Electric vehicle regenerative braking distribution method based on neural network

Abstract

The invention relates to the technical field of electric vehicle control, in particular to an electric vehicle regenerative braking distribution method based on a neural network, and the method comprises the following steps: S1, collecting the state information of a vehicle in real time; s2, the collected state information is preprocessed, and a multi-dimensional feature vector X is constructed; s3, inputting the multi-dimensional feature vector into a pre-trained multi-layer perceptron neural network model, and outputting a motor braking ratio alpha between 0 and 1 by the neural network model; and S4, calculating distribution values of motor braking torque and mechanical braking torque according to the motor braking proportion alpha and the total required braking torque of the vehicle, and issuing a corresponding control instruction. According to the method, the pre-trained neural network model is used for replacing complex online optimization calculation, the requirement for the calculation power of the vehicle-mounted controller is lowered, and therefore the real-time requirement for brake control is met.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle control technology, and in particular to a neural network-based method for regenerative braking distribution in electric vehicles. Background Technology

[0002] With the increasing popularity of electric vehicles, maximizing the use of braking energy to extend driving range has become a key technology. Regenerative braking systems recover energy by converting the vehicle's kinetic energy into electrical energy during braking and storing it. Existing regenerative braking allocation strategies have the following limitations: rule-based or lookup-based methods have poor adaptability and struggle to achieve optimal balance among multiple objectives; while model predictive control (MPC)-based optimization methods offer superior performance, their complex online rolling optimization calculations place extremely high demands on the computing power of the onboard controller, making it difficult to meet millisecond-level real-time control requirements and limiting their widespread application in actual vehicles. Therefore, there is an urgent need for a regenerative braking allocation solution that can maintain near-optimal control performance while meeting the constraints of real-time onboard computation. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for regenerative braking distribution of electric vehicles that has high computational efficiency, good real-time performance, and can achieve near-optimal performance.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A neural network-based regenerative braking distribution method for electric vehicles, characterized by comprising the following steps: S1. Real-time collection of vehicle status information, including current vehicle speed v(t), battery state of charge SOC(t), motor temperature T(t), motor speed n(t), brake pedal travel or brake handle angle θ(t), and brake pedal or brake handle action rate θ'(t). S2. Preprocess the collected state information to construct a multidimensional feature vector X; S3. Input the multidimensional feature vector into a pre-trained multilayer perceptron neural network model, and output a motor braking ratio α between 0 and 1 from the neural network model; S4. Based on the motor braking ratio α and the total required braking torque of the vehicle, calculate the distribution value of the motor braking torque and the mechanical braking torque, and issue corresponding control commands.

[0006] Based on the aforementioned technical means, this invention introduces a pre-trained neural network model as the core decision-making unit, transforming the regenerative braking allocation problem into a deterministic forward computation. Since the model parameters are fixed, only one forward propagation is required during online application. Its computational load is significantly lower than methods such as model predictive control, which require online iterative optimization. This reduces the demand on the onboard controller's computing resources, enabling the use of neural networks to handle complex nonlinear relationships while meeting the real-time requirements of vehicle braking control.

[0007] Furthermore, the preprocessing in S2 includes the following steps: Based on the timestamp of brake pedal travel or brake lever angle θ(t), the current vehicle speed v(t), battery state of charge SOC(t), motor temperature T(t), motor speed n(t), and brake pedal or brake lever action rate θ'(t) data are aligned to the same time series using linear interpolation. The validity of the status information is verified, and abnormal data that exceeds the preset reasonable range is removed; The valid state information is normalized to form the multidimensional feature vector X.

[0008] Based on the aforementioned technical means, the preprocessing step in this invention first ensures, through timestamp alignment, that all state information input to the neural network corresponds to the same decision-making moment, improving the spatiotemporal consistency of the operating condition description. Secondly, the removal of outlier data prevents transient sensor malfunctions or noise from interfering with the model input, enhancing the system's robustness. Finally, normalization provides a scale-uniform input for the subsequent neural network model, which helps improve the model's training stability and final performance.

[0009] Furthermore, the normalization process maps each dimension of state information to the [0,1] interval, specifically as follows: , Where x represents the original data. and The corresponding minimum and maximum values ​​are preset; the multidimensional feature vector X is: , in, These are the normalized vehicle speed, battery SOC, motor temperature, motor speed, brake pedal travel or brake lever angle, and brake pedal or brake lever actuation rate, respectively.

