mot,t ,SOC t ,T env,t ]; Output target definition: Define the load forecast value for the next time step t+1 Among them, h t is the hidden state of RNN; The model structure adopts the long short-term memory LSTM network structure: Forget Gate: where f t is the forget gate vector, σ is the Sigmoid activation function, W f represents the weight matrix of the forget gate, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous time step; Input Gate: where i t is the input gate vector, is the candidate cell state vector, W i , W C is the weight matrix of the input gate and cell state, b i , b C is the bias vector for the input gate and cell state; Cell status update: Among them C t is the cell state at the current time step, C t-1 is the cell state at the previous time step; Output Gate: Among them t is the output gate vector, W o is the weight matrix of the output gate, b o represents the bias vector of the output gate; Hide status update:h t =o t tanh(C t ), where h t is the hidden state of the current time step; Use mean square error (MSE) as the loss function to calculate the error between the predicted value and the actual load. Here n is the number of samples, Indicates the predicted load value, L t is the real load value; Use the back-propagation algorithm to train the LSTM network through time BPTT; The trained LSTM model is deployed to the vehicle control system and sensor data is input in real time for load prediction.

2. A hybrid vehicle power domain control method according to claim 1, characterized in that: In step S3, data including the current speed, acceleration, load prediction value, and battery power of the vehicle are defined as a state space; The data including how the power of the engine and the electric motor is distributed is defined as the action space.

3. A hybrid vehicle power domain control method according to claim 2, characterized in that: In step S3, a reward function is constructed based on energy efficiency and driving experience. The goal is to maximize the total reward. The deep Q network DQN is used to train the scheduling strategy and dynamically adjust the load distribution of the engine and the motor.

4. A hybrid vehicle power domain control method according to claim 3, characterized in that: In step S3, the steps of using the deep Q network DQN algorithm to train the scheduling strategy include: Define the environment and state, state space S: define the current state of the vehicle, expressed as: Where V t is the vehicle speed, A t is the acceleration, To predict load, SOC t is the battery capacity, T env,t is the ambient temperature, L eng,t is the current engine load, L mot,t is the current motor load; Action space A: defined as the power distribution between the engine and the motor, expressed as: A t =[P eng,t ,P mot,t ], where P eng,t is the engine power, P mot,t is the motor power; Instant Rewards t To optimize energy efficiency and driving experience, with the goal of minimizing fuel consumption and emissions: t =-(α1·FC t +α2·EM t -β·DE t ), where FC t Indicates fuel consumption, EM t Denotes emissions, DE t It represents the driving experience, expressed by the acceleration smoothness. α1, α2, β all represent the corresponding weight coefficients, and t is the time step.

5. A hybrid vehicle power domain control method according to claim 4, characterized in that: In step S3, the deep Q network DQN training scheduling strategy is used as follows: Q network structure: Use a deep neural network to estimate the state action value function, with the input being the state S t , the output is the Q value Q(S t ,A t ;θ): Here θ is the Q network parameter, θ′ represents the target Q network parameter, and γ is the discount factor; Stores experience fragments of the agent's interaction with the environment (S t ,A t ,r t ,S t+1 ) to the experience replay pool, and randomly select small batches of samples for training; Use a delayed updated target network to calculate the target Q value yt and stabilize the training process: Here θ - is the target Q network parameter, Q(S t+1 ,A t+1 θ - ) represents the Q value of the target Q network output, that is, the state S at the next moment t+1 Next take action A t+1 The maximum Q value.

6. A hybrid vehicle power domain control method according to claim 5, characterized in that: The training process in step S3 using the deep Q network DQN training scheduling strategy is: a. Initialization: Initialize the Q network parameters θ and the target Q network parameters θ′, set to a copy of θ; b. Collect experience: perform actions in the environment and collect (S t ,A t ,r t ,S t+1 ) and stored in the experience replay pool D; c. Extract small batch samples: randomly extract small batch samples B from the experience replay pool; d. Calculate the target Q value: For each sample, calculate the target Q value y t , e. Update the Q network: minimize the loss function L(θ) and update the Q network parameters θ: Here, E represents the expected value, (S t ,A t ,r t ,S t+1 )~D represent small batch samples drawn from the experience replay pool D, each sample contains the current state S t 、Current Action A t , Instant Rewards t and the next state S t+1 , Q(S t ,A t ; θ) is the Q value of the current Q network output, indicating that in the current state S t Next take action A t Q value; f. Periodically update the target network: copy the Q network parameters θ to the target network parameters θ′ at fixed steps; g. Repeat the training process of steps af until the loss function converges or reaches the predetermined number of training rounds.

7. A hybrid vehicle power domain control method according to claim 6, characterized in that: The global energy allocation optimization method is: Initialize the population: generate an initial energy allocation strategy set; Fitness evaluation: Calculate the fitness value of each strategy based on energy efficiency and driving experience; Selection, crossover, and mutation: Generate new strategy combinations and iterate until convergence.

8. A hybrid vehicle power domain control method according to claim 7, characterized in that: The global energy allocation optimization method in step S5 is: 1) Generate an initial energy allocation strategy set. Each individual represents an energy allocation strategy. Let the population size be N. Each individual xi represents the power allocation strategy of the engine and the motor. i =[P eng,i ,P mot,i ], P eng,i is the engine power, P mot,i is the motor power; Randomly generate individuals in the population so that each individual satisfies the power constraint, P eng,i +P mot,i =P total , P total represents the total power demand; 2) Define the fitness function to evaluate the quality of each individual's energy allocation strategy. The fitness function F(x i ): Where E eng,i is the engine energy output, E mot,i is the motor energy output, DE i Indicates driving experience evaluation, FC i is the fuel consumption, EM i is emission, η, β, α are weight coefficients, balancing energy efficiency, driving experience and emission; 3) According to the fitness value, use the roulette wheel selection method to select individuals as parents for crossover and mutation, and the selection probability is: Here p i That is, individual x i The probability of being selected, N represents the population size, and j represents the index of the jth individual; 4) Use single-point crossover to combine two parent individuals to generate new individual offspring. The crossover formula is: Here x new,k is the newly generated individual, x parent,k ,x parent2,k For two parent individuals, Represents a single-point crossover operation; 5)Use uniform mutation to randomly change part of the individual. The mutation formula is: mut,k =x orig,k +δ, where x mut,k For the individual after mutation, x orig,k represents the original individual, δ is a small random perturbation, satisfying x orig,k +δ≤P total ; 6) Combine the new individuals generated by selection, crossover and mutation to form a new population and update the population: p new ={x new,1 ,x new,2 ,...,x new,N }, p new For new populations; Repeat steps 2)-6) until the maximum number of iterations is reached or the fitness value converges; The individual with the highest fitness value is selected from the final population as the optimal energy allocation strategy.

Citation Information

Patent Citations

  • Parallel hybrid electric vehicle power domain control system

    CN112918461A

  • Recursive deep reinforcement learning hybrid vehicle energy management strategy adaptive to multiple working conditions

    CN116561579A

  • Hybrid drive energy conversion management method and system

    CN117962863A

  • Model-predictive control of a powertrain system using preview information

    US20200223423A1

Cited By

  • Hybrid electric vehicle energy distribution method based on dynamic load prediction

    CN121316811A