Hybrid vehicle energy management method and system considering driving condition changes

Abstract

The application discloses a hybrid electric vehicle energy management method and system considering driving condition changes, which can avoid battery health state damage caused by excessive charging and discharging or electric quantity sharp fluctuation, and effectively improve the economy of the hybrid electric vehicle. The method comprises the following steps: inputting vehicle speed information into a trained driving condition recognition model, and outputting a driving condition recognition result, wherein the driving condition recognition result is an urban driving condition or a suburban driving condition; the driving condition recognition model is built based on a sequence network, and parameters are initialized by using a glorot normal distribution, an Adam optimization algorithm is used as an optimizer, and a cross-entropy loss function is used for model training; the sequence network comprises a recurrent neural network (RNN), a long short-term memory (LSTM) or a neural network based on an attention mechanism (Transformer); an energy optimization controller is selected according to the driving condition recognition result, the driving condition recognition result is input into the selected energy optimization controller, and an instantaneous energy optimization control result is output.

Description

Technical Field

[0001] This application relates to the field of hybrid electric vehicle energy management technology, and more particularly to a hybrid electric vehicle energy management method and system that takes into account changes in driving conditions. Background Technology

[0002] Hybrid electric vehicles, as a solution for the transition from traditional gasoline vehicles to pure electric vehicles, mostly use internal combustion engines to support long-distance travel while utilizing battery packs to improve fuel economy. Energy management strategies allocate power output from multiple power sources during vehicle operation, playing a crucial role in the vehicle's economy and stability.

[0003] With the rapid development of various sensor technologies, adaptive energy management strategies can further improve vehicle fuel economy by analyzing external vehicle data and adjusting underlying strategies. In recent years, adaptive energy management strategies have received considerable attention. However, issues such as the accuracy of variable parameter identification and the performance of the underlying control strategy significantly limit overall vehicle performance. How to combine road condition information with energy management strategies to maximize energy utilization has become a major challenge.

[0004] Vehicle driving condition recognition often employs fuzzy rule-based or clustering methods to extract multi-dimensional features from data as input, resulting in low accuracy. Low-level control strategy development methods can be categorized into rule-based, optimization-based, and learning-based approaches. Learning-based strategies are increasingly prominent in the literature. However, rule-based or optimization-based methods are still commonly used to build the low-level strategy in adaptive energy management strategy development. In summary, constructing a road recognition model based on a sequence network architecture and developing the low-level strategy based on deep reinforcement learning can better integrate road condition information with energy management strategies, thereby avoiding overcharging / discharging or drastic power fluctuations that could damage battery health and effectively improve the economic efficiency of hybrid vehicles. Summary of the Invention

[0005] To overcome the aforementioned technical deficiencies, this application provides a deep learning-based energy management method and system for hybrid electric vehicles that considers changes in driving conditions. The main technical solutions are as follows:

[0006] S1. Input the vehicle speed information into the trained driving condition recognition model and output the driving condition recognition result, which is either urban driving condition or suburban driving condition.

[0007] The driving condition recognition model is built on a sequence network and initializes the parameters using the glorot normal distribution. It uses the Adam optimization algorithm as the optimizer and the cross-entropy loss function for model training. The sequence network includes recurrent neural networks (RNN), long short-term memory networks (LSTM), or neural networks based on the attention mechanism Transformer.

[0008] S2. Select an energy optimization controller based on the driving condition identification result, input the driving condition identification result into the selected energy optimization controller, and output the instantaneous energy optimization control result;

[0009] The energy optimization controller achieves instantaneous energy optimization control through the energy management strategy of the hybrid vehicle under different driving conditions. The energy management strategy is obtained by training the driving cycle data with a deep reinforcement learning algorithm to adjust the weights of the multi-objective reward function. The multi-objective reward function is established based on fuel consumption and battery SOC changes.

[0010] In addition, this application also provides a corresponding hybrid electric vehicle energy management system for executing the above-mentioned deep learning hybrid electric vehicle energy management method that takes into account changes in driving conditions.

