Intelligent Control Method for Vehicle Energy Management Based on Multi-Level Prediction

Abstract

This invention discloses a multi-layer prediction-based intelligent control method and system for vehicle energy management. The method includes acquiring multi-dimensional vehicle data information; constructing a hybrid 1D-CNN-LSTM one-dimensional convolutional long short-term memory neural network model based on the vehicle data information to predict the vehicle speed in a finite future time domain, and obtaining the vehicle's driving power demand in a rolling time domain based on the predicted vehicle speed; optimizing the vehicle's driving power demand using a deep Q-learning algorithm (DQN) to obtain the optimal torque allocation sequence corresponding to the minimum fuel consumption in the rolling time domain; applying the first value of the optimal torque allocation sequence to the vehicle, and at the next time step, detecting the vehicle's actual speed and acceleration output value to refresh the prediction model; re-optimizing the torque allocation based on the updated prediction value, repeating the above steps at each time step until the end of the prediction time domain, and outputting the optimal torque allocation sequence in the prediction time domain to achieve intelligent control of vehicle energy management.

Description

Technical Field

[0001] This invention belongs to the field of automotive energy optimization control technology, and more specifically, relates to an intelligent control method for automotive energy management based on multi-layer prediction. Background Technology

[0002] Hybrid electric vehicles (non-range-extended) have two or more power sources, such as an engine and an electric motor. The power output from these power sources is coordinated and then used to drive the vehicle through the transmission system. Plug-in hybrid electric vehicles (PHEVs) are complex, nonlinear, multivariable, and time-varying systems. Due to their unique energy distribution structure, the control of power sources such as the engine and electric motor is more complex than that of traditional hybrid systems. Energy management strategies, as one of the key technologies of hybrid systems, aim to optimize the control of each power source while meeting the vehicle's driving and driver power requirements. This will enable each power source to operate efficiently and rationally, further tapping its energy-saving potential and improving the overall fuel economy of the vehicle. This is currently a key focus of PHEV research and an urgent need for the development of the new energy vehicle industry.

[0003] Energy management for a vehicle aims to optimize the distribution of torque from the power source while meeting the vehicle's power requirements, thereby maximizing overall vehicle performance. However, current technologies have several shortcomings in energy management, making it difficult to effectively solve the optimization problem in this complex system. These shortcomings are as follows:

[0004] Insufficient global optimization capability: Existing technologies often struggle to effectively address global optimization issues across the entire driving cycle in terms of energy management strategies. Traditional methods tend to focus on local optimization, failing to comprehensively consider the energy distribution needs of the vehicle at different driving stages from a macroscopic perspective. This results in the inability to achieve optimal energy management under complex conditions, failing to fully utilize the potential of each power source, and consequently affecting the vehicle's fuel economy and performance.

[0005] Poor adaptability to complex operating conditions: Plug-in hybrid electric vehicles (PHEVs) are complex, nonlinear, multivariable, and time-varying systems with unique energy distribution structures. Existing technologies lack effective prediction and dynamic adjustment capabilities when facing such complex conditions. They cannot accurately predict the vehicle's future driving state and energy demands, and struggle to flexibly adjust energy management strategies based on real-time changing conditions. This results in the inability to achieve efficient and rational operation of each power source in actual operation, failing to meet the vehicle's power demands and energy-saving goals under different driving conditions.

[0006] Lack of hierarchical control and collaborative optimization: Existing technologies lack a systematic hierarchical control architecture in energy management, failing to decompose complex global optimization problems into more manageable local optimization problems. Furthermore, insufficient collaborative optimization between power sources hinders efficient coordination between engines, motors, and other power sources, resulting in a complex and inefficient energy management process and poor overall vehicle performance and energy-saving effects.

[0007] Insufficient dynamic update capability: During vehicle operation, operating conditions are constantly changing, and existing technologies are inadequate in dynamically updating energy management strategies. Energy management often relies on expert knowledge and experience, which has certain limitations and cannot achieve online real-time control. It cannot quickly and accurately update energy management strategies based on real-time feedback of vehicle status and driving conditions, resulting in poor adaptability and timeliness of the strategies. This makes it impossible for the vehicle to adjust energy distribution in a timely manner when facing rapidly changing operating conditions, and it cannot achieve optimal power source torque distribution, thus affecting the overall vehicle performance and energy-saving effect. Summary of the Invention

[0008] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a multi-layer prediction-based intelligent control method and system for automotive energy management. By integrating onboard sensors, intelligent transportation systems, and vehicle-to-everything (V2X) technology to acquire multi-source vehicle data, it utilizes a 1D-CNN-LSTM hybrid model for accurate prediction of vehicle speed and power demand. Furthermore, it combines this with a deep Q-learning algorithm (DQN) to achieve online real-time optimization of torque distribution. Compared to existing technologies, this significantly improves the real-time performance, adaptability, and fuel economy of energy management. This invention forms closed-loop control through real-vehicle condition feedback, effectively reducing control errors and enhancing the robustness of the control system. Simultaneously, through hierarchical control and collaborative optimization strategies, it achieves global optimization, reducing reliance on expert knowledge and improving driving performance. Therefore, it provides an intelligent and automated energy management solution for plug-in hybrid electric vehicles.

[0009] To achieve the above objectives, one aspect of the present invention provides a vehicle energy management intelligent control method based on multi-layer prediction, characterized by comprising the following steps:

[0010] S1. Obtain multi-dimensional vehicle data information, including current vehicle status data, road slope, information of vehicles ahead, and multi-source real-time traffic information; current vehicle status data includes vehicle speed, acceleration, engine speed, and battery SOC;

[0011] S2. Based on the multi-dimensional vehicle data information, construct a 1D-CNN-LSTM one-dimensional convolutional long short-term memory neural network hybrid model to predict the vehicle speed in the future finite time domain, and obtain the vehicle driving power demand in the rolling time domain based on the predicted vehicle speed.

[0012] S3. Optimize the vehicle drive demand power in the rolling time domain using the deep Q-learning algorithm DQN to obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain.

[0013] S4. Apply the first value of the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain to the vehicle. At the next moment, detect the actual vehicle speed and acceleration output value, and refresh the prediction model. Optimize the torque distribution again based on the updated prediction value.

[0014] S5. Repeat steps S1 to S4 at each time point until the prediction time domain ends, and output the optimal torque distribution sequence in the prediction time domain to realize intelligent control of vehicle energy management.

[0015] Further, step S2 includes:

[0016] S21: Construct a one-dimensional convolutional neural network 1D-CNN; the one-dimensional convolutional neural network 1D-CNN adopts a 6-layer network structure, including 3 convolutional layers and 3 pooling layers;

[0017] S22: Construct a Long Short-Term Memory (LSTM) network; the LSTM network consists of an input layer, a hidden layer, and an output layer. Memory units are added to each neuron in the hidden layer to extract the feature values ​​of the time series and perform selective output.