[0010] Based on the aforementioned technical means, this invention employs a specific normalization formula to map sensor data with different physical dimensions and numerical ranges to a unified [0,1] interval, simplifying the data scale and making the update gradients of the weight parameters of each layer of the neural network more stable, which is beneficial to the model's learning and convergence. Simultaneously, when the output layer uses the Sigmoid activation function, its input matches the normalized data range more closely, further optimizing the model's output characteristics.

[0011] Furthermore, the multilayer perceptron neural network model includes an input layer, two hidden layers, and an output layer, wherein: The number of neurons in the input layer is consistent with the dimension of the multidimensional feature vector X; The two hidden layers contain 128 and 64 neurons respectively, and use the ReLU activation function; The output layer contains one neuron and outputs the α value using the Sigmoid activation function.

[0012] Based on the above technical means, the present invention adopts a network structure containing two hidden layers, which can effectively learn and express the complex relationship between multiple vehicle states and braking distribution ratios. This ensures that the model has the necessary learning ability while controlling the model size, so that the computational load required for forward computation of the trained network is within the range that the vehicle controller can easily handle, thereby supporting real-time control.

[0013] Furthermore, the pre-trained multilayer perceptron neural network model is obtained in the following way: Obtain the training dataset, which contains multiple sets of state information samples collected under various driving conditions and the corresponding optimal motor braking ratio label α. optimal The α optimal Calculated by model predictive control algorithm; Using the state information sample as input, the α optimal The neural network model is trained under supervised labeling.

[0014] Based on the aforementioned technical means, the key to the training method lies in using model predictive control algorithms to provide training targets for the neural network. Because model predictive control algorithms can fully consider vehicle dynamics, safety constraints, and multiple optimization objectives during the design phase, without being limited by real-time computation, their calculated braking distribution ratios possess high theoretical optimization performance. Using this as the target to train the neural network allows the final fixed network model to learn and approximate this deeply optimized decision-making pattern.

[0015] Furthermore, the supervised training uses mean squared error (MSE) as the loss function and employs the Adam optimizer to update the network parameters.

[0016] Based on the aforementioned techniques, using mean squared error as the evaluation criterion during training makes the model more sensitive to larger deviations in the prediction results. This allows it to make more cautious and accurate predictions at critical decision points related to braking safety (such as when the allocation ratio is close to its limit). Simultaneously, employing the Adam optimizer can automatically adjust the magnitude of parameter updates based on the actual situation during training. This makes the training process of the neural network smoother and more efficient, contributing to obtaining a final model with stable performance.

[0017] Furthermore, during training, the dataset is divided into training, validation, and test sets; the validation set loss is monitored during training, and an early stopping mechanism is used to prevent overfitting.

[0018] Based on the above technical means, by setting up an independent validation set during the training process to monitor the model's performance and adopting an "early stop" strategy, it is possible to effectively prevent the model from merely mechanically memorizing specific details in the training data (i.e., "overfitting"). This ensures that the trained neural network can not only handle the scenarios seen during training but also acquire the ability to cope with unseen and new driving conditions, thereby improving the reliability and adaptability of the model in actual vehicle applications.

[0019] Furthermore, after training is completed, the model performance is evaluated using a test set. When the model's mean absolute error (MAE) on the test set is lower than a preset threshold, the model is deemed qualified and can be deployed.

[0020] Based on the above technical means, after the model training is completed, a completely unused test set is used for final evaluation, and a clear mean absolute error threshold is set, providing an objective and quantitative basis for judging whether the model has reached the practical standard.

[0021] Furthermore, the model predictive control algorithm, based on the vehicle dynamics model, motor model, and battery model, solves for the optimal motor braking ratio label α by optimizing energy recovery efficiency, braking tracking accuracy, and ride comfort under the conditions of braking safety and battery charging power constraints. optimal .

[0022] Based on the aforementioned technical means, the specific settings of the Model Predictive Control (MPC) algorithm used to generate training labels are explicitly required to be based on the physical model of key vehicle components, and to include braking safety and battery safety as inviolable hard constraints. This ensures that the optimal allocation ratio calculated by MPC theoretically balances multiple requirements such as safety, recovery efficiency, and smoothness. High-quality label data is fundamental to the learning effect of the neural network and lays a solid foundation for the overall performance of the final neural network controller.