[0011] The technical effects achieved by using the above-described technical solution in this application embodiment are as follows:

[0012] This application improves the accuracy of driving condition recognition by extracting long-term dependency information from vehicle speed data through sequence networks. By adjusting the weights of the reward function in deep reinforcement learning and training driving cycle data, a low-level control strategy applicable to different driving conditions is obtained. This application combines changes in driving conditions with energy management strategies, which can reduce unreasonable energy allocation caused by driving condition switching, further optimize the energy planning effect of hybrid vehicles, and improve vehicle fuel economy and battery health. Attached Figure Description

[0013] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0014] Figure 1 This is a flowchart illustrating the hybrid vehicle energy management method that takes into account changes in driving conditions in this application.

[0015] Figure 2 This is an overall framework diagram of the hybrid vehicle energy management method that takes into account changes in driving conditions in this application;

[0016] Figure 3 This is a comparison chart of the holdout test results of the driving condition recognition model based on LSTM provided in this application and the recognition model developed by others on different standard driving cycles;

[0017] Figure 4 This is a comparison chart of the control results of multiple reinforcement learning agents trained with different multi-objective reward function weights under different driving conditions, provided in this application.

[0018] Figure 5 This is a comparison chart of the control results of the controller Agent-S for suburban conditions and the controller Agent-U for urban conditions, obtained through training according to this application, under different standard driving cycles.

[0019] Figure 6 This is a comparison chart of the driving condition adaptive deep reinforcement learning energy control strategy provided in this application with the control results of traditional deep learning algorithms and rule-based control strategies in a dynamic and complex environment obtained by combining multiple standard driving cycles. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that the descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0022] In the description of this application, it should be understood that the numerical labels before the steps do not indicate the order of the steps, but are only used to facilitate the description of this application and to distinguish each step, and therefore should not be construed as a limitation of this application.

[0023] First, a definition of the terminology used in this application is provided:

[0024] Glorot Normal Distribution: Also known as Xavier initialization, the Glorot normal distribution is a weight initialization method commonly used in the training of deep neural networks. This method helps the network train more effectively by initializing the network weights with a normal distribution, avoiding the vanishing or exploding gradient problems. In neural networks, Glorot initialization chooses a weights with a mean of 0 and a standard deviation of 0. Values ​​are drawn from a normal distribution, where n in and n out These are the number of input nodes and the number of output nodes in the current layer, respectively.

[0025] Adam Optimization Algorithm: Adam (Adaptive Moment Estimation) is a widely used gradient optimization algorithm that combines the advantages of Momentum and RMSprop to accelerate the training process of deep learning models. The Adam algorithm adjusts the learning rate for each parameter by calculating the first moment (the mean of the gradient) and the second moment (the mean of the squared gradients). It adaptively adjusts the learning rate and can effectively handle sparse gradients, noisy data, and non-stationary objective functions.

[0026] DQN Algorithm: DQN (Deep Q-Network) is a method in deep reinforcement learning that uses deep neural networks to approximate the Q-value function, solving the computational challenges of traditional Q-learning algorithms in high-dimensional state spaces. DQN trains a neural network to estimate the action value (Q-value), i.e., the expected reward obtained by taking a certain action in a specific state. DQN combines techniques such as experience replay and target network, enabling the algorithm to stably converge and optimize the policy in complex environments.

[0027] DDPG (Deep Deterministic Policy Gradient) is a deep reinforcement learning algorithm used to solve decision-making problems in a continuous action space. It combines deep learning with policy gradient methods, using deterministic policies instead of stochastic ones. DDPG is based on the Actor-Critic framework, where the Actor generates the action policy and the Critic evaluates the value of the current policy. It further stabilizes the learning process by using a target network and an experience replay mechanism.

[0028] TD3 Algorithm: TD3 (Twin Delayed Deep Deterministic Policy Gradient) is an improved version of the DDPG algorithm, designed to address the overestimation of Q-values ​​that occurs during DDPG training. Building upon DDPG, it: 1) uses two Q-networks to reduce Q-value overestimation; 2) reduces the update frequency of the Critic network, thereby reducing variance during training; and 3) adds noise to the target Q-value to increase training robustness.