[0018] S23: The features extracted by 1D-CNN are input into the Long Short-Term Memory (LSTM) network to extract the temporal features of the features. The processed features are then input into a fully connected layer. Two fully connected layers are used to reduce the dimensionality to the same level as the output label category. The softmax classifier is then used to output the classification label. The road slope, real-time traffic information, and current vehicle speed are used as input feature values ​​for the 1D-CNN model, and the output is the predicted vehicle speed within a finite time domain.

[0019] S24: Based on the predicted vehicle speed within the future finite time domain, obtain the vehicle driving power demand within the rolling time domain.

[0020] Further, step S3 includes:

[0021] S31: Calculate the demand power transfer probability matrix for each discrete vehicle speed in the prediction time domain;

[0022] S32: Determine the state variable S(t) and action variable a(t), and construct a reward function with the objective of minimizing the equivalent fuel consumption of the whole vehicle;

[0023] S33: The battery SOC safety range, engine / motor torque, and speed-power limits within the prediction time domain are used as constraints on the reward function;

[0024] S34: Solve the optimal torque distribution sequence corresponding to the minimum fuel consumption in the predicted time domain using the DQN algorithm.

[0025] Further, step S32 includes:

[0026] Select the battery state of charge C and the required power P in the prediction time domain. req The motor torque T is a state variable. m To control action variables;

[0027] By equating battery energy changes to fuel consumption using an equivalent factor, a reward function is constructed with the objective of minimizing the equivalent fuel consumption of the entire vehicle. The expected value of the minimum cumulative reward is used as the objective function. The expression for the objective function is:

[0028]

[0029] Among them, J k The optimization objective at time step k, i.e., minimizing the expected value of the cumulative return, is obtained by solving for the expected value of the cumulative return; Q * k (s,a) is the optimal state-action value function when taking action a in state s; k~k+n p For prediction in the time domain; n p The length of the prediction time domain, i.e., the number of steps for future prediction; This means minimizing the time from the current time step k to the end of the prediction time domain at time step k+n. p The expected value of the cumulative return; γ is the discount factor, r t+1 The reward obtained at time step t+1;

[0030] The reward function includes the sum of fuel consumption and the equivalent fuel consumption of electricity. To maintain the balance of State of Charge (SOC), a SOC penalty function is added to the reward function, expressed as follows:

[0031] r(s,a)=m fuel +m ele +β(SOC(t)-SOC ref (t)) 2 (7)

[0032] Where r(s,a) represents the reward function for the current state s and action a; m fuel The engine fuel consumption for the current state and action, in g; m ele The equivalent fuel energy for electrical energy is expressed in g / (kW·h); β is a weighting coefficient; SOC(t) represents the battery state of charge at time t; SOC ref This is a reference value for SOC.

[0033] Further, step S33 includes: in the prediction time domain k~k+n p The following constraints are applied to the variables related to SOC, torque, and speed:

[0034]

[0035] Where, n e (k) represents the engine speed at time k, in r / min; n e_max (k), n e_min (k) represents the maximum and minimum engine speeds at time k, in r / min; T m (k) represents the motor torque at time k, in N·m; T m_max (k), T m_min (k) represents the maximum and minimum values ​​of the motor torque at time k, in N·m; T e (k) represents the engine torque at time k; T e_max (k), T e_min (k) represents the maximum and minimum engine torque at time k, respectively, in N·m; P m (k) represents the motor power at time k, in kW; P m_max (k), P m_min (k) represents the maximum and minimum motor power at time k, in kW; P e (k) represents the engine power at time k, in kW; P e_max (k), P e_min (k) represents the maximum and minimum engine power at time k, in kW; SOC(k) represents the state of charge (SOC) of the battery at time k.

[0036] Further, step S34 includes:

[0037] S341: Initialize the basic parameters of DQN, including the experience pool size N and the number of randomly drawn samples n. b The number of iterations n, the evaluation network parameter θ, the target network parameter θ′, the exploration rate ε, and the discount factor γ;

[0038] S342: Based on the current state of the vehicle, a greedy algorithm is used to select an action. The reward value and the state at the next moment are calculated based on the current state of the vehicle and the selected action, and stored in the experience pool.

[0039] S343: Network training and network parameter update. Randomly select samples from the experience pool for network training, and use the loss function gradient descent method to update and evaluate the network parameters in reverse.

[0040] S344: Update the target network. At regular intervals, copy the updated evaluation network parameters to the target network.

[0041] S345: Determine if the prediction time domain has been reached, and determine if the current time step t exceeds the prediction time domain T. pre =k~k+n p If yes, proceed to the next step; otherwise, repeat steps S342 to S344 until the prediction time domain ends.

[0042] S346: Determine if the iteration has ended by comparing the difference |Q| between the cumulative Q-values ​​of adjacent iterations. n+1 -Q n Is it less than the set threshold (e.g., |Q)? n+1 -Q n If |<0.01), the iteration ends and proceeds to the next step; otherwise, repeat steps S342~S345 to continue iterative training.

[0043] S347: Take the strategy corresponding to the action with the smallest Q value after convergence as the optimal strategy, calculate the equivalent fuel consumption of the whole vehicle, and obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the prediction time domain.

[0044] Further, step S342 includes:

[0045] Choose the currently estimated optimal action a with a probability of 1-ε. t =argmaxQ(s t (a, θ); explore other actions randomly with probability ε to explore the environment; the expression is as follows:

[0046]

[0047] Where s is the state; a is the action taken in state s; π(a|s) represents the probability of choosing action a in state s; ε is the exploration rate; |M(s)| is the size of the set of actions available in state s; and Q(s,a) represents the Q value of taking action a in state s, i.e., the expected reward.

[0048] Further, step S343 includes:

[0049] Once the amount of data in the experience pool reaches a certain quantity, n data points are randomly selected from the experience pool. b Training is performed on 100 samples;

[0050] Calculate the minimum Q-value at the next time step using the target network, and combine it with the immediate reward r. t Calculate the target Q value; the expression is as follows:

[0051] Q target =r t +γminQ(s t+1 ,a t+1 ;θ′) (11)

[0052] Among them, Q target This represents the target Q-value, which is the expected Q-value used to update and evaluate the network parameters during training; r t For immediate feedback, it means taking action a at time step t. t The reward obtained later; γ is a discount factor, a value between 0 and 1, used to weigh the importance of current and future returns; min Q(s t+1 ,a t+1 ;θ′) represents all possible actions a in the next state st+1 at time step t+1. t+1 The minimum Q value; this minimum value is calculated from the target network, where θ′ is the target network parameter;

[0053] The network parameters are updated and evaluated using gradient descent by minimizing the loss function; the expression for the loss function is:

[0054] L(θ)=E{r t +γmin Q(s t+1 ,a t+1 ;θ′)-Q(s t ,a t ;θ)] 2} (12)

[0055] Where L(θ) is the loss function, used to measure the difference between the Q-value predicted by the evaluation network and the target Q-value; E is the expected value, representing the average over all samples; Q(s) t ,a t ;θ) represents the state s t Take action a t The Q value.