[0023] The beneficial effects achieved by this invention are as follows: This invention transforms the regenerative braking allocation problem into a deterministic forward computation by introducing a pre-trained neural network model as the core decision-making unit. Since the model parameters are fixed, only one forward propagation is required for online application. The neural network model (6-128-64-1) requires approximately (6*128+128)+(128*64+64)+(64*1+1)≈9220 multiply-accumulate operations and corresponding activation function calculations in a single forward propagation. On mainstream automotive microcontrollers, after optimization, the computation can be completed within 1ms. Its computational load is significantly lower than methods such as model predictive control that require online iterative optimization, reducing the demand on the computational resources of the automotive controller. This allows for the utilization of neural networks to handle complex nonlinear relationships while meeting the real-time requirements of vehicle braking control. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0025] The accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. To better illustrate the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings. The same or similar reference numerals correspond to the same or similar parts. The terms describing positional relationships in the drawings are for illustrative purposes only and should not be construed as limiting this patent. Detailed Implementation

[0026] It should be noted that, unless otherwise specified, the embodiments and technical features in the embodiments of this application can be combined with each other, and the detailed descriptions in the specific embodiments should be understood as explanations of the purpose of this application and should not be regarded as undue limitations on this application.

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0028] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0029] In the embodiments of this application, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium.

[0030] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0031] The technical solution of this embodiment will be described in detail below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, one embodiment of the present invention provides a complete implementation process for a neural network-based regenerative braking distribution method for electric vehicles. This method is implemented based on a data acquisition and preprocessing module, a neural network calculation module, and a braking distribution control module, which can be integrated into the vehicle's onboard electronic control unit (ECU).

[0033] S1. Real-time collection of vehicle status information.

[0034] The following real-time signals are synchronously acquired within each control cycle (e.g., 10ms) using the vehicle sensor network: Current vehicle speed v(t) from wheel speed sensor, in km / h.

[0035] Battery state of charge (SOC(t)) from the battery management system, in units of %.

[0036] Motor temperature T(t) from the motor temperature sensor, in °C.

[0037] Motor speed n(t) from the motor controller, in rpm.

[0038] Brake pedal travel θ(t) from the brake pedal position sensor, in % (0% represents fully released, 100% represents fully depressed).

[0039] The brake pedal actuation rate θ'(t), calculated in real time based on θ(t), is expressed in % / s and is used to reflect the driver's braking urgency.

[0040] S2. Preprocess the collected state information to construct a multidimensional feature vector X.

[0041] This step is performed by the data acquisition and preprocessing module, and specifically includes three sub-steps: 1. Data Synchronization and Alignment: Due to slight delays in signal transmission and processing from different sensors, in order to ensure that all data correspond to the same physical moment, the timestamp of the brake pedal signal θ(t) is used as a reference. Through linear interpolation, the data points of the current vehicle speed v(t), battery state of charge SOC(t), motor temperature T(t), motor speed n(t), and brake pedal or brake lever action rate θ'(t) are aligned to this reference time series.

[0042] 2. Anomaly Data Handling: Based on the vehicle's physical limits, reasonable ranges are set for each parameter. Data is validated in real time, and data points exceeding the range are considered outliers and removed. For consecutive anomalies, data from the previous valid period is used to fill the gaps, ensuring the continuity of the data stream.

[0043] In this embodiment, the preset ranges for each sensor are as follows: Effective speed range: 0 ≤ v ≤ 200 km / h; Effective SOC range: 10% ≤ SOC ≤ 100%; Effective temperature range for the motor: -20°C ≤ T ≤ 130°C; Effective range of motor speed: 0≤n≤12000rpm; Effective range of brake pedal travel / brake handle angle: 0%≤θ≤100%; Effective range of brake pedal / handle actuation rate: -500% / s ≤ θ' ≤ 500% / s.

[0044] 3. Feature Normalization: The cleaned six-dimensional raw data are normalized and mapped to the [0,1] interval. For example, for vehicle speed v, vspeed is set to... min =0, v max =200, then the normalized value v norm =v / 200. Similar operations are applied to parameters such as SOC and T. Finally, a six-dimensional normalized feature vector X is constructed: .

[0045] S3. Input the multidimensional feature vector into a pre-trained multilayer perceptron neural network model, and output a motor braking ratio α between 0 and 1 from the neural network model.

[0046] This step is performed by the neural network computation module. This module contains a pre-trained multilayer perceptron model. Its structure is configured as follows: Input layer: 6 neurons, each corresponding to a preprocessed 6-dimensional feature vector.

[0047] Hidden layers: There are two layers in total. The first hidden layer has 128 neurons, and the second hidden layer has 64 neurons. Both layers use the ReLU (Rectified Linear Unit) function as the activation function to introduce non-linearity and accelerate training. Specifically: Each neuron in the first hidden layer performs a linear transformation on the input neuron: , in Let be the linear output value of the j-th hidden neuron (j=1,2,...,128). The connection weights are the connection weights from the i-th node in the input layer to the j-th neuron in the first hidden layer. This represents the bias term for the j-th neuron in the first hidden layer. After a linear transformation, the ReLU activation function is applied to the linear output of each neuron. Activate: , in The final output value of the j-th neuron in the first hidden layer forms a 128-dimensional feature vector. .