[0029] HWFET: HWFET is a standardized driving cycle used to test the fuel economy of hybrid vehicles on highways. The HWFET test is a fuel efficiency testing method based on the US EPA (Environmental Protection Agency) standard, simulating the driving conditions of a vehicle at a certain speed on a highway. This test is used to evaluate the vehicle's fuel consumption on highways and the performance of the hybrid system.

[0030] UDDS: UDDS is another standardized driving cycle specifically designed to test a vehicle's fuel economy in urban environments, simulating driving conditions on city streets. The UDDS test includes low-speed driving, frequent acceleration and braking, reflecting common urban driving scenarios. This test is used to evaluate a vehicle's fuel consumption and emissions in urban traffic. It is often used in conjunction with HWFET to provide comprehensive performance data on hybrid vehicles under different driving conditions.

[0031] Secondly, to facilitate understanding of the technical solutions provided in the embodiments of this application by those skilled in the art, the relevant technologies are described below:

[0032] Example 1

[0033] See Figure 1 and Figure 2 The deep learning-based hybrid electric vehicle energy management method considering changes in driving conditions in this application includes:

[0034] S1. Input the vehicle speed information into the trained driving condition recognition model and output the driving condition recognition result, which is either urban driving condition or suburban driving condition.

[0035] The driving condition recognition model is built on a sequence network, using a glorot normal distribution to initialize parameters, the Adam optimization algorithm as the optimizer, and cross-entropy loss function for model training. The sequence network can be a recurrent neural network (RNN), a long short-term memory network (LSTM), or a neural network based on an attention mechanism like the Transformer.

[0036] S2. Select an energy optimization control model based on the driving condition identification results, input the driving condition identification results into the selected energy optimization control model, and output the instantaneous energy optimization control results.

[0037] Among them, the energy optimization control model achieves instantaneous energy optimization control through the energy management strategy of hybrid vehicles under different driving conditions. The energy management strategy is obtained by using deep reinforcement learning to train driving cycle data to adjust the weights of the multi-objective reward function. The multi-objective reward function is established based on fuel consumption and battery SOC changes.

[0038] Multi-objective reward functions play a crucial role in the training process of the underlying energy management controller. They not only influence the learning process of the agent (i.e., the energy management controller) but also determine how the final control strategy balances multiple optimization objectives. In deep reinforcement learning, the reward function provides feedback based on the agent's performance at each step in the environment. For the energy management problem of hybrid vehicles, multiple objectives typically need to be optimized simultaneously. For example: 1) Fuel consumption: minimizing fuel consumption to improve vehicle fuel economy. 2) SOC change: SOC is the battery's state of charge; excessive battery consumption or frequent charging can affect battery life. Therefore, reasonably controlling SOC changes and avoiding overcharging and over-discharging is also an important objective.

[0039] A simple multi-objective reward function can be expressed in the form of: R = ω fuel ·R fuel +ω SOC ·R SOC , where ω fuel and ω SOC R represents the weights of the objectives, used to control the contribution of each objective to the overall reward function. fuel and R SOC These are reward terms related to fuel consumption and SOC changes. Weight adjustment is a crucial mechanism in multi-objective reward functions during training, determining the relative importance of different objectives in the final policy. Multi-objective reward functions guide deep reinforcement learning agents to make optimal decisions during training by combining feedback from multiple objectives (such as fuel consumption and SOC changes). By designing appropriate reward functions and adjusting training data, it is possible to ensure stable performance of the energy management controller in complex and variable driving environments, optimizing vehicle energy efficiency and battery management.

[0040] The aforementioned deep reinforcement learning algorithms can be DQN, DDPG, TD3, etc.

[0041] In the technical solution of this application, the driving condition recognition model constructed by the sequence network can extract long-term dependency information in vehicle speed, thereby improving the accuracy of driving condition recognition. Since the energy optimization control model achieves instantaneous energy optimization control through energy management strategies applicable to different driving conditions, and the energy management strategy is obtained by training driving cycle data using deep reinforcement learning to adjust the weights of the multi-objective reward function, the combination of changes in driving conditions with the energy management strategy to output instantaneous energy optimization control results in real time can reduce unreasonable energy allocation caused by driving condition switching, further optimize the energy planning effect of hybrid vehicles, improve vehicle fuel economy, and extend battery life.