[0056] Further, in step S347, the functional expression of the strategy is:

[0057] π(s,a)=argminQ(s,a) (13)

[0058] Where π(s,a) is the policy function, representing the decision rule for taking action a in state s; argmin represents finding the independent variable that minimizes the subsequent function value; and Q(s,a) is the state-action value function.

[0059] A second aspect of the present invention provides a multi-level predictive intelligent control system for vehicle energy management, used to implement the multi-level predictive intelligent control method for vehicle energy management, comprising:

[0060] The multi-source data acquisition and fusion module is used to acquire vehicle data information from multiple dimensions; the multi-dimensional vehicle data information includes current vehicle status data, road slope, information of the vehicle in front, and multi-source real-time traffic information; the current vehicle status data includes vehicle speed, acceleration, engine speed, and battery SOC;

[0061] The working condition prediction module is used to construct a 1D-CNN-LSTM one-dimensional convolutional long short-term memory neural network hybrid model based on the multi-dimensional set of vehicle data information, predict the vehicle speed in the future finite time domain, and obtain the vehicle driving power demand in the rolling time domain based on the predicted vehicle speed.

[0062] The energy management decision module is used to optimize the vehicle drive demand power in the rolling time domain using the deep Q-learning algorithm DQN to obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain.

[0063] The real-time execution and feedback module applies the first value of the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain to the vehicle. At the next moment, it detects the actual vehicle speed and acceleration output value and refreshes the prediction model. Based on the updated prediction value, it re-optimizes the torque distribution. The above steps are repeated at each moment until the prediction time domain ends, and the optimal torque distribution sequence in the prediction time domain is output to realize intelligent control of vehicle energy management.

[0064] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0065] (1) This invention abandons the traditional mode that relies on predefined rules and uses a 1D-CNN-LSTM hybrid model to predict future operating conditions in real time, combined with deep reinforcement learning (DQN) to dynamically optimize torque distribution online. This data-driven architecture can complete decision-making within milliseconds, forming a closed-loop control of "prediction-optimization-execution-feedback", breaking through the limitations of expert experience and realizing intelligent online real-time control. It completely solves the problem of response lag caused by reliance on human experience in existing technologies. Actual tests show that the system control delay is reduced by more than 90%, meeting the millisecond response requirements of dynamic scenarios such as urban congestion and high-speed cruising, and realizing adaptive real-time control under all operating conditions.

[0066] (2) To address the problem that traditional methods often focus on local optimization and fail to comprehensively consider the energy allocation needs of vehicles at different driving stages from a macro perspective, resulting in the inability to achieve optimal energy management under complex operating conditions and thus affecting the fuel economy and performance of the vehicle, this invention constructs a 1D-CNN-LSTM hybrid model and the DQN algorithm. The 1D-CNN extracts the spatial features of road slope and traffic flow, while the LSTM captures the temporal dependence of vehicle speed. Through long-term and short-term collaborative prediction, it accurately predicts the vehicle's driving power in the future time domain. The DQN algorithm aims to minimize the equivalent fuel consumption in the prediction time domain. It combines SOC penalty to dynamically balance electric energy and fuel consumption, and coordinates the cooperation of multiple power sources from a macro perspective to avoid energy waste caused by short-sighted control. It transforms the global optimization problem into a local optimization problem in the prediction time domain. This invention integrates multi-source information and significantly improves the global optimization capability through long-term and short-term collaborative prediction, reducing the overall fuel consumption of the vehicle by 15%-20% and the battery SOC fluctuation range by 30%, thus completely solving the energy waste caused by local optimization.

[0067] (3) This invention establishes a hierarchical architecture of “upper-level DQN decision-making - lower-level power source execution”, constructs a 1D-CNN-LSTM hybrid model and DQN algorithm, decomposes the global optimization objective into local actions of collaborative control; through hierarchical control and collaborative optimization, it enhances the adaptability to complex working conditions and improves the efficiency and effectiveness of energy management; through the DQN reward function, it strictly constrains the engine / motor power limit to ensure that the action space conforms to the physical boundary; at the same time, it dynamically adjusts the power source output ratio in real time and rolling optimization. Under extreme working conditions such as sudden slope changes and rapid acceleration, the system still maintains an overall efficiency of more than 90%, and the smoothness of power source switching is significantly improved, which significantly breaks through the adaptability bottleneck of existing technologies in nonlinear systems.

[0068] (4) This invention innovatively introduces a real-time feedback-refresh mechanism, which executes only the first value of the optimal torque sequence at each step, and then detects the actual vehicle speed / acceleration and refreshes the prediction model weights. This closed-loop design enables the model to continuously correct prediction deviations caused by environmental disturbances (such as sudden traffic changes and slope fluctuations). Based on the real-time feedback of vehicle status and driving conditions, the energy management strategy is updated quickly and accurately. Through online real-time closed-loop control under all operating conditions, dynamic adjustment of energy management of plug-in hybrid electric vehicles is achieved. This effectively reduces control errors, improves the robustness of the control system, and enables the vehicle to adjust energy distribution in a timely manner when facing rapidly changing operating conditions, thereby achieving optimal power source torque distribution.

[0069] (5) This invention optimizes the torque distribution of the power source by predicting the future driving state and energy demand of the vehicle, enabling the vehicle to meet driving requirements while achieving efficient and reasonable operation of each power source. Through the ε-greedy strategy of the DQN algorithm, the system automatically learns the optimal operating point of the engine and the motor assistance strategy, and optimizes the charging and discharging timing in real time based on the equivalent fuel factor; it also considers battery life protection, breaking through the limitations of traditional rules in the insufficient utilization of the power source's efficient range; thereby tapping the energy-saving potential and improving the overall fuel economy of the vehicle. While improving fuel economy, this invention can also ensure the driving performance of the vehicle under various driving conditions, such as acceleration performance and climbing ability, thereby improving the driver's driving experience. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating an intelligent control method for vehicle energy management based on multi-layer prediction, according to an embodiment of the present invention.

[0071] Figure 2 This is a schematic diagram of the structure of a 1D-CNN convolutional neural network in an intelligent control method for vehicle energy management based on multi-layer prediction, according to an embodiment of the present invention.

[0072] Figure 3 This is a schematic diagram of the structure of a 1D-CNN-LSTM hybrid model in a vehicle energy management intelligent control method based on multi-layer prediction according to an embodiment of the present invention;

[0073] Figure 4 This is a schematic diagram of the DQN learning algorithm in a multi-level prediction-based intelligent control method for vehicle energy management according to an embodiment of the present invention.