[0048] The second hidden layer takes the output A1 of the first hidden layer as input, and performs linear transformation and ReLU activation in the same way as the first hidden layer. The result of the linear transformation is: , in This represents the linear output value of the k-th neuron in the second hidden layer (k=1,2,...,64). Let be the connection weight from the j-th neuron in the first hidden layer to the k-th neuron in the second hidden layer. This is the bias term for the k-th neuron in the second hidden layer.

[0049] Linear output for each neuron ReLU activation is used: , in The final output value of the k-th neuron in the second hidden layer forms a 64-dimensional feature vector. .

[0050] Output layer: 1 neuron, using the sigmoid function as the activation function, restricting its output value to the interval [0,1]. Specifically: The output layer takes the output A2 of the second hidden layer as input and performs a linear transformation: , in The linear output value of the output neuron. The connection weights from the k-th neuron in the second hidden layer to the output neuron are... This is the bias term for the output neuron. For linear output... Activation is performed using the Sigmoid activation function: , Where α is the final output motor braking ratio, i.e., the motor braking ratio, and its value is a continuous value between [0,1].

[0051] α=1 indicates that the requested braking force is entirely provided by regenerative braking of the motor; α=0 indicates that it is entirely provided by mechanical friction braking; 0<α<1 indicates that the two are coordinated braking in proportion.

[0052] S4. Based on the motor braking ratio α and the total required braking torque of the vehicle, calculate the distribution value of the motor braking torque and the mechanical braking torque, and issue corresponding control commands.

[0053] This step is performed by the brake distribution control module.

[0054] Total required braking torque calculation: Based on the brake pedal travel θ(t), a preset pedal feel curve is used, which defines the relationship between the brake pedal travel percentage θ(t) and the vehicle's total required braking torque F. total The mapping relationship between these parameters is determined through calibration to ensure that the driver's braking feel matches expectations. This method yields the total required braking torque F of the vehicle requested by the driver at the current moment. total .

[0055] Based on the neural network output α, the following allocation is made: Electric motor torque F motor =α×F total , Mechanical braking torque F friction =(1-α)×F total , F motor As a torque command, it is sent to the motor controller via the CAN bus. The motor controller then controls the motor to operate in generator mode, producing the corresponding braking torque and feeding the generated electrical energy back to the battery. (The F...) friction As a pressure or torque command, it is sent to the controller (such as ESC) of the hydraulic or pneumatic braking system to control the friction brake to generate the corresponding mechanical braking torque.

[0056] Through the above process, real-time and optimized coordination of regenerative braking and mechanical braking is achieved.

[0057] It should be noted that the neural network model of this invention is pre-trained, which avoids the heavy online optimization burden of algorithms such as MPC. Its training process (offline phase) is as follows: 1. Training Dataset Construction: In a simulation environment or on a real-vehicle data acquisition platform, run driving cycles covering various typical and extreme conditions, including urban roads, suburbs, highways, and long downhill slopes. At each simulation or sampling time point, record the vehicle's real-world state information (i.e., v(t), SOC(t), T(t), n(t), θ(t), θ'(t)). Simultaneously, run a high-precision Model Predictive Control (MPC) optimizer. This MPC optimizer incorporates simplified dynamic models of the vehicle, motor, and battery, with the primary objective of maximizing energy recovery efficiency, while strictly constraining conditions such as braking deceleration tracking accuracy, battery charging power, motor torque limits, and braking stability. At each time point, the MPC solves a rolling time-domain optimization problem, outputting the theoretically optimal motor braking ratio α for the current time point. optimal The feature vector X obtained by preprocessing the collected raw state data in the same way as the online process, and the corresponding α optimal Pairing is performed to form a massive number of training samples (X, α) optimal This constitutes the training dataset.

[0058] 2. Supervised Model Training: The dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio. An MLP network with the aforementioned structure (6-128-64-1) is initialized. During training, mean squared error (MSE) is used as the loss function to measure the difference between the network's predicted value α and the MPC label α. optimal The difference between the two was investigated. The Adam optimizer (initial learning rate set to 0.001, β1=0.9, β2=0.999, β1 and β2 being the decay rates estimated for the first and second moments, respectively) was used to iteratively update the network weights and biases to minimize the loss. During training, the batch size was set to 64. To prevent overfitting, L2 weight decay (coefficient 0.0001) and an early stopping mechanism were introduced—that is, the model's loss on the independent validation set was continuously monitored; if it did not decrease for 10 consecutive training epochs, training was stopped, and the model parameters that minimized the loss on the validation set were rolled back.