[0042] Example 2

[0043] The driving condition recognition model mentioned above requires pre-training. The following will use the Long Short-Term Memory (LSTM) network as an example to explain the technical solution for building and training the driving condition recognition model.

[0044] Specifically, the driving condition recognition model takes a fixed-length speed data sequence as input and outputs the probability that the current driving condition belongs to a certain category, such as the probability of suburban or urban driving conditions.

[0045] The construction of the driving condition recognition model based on the Long Short-Term Memory (LSTM) network specifically includes the following steps S1.1-S1.3, as follows:

[0046] S1.1 Data Processing: Divide the standard driving cycle into data segments based on the vehicle's start and stop points.

[0047] The number of standard driving cycles is at least two, and the data segments are of fixed length.

[0048] Specifically, urban and suburban driving conditions are used as the identification targets. The data is divided into multiple segments based on the start and stop points of the selected cycle, and the data is cleaned according to its changing characteristics. "0" and "1" are used to represent suburban and urban driving conditions, respectively. After segmenting the driving cycle data, the segmented data x can be represented as: x = [v 1+nt v 2+nt , ..., v m+nt ], where m is the segment length and t is the segment interval. n is a non-negative integer and (m+nt) is less than or equal to the total length of the driving cycle.

[0049] Specifically, in one implementation, the number of standard driving cycles can be 12, i.e., 12 standard driving cycles; the data segmentation method is as follows: using a time window of fixed length, the cleaned driving cycle is segmented, and the segment size is generally 60s to 120s.

[0050] The data comes from 12 standard driving cycles. Since vehicle speed is also an important factor in judging driving conditions, and vehicle speed is usually within a certain range, normalization is not required. Standard driving cycles often simulate more than one driving condition; therefore, they are first categorized as a whole driving cycle, then divided into smaller segments based on vehicle start and stop points, and the resulting data segments are further cleaned. A fixed time window of 90 seconds is then used, with a step size of 5 seconds, to segment the cleaned driving cycles. Training data is obtained by sampling without replacement from the obtained data.

[0051] S1.2 Model Framework Construction: The model framework for the driving condition recognition model is built based on the Long Short-Term Memory (LSTM) network.

[0052] The constructed model framework contains multiple memory units, and each memory unit uses the following three gating mechanisms to process information in the sequence:

[0053] Where f, i, and o represent the openings of the forget gate, input gate, and output gate, respectively, and W and b are the trainable matrix parameters and biases.

[0054] The expression for the implicit vector h is as follows: h t =σ(W o ·[h t-1 x t ]+b o )·tanh(C t ), where h is the implicit vector and C is the cell state containing long-term information.

[0055] Because it introduces a hidden vector h and a cell state C, LSTM is able to capture long-term dependency information contained in sequence data very well, thus achieving better results in problems that require processing long sequence data.

[0056] Furthermore, the driving condition recognition model is constructed as follows: First, an input layer receives vehicle speed data and passes the output to an LSTM layer. A ReLU activation function is used to obtain the complete output sequence, and a dropout layer is applied to alleviate overfitting. Then, the output sequence is fed back to another LSTM layer, which is also activated using the ReLU function, and another dropout layer is applied to further alleviate overfitting. This dropout layer only returns the output of the last time step. The result is then passed through a fully connected layer and activated using the sigmoid function, resulting in the final model output. This is the predicted probability. If the value is close to 1, it means that the driving condition is more likely to belong to one category, and vice versa.

[0057] S1.3 Model Training: During the training process, the parameters are initialized using the glorot normal distribution, the Adam optimization algorithm is used as the optimizer, and the cross-entropy loss function is used for training.

[0058] Understandably, initializing parameters using the glorot normal distribution can enhance model convergence, avoid gradient vanishing or exploding, and help accelerate the training process and improve stability. Using the Adam optimization algorithm as the optimizer typically outperforms traditional gradient descent methods and can adaptively adjust the learning rate of each parameter, thereby accelerating convergence, especially suitable for handling large-scale datasets and high-dimensional complex problems. Using the cross-entropy loss function to calculate the loss can effectively measure the gap between the model output and the true label.