[0074] Figure 5 This is a schematic diagram of the DQN algorithm in a multi-level prediction-based intelligent control method for vehicle energy management according to an embodiment of the present invention.

[0075] Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0077] like Figure 1As shown, one aspect of the present invention provides a multi-layer prediction-based intelligent control method for vehicle energy management, used for energy management of plug-in hybrid electric vehicles or other complex vehicle systems that are nonlinear, multivariable, and time-varying, comprising the following steps:

[0078] S1. Obtain multi-dimensional vehicle data information, including current vehicle status data, road slope, information of the vehicle in front, multi-source real-time traffic information, and driver information; current vehicle status data includes current vehicle speed, acceleration, engine speed, and battery SOC;

[0079] S2. Generate predicted operating conditions: Based on the multi-dimensional vehicle data information, construct a 1D-CNN-LSTM (1D-convolutional neural networks-long short-term memory) hybrid model to predict the vehicle speed in the future finite time domain, and obtain the vehicle driving power demand in the rolling time domain based on the predicted vehicle speed.

[0080] S3. Energy Management Algorithm Decision: The Deep Q-Learning (DQN) algorithm is used to optimize the vehicle driving power demand in the rolling time domain to obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain.

[0081] S4. Real-world vehicle status feedback: Apply the first value of the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain to the vehicle. At the next moment, detect the actual vehicle speed and acceleration output value, and refresh the prediction model. Optimize the torque distribution again based on the updated prediction value.

[0082] S5. Repeat steps S1 to S4 at each time point until the prediction time domain ends, and output the optimal torque distribution sequence in the prediction time domain to realize intelligent control of vehicle energy management.

[0083] Further, step S1 includes: using onboard sensors to acquire current vehicle driving status data, including vehicle speed, acceleration, engine speed, and battery state of charge (SOC);

[0084] Utilize intelligent transportation systems (ITS), vehicle-to-vehicle (V2V), or vehicle-to-infrastructure (V2I) communication technologies to obtain external environmental data such as road slope information, information about vehicles ahead, and real-time traffic conditions;

[0085] Vehicle-to-everything (V2X) technology can be used to obtain a wider range of traffic information, such as traffic flow and road construction status.

[0086] By fusing the above multi-source data, a multi-dimensional vehicle data set is formed, providing comprehensive input information for subsequent prediction models;

[0087] Further, step S2 includes:

[0088] S21: Construct a one-dimensional convolutional neural network (1D-CNN);

[0089] One-dimensional convolutional neural networks (CNNs) use multiple convolutional kernels with identical weights to extract features from signals, exhibiting strong robustness. Convolutional layers compute the feature parameters of neurons in the previous layer, activate them, and generate new features that are passed to the next layer. Convolutional operations establish a mapping between layers. To reduce computational cost, pooling layers are introduced to reduce the dimensionality of the convolutional features, thereby mitigating the risk of overfitting. A one-dimensional CNN refers to a network where convolution and pooling operations are performed only on a one-dimensional scale; its structure is as follows: Figure 2 As shown;

[0090] The one-dimensional convolutional neural network 1D-CNN adopts a 6-layer network structure, including 3 convolutional layers and 3 pooling layers;

[0091] Convolutional layers extract features from the input features using convolutional kernels. The calculation formula is as follows:

[0092]

[0093] In the formula, H i For the first Layer input features, H i-1 For the first Layer input features; W represents the convolution operation. i and b i Representing the first The weight matrix and corresponding bias vector of the convolutional kernel, f represents the activation function, and nonlinear factors are introduced into the model;

[0094] Pooling layers are used to reduce the dimensionality of the convolutional features, thereby reducing computational cost and the risk of overfitting. The calculation formula is as follows:

[0095]

[0096] In the formula, H i and H′ i These represent the feature quantities before and after pooling, respectively. down(·) represents the pooling function, and the dimensionality reduction is performed using max pooling.

[0097] S22: Construct a Long Short-Term Memory (LSTM) network;

[0098] The Long Short-Term Memory Network consists of an input layer, a hidden layer, and an output layer. Memory units are added to each neural unit in the hidden layer to extract the feature values ​​of the time series and perform selective output.

[0099] The memory unit structure of the Long Short-Term Memory (LSTM) network controls the updating of cell states through input gates, forget gates, and output gates. The calculation formula is as follows:

[0100]

[0101] In the formula, f t i t o t The calculation results are for the three gate structures: the input gate, the forget gate, and the output gate. t The input at time t; h t C is the LSTM output at time t; t The cell state at time t; C′ t For the new candidate value vector; W f W i W o These are the weight matrices for the input gate, forget gate, and output gate, respectively; b f b i b o W represents the bias vectors for the input gate, forget gate, and output gate; c It is the weight matrix of the candidate value vector, and its bias is b. c σ and tanh represent two activation functions;

[0102] S23: Predict vehicle speed in the future within a finite time domain using a 1D-CNN-LSTM hybrid model; output a rolling time-domain predicted vehicle speed sequence.

[0103] Features extracted by 1D-CNN are input into a Long Short-Term Memory (LSTM) network to extract temporal features. The processed features are then fed into fully connected layers, where two fully connected layers reduce the dimensionality to match the output label category. A Softmax classifier outputs the classification label. Road slope, real-time traffic information, and current vehicle speed are used as input features to the 1D-CNN model, which outputs a predicted vehicle speed over a finite time domain. The prediction principle of the 1D-CNN-LSTM hybrid model is as follows: Figure 3 As shown;

[0104] S24: Based on the predicted vehicle speed within the future finite time domain, obtain the vehicle driving power demand within the rolling time domain; output the driving power demand sequence; the formula for calculating the vehicle driving power demand within the rolling time domain is:

[0105]

[0106] Among them, P reg F represents the vehicle drive power demand in the rolling time domain; f F represents rolling resistance.w For wind resistance, F j ρ is the slope resistance; v is the predicted vehicle speed; q is the rolling resistance coefficient; m is the vehicle mass; g is the acceleration due to gravity; C D δ is the air resistance coefficient; A is the vehicle's frontal area; δ is a coefficient related to vehicle acceleration. It is the vehicle's acceleration.