[0059] 3. Model Validation and Consolidation: The best-performing trained model is evaluated using a reserved test set. The mean absolute error (MAE) of the model on the test set is calculated. When the MAE is below a preset strict threshold (e.g., 0.05), the model accuracy is considered to meet the requirements for engineering applications. At this point, the final structure (number of layers, number of nodes) and all parameters (weights, biases) of the model are consolidated and compiled and deployed to the neural network computing module of the target vehicle ECU for use in the online phase.

[0060] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for regenerative braking distribution in electric vehicles based on neural networks, characterized in that, Includes the following steps: S1. Real-time collection of vehicle status information, including current vehicle speed v(t), battery state of charge SOC(t), motor temperature T(t), motor speed n(t), brake pedal travel or brake handle angle θ(t), and brake pedal or brake handle action rate θ'(t). S2. Preprocess the collected state information to construct a multidimensional feature vector X; S3. Input the multidimensional feature vector into a pre-trained multilayer perceptron neural network model, and output a motor braking ratio α between 0 and 1 from the neural network model; S4. Based on the motor braking ratio α and the total required braking torque of the vehicle, calculate the distribution value of the motor braking torque and the mechanical braking torque, and issue corresponding control commands.

2. The method for regenerative braking distribution of electric vehicles based on neural networks according to claim 1, characterized in that, The preprocessing in S2 includes the following steps: Based on the timestamp of brake pedal travel or brake lever angle θ(t), the current vehicle speed v(t), battery state of charge SOC(t), motor temperature T(t), motor speed n(t), and brake pedal or brake lever action rate θ'(t) data are aligned to the same time series using linear interpolation. The validity of the status information is verified, and abnormal data that exceeds the preset reasonable range is removed; The valid state information is normalized to form the multidimensional feature vector X.

3. The method for regenerative braking distribution of electric vehicles based on neural networks according to claim 2, characterized in that, The normalization process maps each dimension of state information to the [0,1] interval, specifically as follows: ， Where x represents the original data. and These are the preset minimum and maximum values ​​for the corresponding data.

4. The method for regenerative braking distribution of electric vehicles based on neural networks according to claim 3, characterized in that, The multidimensional feature vector X is: ， in, These are the normalized vehicle speed, battery SOC, motor temperature, motor speed, brake pedal travel or brake lever angle, and brake pedal or brake lever actuation rate, respectively.

5. A neural network-based regenerative braking distribution method for electric vehicles according to claim 1 or 4, characterized in that, The multilayer perceptron neural network model includes an input layer, two hidden layers, and an output layer, wherein: The number of neurons in the input layer is consistent with the dimension of the multidimensional feature vector X; The two hidden layers contain 128 and 64 neurons respectively, and use the ReLU activation function; The output layer contains one neuron and outputs the α value using the Sigmoid activation function.

6. The method for regenerative braking distribution of electric vehicles based on neural networks according to claim 5, characterized in that, The pre-trained multilayer perceptron neural network model is obtained through the following method: Obtain the training dataset, which contains multiple sets of state information samples collected under various driving conditions and the corresponding optimal motor braking ratio label α. optimal The α optimal Calculated by model predictive control algorithm; Using the state information sample as input, the α optimal The neural network model is trained under supervised labeling.

7. The method for regenerative braking distribution of electric vehicles based on neural networks according to claim 6, characterized in that, The supervised training uses mean squared error (MSE) as the loss function and employs the Adam optimizer to update the network parameters.

8. The method for regenerative braking distribution of electric vehicles based on neural networks according to claim 6, characterized in that, During training, the dataset is divided into training, validation, and test sets; the validation set loss is monitored during training, and an early stopping mechanism is used to prevent overfitting.

9. A neural network-based regenerative braking distribution method for electric vehicles according to claim 8, characterized in that, After training is completed, the model performance is evaluated using a test set. When the model's mean absolute error (MAE) on the test set is lower than a preset threshold, the model is deemed qualified and can be deployed.

10. A neural network-based regenerative braking distribution method for electric vehicles according to claim 6, characterized in that, The model predictive control algorithm is based on the vehicle dynamics model, motor model, and battery model. Under the conditions of satisfying braking safety and battery charging power constraints, it optimizes energy recovery efficiency, braking tracking accuracy, and ride comfort to obtain the optimal motor braking ratio label α. optimal .

Images