[0059] During model training, a maximum number of training iterations is set, for example, 250 iterations, and the loss value is calculated using the following loss function:

[0060] Where N is the number of samples, and y is the true label value of the sample. i For each iteration, the true label value of the sample. This represents the predicted value output by the model in each iteration.

[0061] For example, the driving condition recognition model built and trained based on the Long Short-Term Memory (LSTM) network in this application is compared with the current mainstream driving condition recognition models developed based on fuzzy rules, clustering, and supervised learning. The recognition accuracy of various methods is shown in Table 1.

[0062] Table 1

[0063]

[0064] As can be seen from Table 1, the accuracy of the driving condition recognition model constructed and trained based on the Long Short-Term Memory (LSTM) network in this application is higher than that of the recognition models constructed by the three methods.

[0065] Furthermore, the above four models were used to conduct holdout tests on WVUSUB, WVUINTER, and WVUCITY, respectively, and the test results are as follows: Figure 3 As shown, it should be noted that Figure 3 The first row shows the speed curves for three operating conditions: WVUSUB, WVUINTER, and WVUCITY. Below are the recognition results for fuzzy rules, clustering, supervised learning, and LSTM-based models, respectively. Figure 3 The leftmost label is provided. The analysis uses the WVUSUB recognition results as an example. This driving cycle includes high-speed driving periods shorter than traditional suburban driving conditions, as well as rapidly changing low-speed ranges. It is a combined driving condition, exhibiting characteristics of both urban and suburban driving conditions. From the recognition results, the LSTM-based model identifies the high-speed driving segments of approximately 150-200 seconds and 800-1450 seconds as suburban driving conditions, while identifying the low-speed fluctuation segments of approximately 250-800 seconds and 1450-1650 seconds as urban driving conditions, demonstrating high accuracy. At the transition points between high-speed and low-speed driving segments (approximately 250 seconds and 800 seconds), the model can promptly identify the transition in driving conditions. In the short-term speed fluctuation segments that maintain suburban driving characteristics (approximately 400-600 seconds), the model maintains recognition stability by considering long-term dependencies in the speed feature sequence, thus avoiding interference from short-term speed changes causing switching in recognition results. For recognition models based on fuzzy rules and clustering, both methods identify most operational conditions as urban driving conditions, which is inconsistent with the real-world situation. For supervised learning-based models, the recognition results exhibit drastic fluctuations across multiple intervals, such as between 150 and 250 seconds and between 1200 and 1600 seconds, demonstrating low stability. The same conclusion can be drawn from the results of WVINTER and WVUCITY. In summary, from... Figure 3 As can be seen, the LSTM-based recognition model has higher recognition accuracy and better stability.

[0066] Example 3

[0067] The energy optimization control model described above mathematically models the vehicle energy management problem, resulting in deep reinforcement learning-based optimization controllers under different driving conditions. The following section uses the DDPG algorithm as an example to illustrate the modeling scheme for the energy optimization control model.

[0068] Specifically, by setting a multi-objective reward function, adjusting the balance between multiple optimization objectives based on the characteristics of different driving conditions, and finally modifying the parameters of the multi-objective reward function and the driving cycle data used in the training process, multiple deep reinforcement learning energy optimization controllers are obtained.

[0069] The construction of the energy optimization control model includes the following steps S2.1-S2.3, as detailed below:

[0070] S2.1 Hybrid Electric Vehicle Simulation Modeling: Modeling the longitudinal dynamics of the whole vehicle, engine, motor, and power battery.

[0071] Building a Hybrid Electric Vehicle Simulation Model: The hybrid electric vehicle simulation model can belong to various architectures. Taking the input power split type as an example, its powertrain architecture includes one engine and two motors, which act as an electric motor and a generator, respectively. This application does not consider vehicle steering and road gradient; the longitudinal dynamics of the entire vehicle can be modeled as follows:

[0072]

[0073] Where C D It is the air drag coefficient, A f is the cross-sectional area of ​​the car, v is the vehicle speed, m is the vehicle mass, g represents the acceleration due to gravity, and f represents the rolling resistance coefficient.

[0074] Modeling key vehicle power components: Ignoring specific implementation details and the effects of temperature, the engine is modeled as an efficiency r function of torque T and speed ω.