[0107] Further, step S3 includes:

[0108] S31: Calculate the demand power transfer probability matrix for each discrete vehicle speed in the prediction time domain;

[0109] S32: Determine the state variable S(t) and action variable a(t), and construct a reward function with the objective of minimizing the equivalent fuel consumption of the whole vehicle;

[0110] S33: The battery SOC safety range, engine / motor torque, and speed-power limits within the prediction time domain are used as constraints on the reward function;

[0111] S34: Solve the optimal torque distribution sequence corresponding to the minimum fuel consumption in the predicted time domain using the DQN algorithm;

[0112] Further, step S32 includes:

[0113] Select the battery state of charge C and the required power P in the prediction time domain. req The motor torque T is a state variable. m To control (action) variables; the expression is:

[0114]

[0115] Among them, P req SOC(t) is the power demand at time t, representing the total power required by the vehicle to meet its driving needs at that moment; this power is calculated based on the vehicle's driving resistance (such as rolling resistance, air resistance, and gradient resistance) and acceleration; SOC(t) is the battery state of charge at time t, representing the percentage of remaining battery charge; T m (t) is the motor torque at time t, representing the torque output by the motor at that moment; this variable is a control variable, and by adjusting the motor torque, the vehicle's power output and energy consumption can be affected;

[0116] By equating battery energy changes to fuel consumption using an equivalent factor, a reward function is constructed with the objective of minimizing the equivalent fuel consumption of the entire vehicle. The expected value of the minimum cumulative reward is used as the objective function. The expression for the objective function is:

[0117]

[0118] Among them, J k The optimization objective at time step k, i.e., minimizing the expected value of the cumulative return, is obtained by solving for the expected value of the cumulative return; Q * k (s,a) is the optimal state-action value function when taking action a in state s; k~k+n p For prediction in the time domain; n p The length of the prediction time domain, i.e., the number of steps for future prediction; This means minimizing the time from the current time step k to the end of the prediction time domain at time step k+n. p The expected value of the cumulative return; γ is the discount factor, r t+1 The reward obtained at time step t+1;

[0119] The reward function includes the sum of fuel consumption and the equivalent fuel consumption of electricity. Meanwhile, to maintain the balance of State of Charge (SOC), a SOC penalty function is added to the reward function, expressed as:

[0120] r(s,a)=m fuel +m ele +β(SOC(t)-SOC ref (t)) 2 (7)

[0121] Where r(s,a) represents the reward function for the current state and action; m fuel The engine fuel consumption for the current state and action, in g; m ele The equivalent fuel energy for electrical energy is expressed in g / (kW·h); β is a weighting coefficient; SOC(t) represents the battery state of charge at time t; SOC ref This is a reference value for SOC;

[0122] Furthermore, step S33 includes: to protect the battery and prevent overcharging or over-discharging, the battery's SOC needs to be limited to a reference range; during the optimization process, factors such as torque and speed can also interfere with the optimization results, therefore, in the prediction time domain k~k+n p The following constraints are applied to the variables related to SOC, torque, and speed:

[0123]

[0124] Where, n e (k) represents the engine speed at time k, in r / min; n e_max (k), n e_min (k) represents the maximum and minimum engine speeds at time k, in r / min; T m(k) represents the motor torque at time k, in N·m; T m_max (k), T m_min (k) represents the maximum and minimum values ​​of the motor torque at time k, in N·m; T e (k) represents the engine torque at time k; T e_max (k), T e_min (k) represents the maximum and minimum engine torque at time k, respectively, in N·m; P m (k) represents the motor power at time k, in kW; P m_max (k), P m_min (k) represents the maximum and minimum motor power at time k, in kW; P e (k) represents the engine power at time k, in kW; P e_max (k), P e_min (k) represents the maximum and minimum engine power at time k, in kW; SOC(k) represents the state of charge (SOC) of the battery at time k.

[0125] Furthermore, Figure 4 This is a schematic diagram of the DQN learning algorithm. Figure 5 The flowchart below shows the DQN algorithm. In step S3, the DQN algorithm obtains the optimal strategy by continuously training and iterating the Q-network. First, based on the current state of the vehicle, an ε-greedy algorithm is used to select actions (selecting the action value corresponding to the largest Q value with probability 1-ε, and selecting other actions with probability ε to avoid getting trapped in local optima). Second, the current reward value and the state at the next moment are calculated based on the current state and the selected actions, and stored in the experience pool. When the experience pool has a certain number of samples, in order to avoid the correlation of training samples and increase training efficiency, samples are randomly selected from the experience pool for network training. The two training networks have the same structure, namely the target network and the evaluation network. The evaluation network is used to obtain the Q value corresponding to the current state-action, and the output is Q(s). t ,a t The target network then outputs Q(s) corresponding to the next state-action sequence, based on the state at the next moment (θ). t+1 ,a t+1 ;θ′), then calculate the target Q target =r t +γminQ(s t+1 ,a t+1 Finally, by continuously updating the parameters of the evaluation network using gradient descent with the goal of minimizing the loss function, the evaluated Q-value is made as close as possible to the target Q-value. To increase the stability of the algorithm, the parameters of the evaluation network are copied to the target network at regular time steps. This iterative solution continues until the prediction time domain ends.

[0126] Step S34 includes:

[0127] S341: Initialize the basic parameters of DQN, including the experience pool size N and the number of randomly drawn samples n. b The number of iterations n, the evaluation network parameter θ, the target network parameter θ′, the exploration rate ε, and the discount factor γ;

[0128] S342: Based on the current state of the vehicle, a greedy algorithm is used to select an action. The reward value and the state at the next moment are calculated based on the current state of the vehicle and the selected action, and stored in the experience pool.

[0129] S343: Network training and network parameter update. Randomly select samples from the experience pool for network training, and use the loss function gradient descent method to update and evaluate the network parameters in reverse.

[0130] S344: Update the target network. At regular intervals, copy the updated evaluation network parameters to the target network.

[0131] S345: Determine if the prediction time domain has been reached, and determine if the current time step t exceeds the prediction time domain T. pre =k~k+n p If yes, proceed to the next step; otherwise, repeat steps S342 to S344 until the prediction time domain ends.

[0132] S346: Determine if the iteration has ended by comparing the difference |Q| between the cumulative Q-values ​​of adjacent iterations. n+1 -Q n Is it less than the set threshold (e.g., |Q)? n+1 -Q n If |<0.01), the iteration ends and proceeds to the next step; otherwise, repeat steps S342~S345 to continue iterative training.

[0133] S347: Take the strategy corresponding to the action with the smallest Q value after convergence as the optimal strategy, calculate the equivalent fuel consumption of the whole vehicle, and obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the prediction time domain.

[0134] Further, in step S342, the step of selecting an action using a greedy algorithm (ε-greedy strategy) based on the current state of the vehicle, calculating the reward value and the state at the next moment based on the current state of the vehicle and the selected action, and storing them in the experience pool includes:

[0135] Select the battery state of charge (SOC) and power demand (P) in the prediction time domain. req For state variable s t =[SOC,P reg Motor torque T m For action variable a t =[Tm At each time step, based on the current state s t Use the ε-greedy strategy to select the optimal action a t Specifically, the currently estimated optimal action a is selected with a probability of 1-ε. t =argmaxQ(s t (a, θ); explore other actions randomly with probability ε to explore the environment; the expression is as follows:

[0136]

[0137] Where s is the state; a is the action taken in state s; π(a|s) represents the probability of choosing action a in state s; ε is the exploration rate, a decimal between 0 and 1, which determines the probability of randomly choosing an action; |M(s)| is the size of the set of actions available in state s; Q(s,a) represents the Q value of taking action a in state s, i.e., the expected reward; Formula (9) defines two conditions, based on which the probability of choosing an action is:

[0138] When a = argmaxQ(s,a), it means that action a is the action with the largest Q value in state s, i.e., the currently estimated optimal action; the probability of choosing this action is... Here, 1-ε represents choosing the optimal action most of the time (with a probability of 1-ε), while This indicates that there is still a small probability (ε is evenly distributed among all actions) of exploring;

[0139] When a ≠ argmaxQ(s,a), it means that action a is not the action with the maximum Q value in state s; the probability of choosing this action is... This ensures that all non-optimal actions have the same probability of being selected, thus promoting exploration;

[0140] Perform the selected action in the environment and record the immediate reward r obtained after the action is performed. t and the state s at the next moment t+1 ;

[0141] The current state s t Optimal action a t Instant rewards t and the state s at the next moment t+1 Stored in the experience pool, the expression for the experience pool is:

[0142] e t =(s t ,a t ,r t ,s t+1 (10);

[0143] The ε-greedy strategy employed in this invention avoids getting trapped in local optima by selecting the currently estimated optimal action most of the time, while retaining a certain probability of random exploration, and ultimately finds the globally optimal strategy.

[0144] Further, in step S343, the step of randomly selecting samples from the experience pool for network training and using the loss function gradient descent method to update and evaluate the network parameters in reverse includes:

[0145] Once the amount of data in the experience pool reaches a certain quantity, n data points are randomly selected from the experience pool. b Training is performed on 100 samples;

[0146] Check if the target network needs to be updated. If the current time step t is an integer multiple of the iteration interval step-c, then update it. That is, perform a modulo operation on the time step t being an iteration interval step-c, with the expression: Mod(t, step-c) = 0, and check if the result is equal to 0. When Mod(t, step-c) = 0, it means that the time step t is an integer multiple of the iteration interval step-c.

[0147] Calculate the minimum Q-value at the next time step using the target network, and combine it with the immediate reward r. t Calculate the target Q-value; use the evaluation network to calculate the evaluation Q-value Q. eval ;

[0148] The expression for the target Q value is as follows:

[0149] Q target =r t +γminQ(s t+1 ,a t+1 ;θ′) (11)

[0150] Among them, Q target This represents the target Q-value, which is the expected Q-value used to update and evaluate the network parameters during training; r t For immediate feedback, it means taking action a at time step t. t The reward obtained later; γ is a discount factor, a value between 0 and 1, used to weigh the importance of current and future returns; minQ(s t+1 ,a t+1 ;θ′) represents all possible actions a in the next state st+1 at time step t+1. t+1 The minimum Q value; this minimum value is calculated from the target network, where θ′ is the target network parameter;

[0151] The expression for evaluating the Q value is: Q eval =Q(s) t ,at ,θ);

[0152] The core idea of ​​formula (11) is to make the immediate return r t The minimum expected value of future rewards minQ(s) t+1 ,a t+1 Combined with θ′, the target Q value Q is formed. target This objective Q-value is used to train the evaluation network, by minimizing the Q-value predicted by the evaluation network. eval With the target Q value Q target The differences between them are used to update the parameters of the evaluation network;

[0153] The network parameters are updated and evaluated using gradient descent by minimizing the loss function; the expression for the loss function is:

[0154] L(θ)=E{r t +γminQ(s t+1 ,a t+1 ;θ′)-Q(s t ,a t ;θ)] 2} (12)

[0155] Where L(θ) is the loss function, used to measure the difference between the Q-value predicted by the evaluation network and the target Q-value; E is the expected value, representing the average over all samples; Q(s) t ,a t ;θ) represents the state s t Take action a t The Q-value is the Q-value predicted by the evaluation network based on the current evaluation network parameters θ, i.e., Q0 eval ;

[0156] By minimizing the difference between the Q-value predicted by the evaluation network and the target Q-value, the evaluation network's predictions become more accurate.

[0157] Further, in step S347, the functional expression of the strategy is:

[0158] π(s,a)=argminQ(s,a) (13)

[0159] Where π(s,a) is the policy function, which defines the decision rule for taking action a in state s; argmin represents finding the independent variable that minimizes the value of the subsequent function; Q(s,a) is the state-action value function, which estimates the expected reward of taking action a from state s and following a certain policy.

[0160] In step S347, the expression for the optimal motor torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain is:

[0161] L = [T] m (k), T m (k+1 / k), ..., T m (k+n p / k)] (14)

[0162] Where L is the optimal motor torque distribution sequence at time k;

[0163] During the training process of the DQN algorithm of this invention, the Q-value is continuously updated to reflect the expected reward of different state-action pairs. When the algorithm converges, for each state s, the action s that minimizes the Q-value is selected as the optimal strategy, thereby achieving the goal of energy management. In energy management problems, we usually want to minimize energy consumption (e.g., fuel consumption), so selecting the action with the minimum Q-value means selecting the action with the lowest expected energy consumption.

[0164] Furthermore, in step S4, in actual control, only the first value T of the optimal torque sequence is used. m (k) is applied to the vehicle; at time k+1, the actual vehicle speed and acceleration output value are detected first, the prediction model is refreshed, the vehicle acceleration in the future finite time domain is re-predicted, and the torque distribution is re-optimized based on the updated prediction value; steps S1 to S3 are repeated at each time until the prediction time domain ends, and the optimal torque distribution sequence in the prediction time domain can be obtained.

[0165] Step S4 specifically includes:

[0166] S41: The first value T of the optimal torque distribution sequence at time k obtained during the rolling optimization process. m (k) Applied to electric motors in vehicles;

[0167] S42: At the next time k+1, detect the vehicle's actual speed, acceleration, and other operating data;

[0168] S43: Refresh the prediction model using the collected actual operating data, including updating the current state of the vehicle, to ensure that the prediction model reflects the latest state of the vehicle;

[0169] S44: Based on the updated prediction model, re-predict vehicle acceleration within a finite time domain in the future;

[0170] S45: Based on the re-predicted acceleration and the vehicle's current state, the DQN algorithm is used again to optimize torque distribution; by dynamically adjusting the torque distribution to adapt to the vehicle's actual operating conditions, optimal energy management is achieved.

[0171] Furthermore, step S5 includes: repeating steps S1 to S4 at each time point until the end of the prediction time domain, and realizing intelligent control of vehicle energy management through continuous feedback and adjustment; when the end point of the prediction time domain is reached, the current rolling optimization process ends, and the energy management optimization within a prediction time domain is completed.