[0075] r f =f e (ω eng T eng ).

[0076] The power P and efficiency η of the two motors are calculated and modeled as follows:

[0077]

[0078] The battery is modeled based on equivalent circuit theory as follows:

[0079]

[0080] Where P bat This refers to the battery output power, where V represents the open-loop voltage, I represents the current, R is the battery internal resistance, and SOC is the state of charge. t It is the remaining charge at time t, and C bat The total battery capacity is a function of the power demand P during vehicle operation, expressed as a function of time (t).

[0081]

[0082] In summary, the function relating fuel consumption and SOC change with respect to t during vehicle operation is as follows:

[0083]

[0084] Improving vehicle fuel economy requires rational power distribution to comprehensively enhance the output efficiency of the engine and battery.

[0085] S2.2. The vehicle energy management problem is modeled based on the DDPG algorithm to obtain the corresponding state vector, action vector and real-time reward function.

[0086] Specifically, using the DDPG algorithm as an example, it employs two deep networks to simulate the action function and evaluation function respectively, effectively handling complex control problems in continuous action spaces. In DDPG, the action network μ and the evaluation network Q are updated at each time step t using the following formula:

[0087]

[0088] Where J represents the objective function, N is the number of samples, and θ μ and θ Q These are the parameters for the action network and the evaluation network, respectively.

[0089] Specifically, the DDPG algorithm consists of three parts: state space, action space, and reward function. In this application, the state space S = {v, acc, SOC} includes the current velocity v, the current acceleration acc, and the remaining SOC, where SOC ranges from [0,1]. Engine output power is chosen as the action space A = {P...} eng}, where P eng The output range is defined between 0kW and 56kW. Fuel consumption and battery SOC changes are used to establish the reward function. In summary, the state vector s, action vector a, and real-time reward function r are defined as follows:

[0090]

[0091] Here, α and β are constant factors affecting the reward value change due to fuel consumption and battery charge variations, respectively. Using the state vector s as input to the action network, gradually weakening noise is added to its output action a to enhance policy exploration capabilities. The action is executed, and the next state s' and the immediate reward value r returned by the environment are recorded. (s, a, s', r) is stored as a sample data point in the experience replay buffer. When there is sufficient data in the buffer, a batch of data is randomly sampled, and the target Q-value is calculated.

[0092] y t =r t +γQ'(s t+1 ,μ′(s t+1 |θ μ′ )|θ Q );

[0093] Where y represents the target Q-value, which is typically calculated as the current reward plus a discount to the Q-value of the next state. The evaluation network is updated by minimizing the error between the current Q-value and the target Q-value.

[0094] L(θ Q )=E[(Q(s,a|θ Q)-y) 2 ];

[0095] The action network is optimized using the gradient ascent algorithm, and the parameters of the target network are finally updated using a soft update method.

[0096] Where τ is the soft update parameter.

[0097] S2.3 Classify vehicle driving conditions according to the characteristics of vehicle speed and power demand changes, and adjust the weight selection of multi-objective reward function and training data in the deep reinforcement learning training process according to the characteristics of different categories.

[0098] To obtain controllers suitable for suburban and urban driving conditions respectively, different driving cycles were used as training data, and the relative importance of different optimization objectives in the final policy was adjusted by modifying the multi-objective reward function. Deep reinforcement learning algorithms were used to train models (such as energy management controllers) based on vehicle simulation models.

[0099] Specifically, S2.1 builds a vehicle simulation model, S2.2 is the specific training process and training settings (action space, state space, etc.) for a single controller, and S2.3 adjusts the training data and reward function weights based on the set driving conditions. Deep reinforcement learning is used to obtain two controllers for suburban driving conditions and urban driving conditions, respectively. During the training process, the vehicle model built in S2.1 is used as the external environment. The specific training process for a single controller is as shown in S2.2.

[0100] (1) Multi-objective reward function weight setting: Multiple controllers are trained by using different reward function weights. These controllers are then used in suburban and urban driving conditions respectively. The control results are compared to select more suitable multi-objective reward function weights for different driving conditions.