[0172] This invention accurately obtains vehicle speed and power demand in the future time domain by constructing a 1D-CNN-LSTM hybrid prediction model. Based on the rolling optimization mechanism of deep reinforcement learning (DQN), it solves the optimal torque allocation sequence in real time with the goal of minimizing the equivalent fuel consumption of the entire vehicle. This completely solves the core defects of traditional methods, such as insufficient global optimization capability and poor adaptability to complex operating conditions due to reliance on expert rules. Through an innovative chain of multi-source data fusion, long-term collaborative prediction, reinforcement learning decision-making, and closed-loop rolling optimization, it solves five key challenges in energy management of plug-in hybrid vehicles: real-time performance, global impact, adaptability, robustness, and economy, providing an industrial-grade solution for intelligent control of new energy vehicles.

[0173] like Figure 2 As shown, a second aspect of the present invention provides a multi-layer prediction-based intelligent control system for automotive energy management, used to implement the above-described design method, comprising:

[0174] A multi-source data acquisition and fusion module is used to acquire vehicle data information from multiple dimensions. This multi-dimensional vehicle data information includes current vehicle status data, road slope, information about vehicles ahead, and multi-source real-time traffic information. Current vehicle status data includes vehicle speed, acceleration, engine speed, and battery state of charge (SOC). Specifically, onboard sensors are used to acquire current vehicle driving status data, including vehicle speed, acceleration, engine speed, and battery SOC. Intelligent Transportation Systems (ITS), vehicle-to-vehicle (V2V), or vehicle-to-infrastructure (V2I) communication technologies are used to acquire external environmental data such as road slope information, information about vehicles ahead, and real-time traffic conditions. Vehicle-to-everything (V2X) technology is used to acquire broader traffic information, such as traffic flow and road construction conditions.

[0175] The working condition prediction module is used to construct a 1D-CNN-LSTM one-dimensional convolutional long short-term memory neural network hybrid model based on the multi-dimensional set of vehicle data information, predict the vehicle speed in the future finite time domain, and obtain the vehicle driving power demand in the rolling time domain based on the predicted vehicle speed.

[0176] The energy management decision module is used to optimize the vehicle drive demand power in the rolling time domain using the deep Q-learning algorithm (DQN) to obtain the optimal torque allocation sequence corresponding to the minimum fuel consumption in the rolling time domain.

[0177] The real-time execution and feedback module applies the first value of the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain to the vehicle. At the next moment, it detects the actual vehicle speed and acceleration output value and refreshes the prediction model. Based on the updated prediction value, it re-optimizes the torque distribution. The above steps are repeated at each moment until the prediction time domain ends, and the optimal torque distribution sequence in the prediction time domain is output to realize intelligent control of vehicle energy management.

[0178] It should be noted that the multi-level prediction-based intelligent control system for vehicle energy management provided in this embodiment can be a computer program (including program code) running on a computer device. For example, the multi-level prediction-based intelligent control system for vehicle energy management is an application software. The multi-level prediction-based intelligent control system for vehicle energy management can be used to execute the corresponding steps in the methods provided in the embodiments of this application.

[0179] In some feasible implementations, the multi-level prediction-based intelligent control system for vehicle energy management provided in this embodiment can be implemented using a combination of hardware and software. As an example, the multi-level prediction-based intelligent control system for vehicle energy management provided in this application embodiment can be a processor in the form of a hardware decoding processor, which is programmed to execute the multi-level prediction-based intelligent control method for vehicle energy management provided in this application embodiment. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0180] In some feasible implementations, the intelligent control system for vehicle energy management based on multi-level prediction provided in this embodiment can be implemented in software. It can be software in the form of programs and plug-ins, and includes a series of modules to implement the intelligent control method for vehicle energy management based on multi-level prediction provided in this embodiment of the invention.

[0181] A third aspect of the present invention also provides an electronic device, Figure 6 This is a schematic diagram of the electronic device in this embodiment, as shown below. Figure 6As shown, the electronic device 1000 in this embodiment may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the electronic device 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a display screen and a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 6 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0182] like Figure 6 In the electronic device 1000 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface for users; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the various steps of the intelligent control method for vehicle energy management based on multi-layer prediction.

[0183] It should be understood that in some feasible implementations, the processor 1001 described above may be a central processing unit (CPU), which may also be other general-purpose processors, DSPs, ASICs, FPGAs, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0184] In specific implementation, the aforementioned electronic device 1000 can perform the above-described actions through its built-in functional modules. Figure 1 The implementation methods provided for each step are detailed in the above-mentioned implementation methods, and will not be repeated here.

[0185] This application also provides a computer-readable storage medium storing a computer program that is executed by a processor to perform... Figure 1 The methods provided in each step are detailed in the implementation methods provided in the above steps, and will not be repeated here.

[0186] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0187] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A vehicle energy management intelligent control method based on multi-layer prediction, characterized in that, Includes the following steps: S1. Obtain multi-dimensional vehicle data information, including current vehicle status data, road slope, information of vehicles ahead, and multi-source real-time traffic information; current vehicle status data includes vehicle speed, acceleration, engine speed, and battery SOC; S2. Based on the multi-dimensional vehicle data information, construct a 1D-CNN-LSTM one-dimensional convolutional long short-term memory neural network hybrid model to predict the vehicle speed in the future finite time domain, and obtain the vehicle driving power demand in the rolling time domain based on the predicted vehicle speed. S3. Optimize the vehicle drive demand power in the rolling time domain using the deep Q-learning algorithm DQN to obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain. S4. Apply the first value of the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain to the vehicle. At the next moment, detect the actual vehicle speed and acceleration output value, and refresh the prediction model. Optimize the torque distribution again based on the updated prediction value. S5. Repeat steps S1 to S4 at each time point until the prediction time domain ends, and output the optimal torque distribution sequence in the prediction time domain to realize intelligent control of vehicle energy management. Step S2 includes: S21: Construct a one-dimensional convolutional neural network 1D-CNN; the one-dimensional convolutional neural network 1D-CNN adopts a 6-layer network structure, including 3 convolutional layers and 3 pooling layers; S22: Construct a Long Short-Term Memory (LSTM) network; the LSTM network consists of an input layer, a hidden layer, and an output layer. Memory units are added to each neuron in the hidden layer to extract the feature values ​​of the time series and perform selective output. S23: The features extracted by 1D-CNN are input into the Long Short-Term Memory (LSTM) network to extract the temporal features of the features. The processed features are then input into a fully connected layer. Two fully connected layers are used to reduce the dimensionality to the same level as the output label category. The softmax classifier is then used to output the classification label. The road slope, real-time traffic information, and current vehicle speed are used as input feature values ​​for the 1D-CNN model, and the output is the predicted vehicle speed within a finite time domain. S24: Based on the predicted vehicle speed within the future finite time domain, obtain the vehicle driving power demand within the rolling time domain; Step S3 includes: S31: Calculate the demand power transfer probability matrix for each discrete vehicle speed in the prediction time domain; S32: Determine the state variables Action variables Construct a reward function with the objective of minimizing the equivalent fuel consumption of the entire vehicle; S33: The battery SOC safety range, engine / motor torque, and speed-power limits within the prediction time domain are used as constraints on the reward function; S34: Solve the optimal torque distribution sequence corresponding to the minimum fuel consumption in the predicted time domain using the DQN algorithm; Step S33 includes: in the prediction time domain The following constraints are applied to the variables related to SOC, torque, and speed: （8） in, Represents the engine speed at time k, in r / min; , are the maximum and minimum engine speeds at time k, respectively, in r / min; Let be the motor torque at time k, in N·m; , Let N and M be the maximum and minimum values ​​of the motor torque at time k, respectively; Let be the engine torque at time k; , are the maximum and minimum engine torque values ​​at time k, respectively, in N·m; Let k be the motor power at time k, in kW; , Let be the maximum and minimum motor power at time k, in kW; Let the engine power at time k be kW; , Let K be the maximum and minimum engine power at time k, in kW; Let S be the state of charge (SOC) of the battery at time k. Step S34 includes: S341: Initialize DQN basic parameters, including experience pool capacity. Number of randomly selected samples Number of iterations Evaluate network parameters Target network parameters Exploration rate and discount factor ; S342: Based on the current state of the vehicle, a greedy algorithm is used to select an action. The reward value and the state at the next moment are calculated based on the current state of the vehicle and the selected action, and stored in the experience pool. S343: Network training and network parameter update. Randomly select samples from the experience pool for network training, and use the loss function gradient descent method to update and evaluate the network parameters in reverse. S344: Update the target network. At regular intervals, copy the updated evaluation network parameters to the target network. S345: Determine if the prediction time domain has been reached, and determine the current time step. Does it exceed the prediction time domain? If yes, proceed to the next step; otherwise, repeat steps S342 to S344 until the prediction time domain ends. S346: Determine if the iteration has ended, and evaluate the cumulative number of adjacent iterations. Difference between values If the value is less than the set threshold, the iteration ends and proceeds to the next step; otherwise, repeat steps S342 to S345 to continue iterative training. S347: After convergence The strategy corresponding to the action with the smallest value is taken as the optimal strategy. The equivalent fuel consumption of the whole vehicle is calculated, and the optimal torque distribution sequence corresponding to the minimum fuel consumption in the prediction time domain is obtained.

2. The intelligent control method for vehicle energy management based on multi-layer prediction according to claim 1, characterized in that: Step S32 includes: The battery state of charge C and the required power P in the prediction horizon are selected as state variables req The motor torque T is selected as a control action variable m ​ By equating battery energy changes to fuel consumption using an equivalent factor, a reward function is constructed with the objective of minimizing the equivalent fuel consumption of the entire vehicle. The expected value of the minimum cumulative reward is used as the objective function. The expression for the objective function is: （6） in, In time step The optimization objective, namely the minimum expected value of the cumulative return, is obtained by solving for the expected value of the cumulative return; It is in state Take action below The optimal state-action value function at that time; For prediction in the time domain; The length of the prediction time domain, i.e., the number of steps for future prediction; This means minimizing from the current time step. End time step of the prediction time domain The expected value of cumulative returns; As a discount factor, In time step The rewards received; The reward function includes the sum of fuel consumption and the equivalent fuel consumption of electricity. To maintain the balance of State of Charge (SOC), a SOC penalty function is added to the reward function, expressed as follows: （7） in, Indicates the current state With action The reward function; The engine fuel consumption, in grams, is the current state and the amount of fuel consumed during the current action. The equivalent energy of fuel for electrical energy is expressed in g / (kW·h). These are the weighting coefficients; Indicates the state of charge of the battery over time; This is a reference value for SOC.

3. The intelligent control method for vehicle energy management based on multi-layer prediction according to claim 1 or 2, characterized in that: Step S342 includes: by The probability of choosing the currently estimated optimal action. ; in terms of probability Randomly explore other actions to explore the environment; the expression is as follows: （9） in, It is a state; It is a state The following actions were taken; Indicates the state Select action The probability of; For exploration rate; In state The size of the available action set; Indicates the state Take action below of Value, i.e., expected return.

4. A vehicle energy management intelligent control method based on multi-layer prediction according to claim 1 or 2, characterized in that: Step S343 includes: Once the amount of data in the experience pool reaches a certain level, data is randomly drawn from the experience pool. Training is performed on 100 samples; Use the target network to compute the minimum value at the next time step. Value, combined with instant rewards Calculation target Value; the expression is as follows: （11） in, Indicate target This value is the expectation used to update and evaluate the network parameters during training. value; For immediate feedback, it indicates the time step. Take action The reward received later; The discount factor is a value between 0 and 1 used to weigh the importance of current returns against future returns. Indicates at time step The next state In, all possible actions of The minimum value; this minimum value is calculated from the target network. For target network parameters; The network parameters are updated and evaluated using gradient descent by minimizing the loss function; the expression for the loss function is: （12） in, It is a loss function used to measure and evaluate the network's predictions. Values ​​and Targets Differences between values; Let be the expected value, representing the average over all samples; Indicates the state Take action below of value.

5. A vehicle energy management intelligent control method based on multi-layer prediction according to claim 1 or 2, characterized in that: In step S347, the functional expression of the strategy is: （13） in, Let be the policy function, representing the state. Take action below Decision-making rules; This means finding the independent variable that minimizes the value of the subsequent function; This is a state-action value function.

6. A multi-layer prediction-based intelligent control system for automotive energy management, characterized in that, The method for implementing the intelligent control method for vehicle energy management based on multi-layer prediction as described in any one of claims 1-5 includes: The multi-source data acquisition and fusion module is used to acquire vehicle data information from multiple dimensions; the multi-dimensional vehicle data information includes current vehicle status data, road slope, information of the vehicle in front, and multi-source real-time traffic information; the current vehicle status data includes vehicle speed, acceleration, engine speed, and battery SOC; The working condition prediction module is used to construct a 1D-CNN-LSTM one-dimensional convolutional long short-term memory neural network hybrid model based on the multi-dimensional set of vehicle data information, predict the vehicle speed in the future finite time domain, and obtain the vehicle driving power demand in the rolling time domain based on the predicted vehicle speed. The energy management decision module is used to optimize the vehicle drive demand power in the rolling time domain using the deep Q-learning algorithm DQN to obtain the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain. The real-time execution and feedback module applies the first value of the optimal torque distribution sequence corresponding to the minimum fuel consumption in the rolling time domain to the vehicle. At the next moment, it detects the actual vehicle speed and acceleration output value and refreshes the prediction model. Based on the updated prediction value, it re-optimizes the torque distribution. The above steps are repeated at each moment until the prediction time domain ends, and the optimal torque distribution sequence in the prediction time domain is output to realize intelligent control of vehicle energy management.

Images