[0101] (2) Training Driving Cycle Setup: Driving cycle data representative of the corresponding driving conditions are used as training data for the deep reinforcement learning energy optimization controller. These driving conditions include suburban and urban driving conditions. Taking HWFET and UDDS as examples, these two standard driving cycles are often used to simulate suburban and urban conditions, respectively. For the instantaneous energy optimization controller used under suburban driving conditions, HWFET is set as the training driving cycle; similarly, for urban driving conditions, UDDS is used for training.

[0102] In this scheme, a controller is trained using weights from different multi-objective reward functions, and the resulting controller is applied to suburban and urban driving conditions, respectively. The control results are as follows: Figure 4As shown in (a) and (b), based on this, 1×200 were selected to train the suburban condition controller Agent-S, and 1×150 were selected to train the urban condition controller Agent-U.

[0103] Figure 5 This is a comparison chart of the control results of the controller Agent-S for suburban conditions and the controller Agent-U for urban conditions, obtained through training according to this application, under different standard driving cycles.

[0104] Multiple standard driving cycles are combined as a dynamic and complex operating condition. The hybrid electric vehicle energy management system considering changes in driving conditions proposed in this application, the traditional deep hybrid electric vehicle energy management strategy without driving condition recognition, and the rule-based strategy are applied to this operating condition, respectively. The control results are as follows: Figure 6 As shown in Table 2.

[0105] Table 2: Comparison of Results of Different Control Strategies under Complex Operating Conditions

[0106]

[0107] Combination Figure 6 As shown in Table 2, this application can stabilize the battery charge / discharge state trajectory, avoiding damage to battery health caused by overcharging / discharging or drastic fluctuations in charge level. Simultaneously, it achieves the highest MPGe (miles per gallon) value among multiple strategies, effectively improving fuel economy. Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of this application can be implemented using general-purpose computer equipment. They can be centralized on a single computer device or distributed across a network of multiple computer devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by the computer device. Furthermore, in some cases, the steps shown or described can be performed in a different order than presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of this application are not limited to any particular hardware and software combination.

[0108] It should be noted that the above are merely preferred embodiments of this application and do not limit the scope of patent protection of this application. Any equivalent structural or procedural changes made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of this application.

Claims

1. A hybrid electric vehicle energy management method that takes into account changes in driving conditions, characterized in that, include: S1. Input the vehicle speed information into the trained driving condition recognition model and output the driving condition recognition result, which is either urban driving condition or suburban driving condition. The driving condition recognition model is built on a sequence network and initializes parameters using a glorot normal distribution. It uses the Adam optimization algorithm as the optimizer and the cross-entropy loss function for model training. The sequence network includes recurrent neural networks (RNN), long short-term memory networks (LSTM), or neural networks based on the attention mechanism Transformer. When the sequence network is a Long Short-Term Memory (LSTM) network, the construction of the driving condition recognition model based on the LSTM network includes: S1.1 Data Processing: Divide the standard driving cycle into data segments based on the vehicle's start and stop points; S1.2 Model Framework Construction: The model framework for the driving condition recognition model is built based on the Long Short-Term Memory (LSTM) network. The constructed model framework includes at least two memory units, each of which uses the gating mechanism shown in the following formula: ； Where f, i, and o represent the openings of the forget gate, input gate, and output gate, respectively; W and b are the trainable matrix parameters and biases; and h is the latent vector, expressed as: C represents the cell state containing long-term information; by segmenting the driving cycle data, the segmented data x can be represented as: , where m is the segment length, t is the segment interval; n is a non-negative integer and m+nt is less than or equal to the total length of the driving cycle; S1.3 Model Training: During the training process, the parameters are initialized using the glorot normal distribution, the Adam optimization algorithm is used as the optimizer, and the cross-entropy loss function is used for training. S2. Select an energy optimization controller based on the driving condition identification result, input the driving condition identification result into the selected energy optimization controller, and output the instantaneous energy optimization control result; The energy optimization controller achieves instantaneous energy optimization control through the energy management strategy of the hybrid vehicle under different driving conditions. The energy management strategy is obtained by training the driving cycle data with a deep reinforcement learning algorithm to adjust the weights of the multi-objective reward function. The multi-objective reward function is established based on fuel consumption and battery SOC changes.

2. The hybrid electric vehicle energy management method according to claim 1, characterized in that, The standard driving cycle consists of 12 segments. The data segments are divided using a fixed-length time window, with each segment typically ranging from 60 to 120 seconds.

3. The hybrid electric vehicle energy management method according to claim 1, characterized in that, When the deep reinforcement learning algorithm is the DDPG algorithm, the construction of the energy optimization control model includes: S2.1 Hybrid Electric Vehicle Simulation Modeling: Modeling the longitudinal dynamics of the whole vehicle, engine, motor, and power battery; S2.

2. The vehicle energy management problem is modeled based on the DDPG algorithm to obtain the corresponding state vector, action vector and real-time reward function; S2.3 Classify vehicle driving conditions according to the characteristics of vehicle speed and power demand changes, and adjust the weight selection of multi-objective reward function and training data in the deep reinforcement learning training process according to the characteristics of different categories.

4. The hybrid electric vehicle energy management method according to claim 3, characterized in that, Adjusting the selection of multi-objective reward function weights specifically includes: training multiple controllers by using different reward function weights, applying these controllers to suburban and urban driving conditions respectively, comparing the control results, and selecting more suitable multi-objective reward function weights for different driving conditions; Adjusting the training data specifically includes using driving cycle data that can represent the corresponding driving conditions as training data for deep reinforcement learning.

5. The hybrid electric vehicle energy management method according to claim 3, characterized in that, The longitudinal dynamics of the vehicle can be modeled as follows: ， In the formula, It is the air drag coefficient. is the cross-sectional area of ​​the car, v is the vehicle speed; m is the vehicle mass, g represents the acceleration due to gravity, and f represents the rolling resistance coefficient. The engine is modeled as an efficiency r as a function of torque T and speed ω: , The power P and efficiency η of the two motors are calculated and modeled as follows: ， The battery is modeled based on equivalent circuit theory as follows: ， In the formula, This is the battery output power, where V represents the open-loop voltage, I represents the current, and R is the battery internal resistance. It is the remaining battery power at time t. It is the total battery capacity; Let the state vector s, action vector a, and real-time reward function r be defined as follows: ； In the formula, v represents the vehicle speed, acc represents the current acceleration, and α and β are constant factors for the change in reward value caused by changes in fuel consumption and battery charge, respectively.

6. The hybrid electric vehicle energy management method according to claim 1, characterized in that, During model training, a maximum number of training iterations is set, and the loss value is calculated using the following loss function: ； In the formula, N is the number of samples, and y is the true label value of the sample. For each iteration, the true label value of the sample. This represents the predicted value output by the model in each iteration.

7. A hybrid electric vehicle energy management system, characterized in that, The hybrid electric vehicle energy management system is used to implement the hybrid electric vehicle energy management method that takes into account changes in driving conditions as described in claim 1. The hybrid vehicle energy management system includes: a driving condition recognition layer and an instantaneous energy optimization control layer; The driving condition recognition layer is used to: input vehicle speed information into the trained driving condition recognition model and output driving condition recognition results, which are urban driving conditions or suburban driving conditions. The driving condition recognition model is built on a sequence network and initializes parameters using a glorot normal distribution. It uses the Adam optimization algorithm as the optimizer and the cross-entropy loss function for model training. The sequence network includes recurrent neural networks (RNN), long short-term memory networks (LSTM), or neural networks based on the attention mechanism Transformer. The instantaneous energy optimization control layer is used to: select an energy optimization controller based on the driving condition identification result, input the driving condition identification result into the selected energy optimization controller, and output the instantaneous energy optimization control result; The energy optimization controller achieves instantaneous energy optimization control through the energy management strategy of the hybrid vehicle under different driving conditions. The energy management strategy is obtained by training the driving cycle data with a deep reinforcement learning algorithm to adjust the weights of the multi-objective reward function. The multi-objective reward function is established based on fuel consumption and battery SOC changes.

8. The hybrid electric vehicle energy management system according to claim 7 is further configured to perform the hybrid electric vehicle energy management method as described in any one of claims 2-6 above.

Images