Output power control method and vehicle

Abstract

The invention discloses an output power control method and a vehicle, and belongs to the technical field of vehicle control. Through the technical scheme provided by the embodiment of the invention, the problem that thermal damage cannot be prevented due to intervention lag in the related technology is solved, active thermal management is realized by predicting the temperature change trend in advance, and power limitation can be dynamically adjusted before the temperature sudden rise risk occurs; and meanwhile, excessive limitation of power output caused by conservative protection allowance is avoided, and unnecessary power limitation is reduced under the low-load working condition, so that the adaptive capacity of the electric drive system under the complex working condition is improved, and balance of thermal safety and power performance is considered.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and more specifically, to a method for controlling output power and a vehicle in the field of vehicle control technology. Background Technology

[0002] As the core power source of electric vehicles, the effectiveness of the electric drive system's thermal management directly affects the performance and reliability of the entire vehicle.

[0003] Currently, the industry generally adopts a power limiting strategy based on a fixed temperature threshold for motor overheat protection. When the motor temperature exceeds the preset threshold, the electric drive system triggers power reduction protection. This passive response mechanism has inherent defects: it is delayed in intervention under conditions of sudden temperature rise, and may not be able to effectively prevent thermal damage; while under fluctuating thermal load, it may lead to excessive limitation of power output due to conservative protection margin settings, affecting driving experience and transportation efficiency.

[0004] Therefore, there is an urgent need for an output power control method that can improve the adaptability of electric drive systems under complex operating conditions. Summary of the Invention

[0005] This application provides a method for controlling output power and a vehicle, which can improve the adaptability of electric drive systems under complex operating conditions. The technical solution is as follows: On the one hand, a method for controlling output power is provided, the method comprising: When the target vehicle is in motion, based on the vehicle status parameter set of the target vehicle, the motor operation parameter set of the drive motor, and the environmental parameter set of the environment, the predicted temperature change trend of the drive motor and the load level of the target vehicle are determined. Based on the predicted temperature change trend, the load level, and a subset of motor state parameters in the set of motor operating parameters, a power limiting factor for the target vehicle is determined, which is used to limit the output power of the drive motor. The output power of the drive motor is limited based on the power limiting factor and the initial power limit of the drive motor.

[0006] On the one hand, an output power control device is provided, the device comprising: The trend level determination module is used to determine the predicted temperature change trend of the drive motor and the load level of the target vehicle based on the vehicle state parameter set of the target vehicle, the motor operation parameter set of the drive motor, and the environmental parameter set of the environment when the target vehicle is in motion. The factor determination module is used to determine the power limiting factor of the target vehicle based on the predicted temperature change trend, the load level, and a subset of motor state parameters in the set of motor operating parameters. The power limiting factor is used to limit the output power of the drive motor. A power limiting module is used to limit the output power of the drive motor based on the power limiting factor and the initial power limit of the drive motor.

[0007] On one hand, a vehicle is provided, the vehicle including one or more processors and one or more memories, the one or more memories storing at least one piece of program code, the program code being loaded and executed by the one or more processors to implement the operations performed by the output power control method.

[0008] On one hand, a computer-readable storage medium is provided, wherein at least one piece of program code is stored in the computer-readable storage medium, the program code being loaded and executed by a processor to implement the operations performed by the output power control method.

[0009] The technical solution provided by the embodiments of this application solves the problem of thermal damage that cannot be prevented due to delayed intervention in related technologies. By predicting the temperature change trend in advance, active thermal management is achieved, which can dynamically adjust the power limit before the risk of a sudden temperature rise occurs. At the same time, it avoids excessive power output restriction caused by conservative protection margin, reduces unnecessary power restriction under low load conditions, thereby improving the adaptability of the electric drive system under complex conditions and balancing thermal safety and power performance. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the implementation environment of an output power control method provided in an embodiment of this application; Figure 2 This is a flowchart of an output power control method provided in an embodiment of this application; Figure 3 This is a flowchart of another output power control method provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an output power control device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application. Detailed Implementation

[0011] The technical solutions in this application will be clearly and thoroughly described below with reference to the accompanying drawings. In the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of the embodiments of this application, "multiple" refers to two or more than two.

[0012] In the following text, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features reflected. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0013] In order to illustrate the technical solutions provided in the embodiments of this application, some terms involved in the embodiments of this application will be introduced below.

[0014] Multi-Task Learning (MTL) is a machine learning paradigm that aims to enable a single model to learn how to solve multiple related tasks simultaneously. Its core idea is to leverage shared information and correlations between different tasks by sharing a subset of representations (features) or parameters within the model to improve its generalization ability and overall performance across all tasks. Compared to training a separate model for each task, MTL is generally more efficient (in terms of computational and storage resources) and achieves better results, especially when data for individual tasks is limited.

[0015] Gated Networks: These are a class of neural network structures whose core feature is the inclusion of gating units / mechanisms. These gating units are essentially learnable switches (typically implemented by functions like the sigmoid function, outputting values ​​between 0 and 1) that dynamically and selectively control the flow of information within the network. The gating mechanism determines which information should be retained, passed on, updated, or forgotten. Typical examples include Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), which are the most well-known examples of gated networks. They address the long-term dependency problem in recurrent neural networks through input gates, forget gates, and output gates (LSTM) or update gates and reset gates (GRUs).

[0016] State space dimensionality reduction refers to the process or technique of transforming (mapping) a high-dimensional state representation (containing a large number of features or variables) to a lower-dimensional space. The goal is to reduce the dimensionality of the data while preserving the most important information in the original state space (such as the ability to distinguish different states, key dynamic characteristics of an electric drive system). The aim is to reduce computational complexity, reduce storage requirements, mitigate the curse of dimensionality, remove redundancy and noise, visualize high-dimensional data, and reveal underlying structure or principal components.

[0017] Activation function: A mathematical function in a neural network that acts on the output of a neuron (or node). It receives a weighted sum of the inputs from the neurons in the previous layer (usually with a bias term added) and determines the output value (activation value) of that neuron, which is then passed to the next layer in the network. Its core function is to introduce nonlinearity. Without a nonlinear activation function, no matter how many layers are stacked, the entire network can only represent linear transformations and cannot learn complex patterns. Activation functions enable neural networks to approximate arbitrarily complex nonlinear functions.

[0018] Loss Function (or Cost Function): In machine learning and deep learning, the loss function is a mathematical function that quantifies the difference (error) between the model's predicted output and the true target value (label). It measures the model's performance on a single data point or the entire training dataset. Its core function is to provide an optimization objective; the goal of model training is to minimize the value of the loss function. It guides learning algorithms (such as gradient descent) on how to adjust the model's parameters to make the predictions increasingly closer to the true values.

[0019] Learning rate: In optimization algorithms (especially those based on gradient descent and its variants, such as SGD and Adam), the learning rate is a hyperparameter. It controls the step size at which the update amount, calculated from the gradient of the loss function, is applied to adjust the model parameters in each parameter update iteration. Its core role is to control convergence speed and stability; the learning rate directly affects the speed of model training and whether a good solution can be found.

[0020] In the thermal management of electric drive systems, relevant technologies generally employ power limiting strategies based on fixed temperature thresholds. When the drive motor temperature exceeds a preset threshold, the electric drive system triggers a power reduction protection mechanism. This mechanism has inherent flaws: under conditions of sudden temperature rise, because it relies solely on real-time temperature monitoring, the intervention lags behind the evolution of the thermal state, failing to effectively prevent thermal damage; and in scenarios of fluctuating thermal loads, the conservative protection margin settings lead to excessive limitation of power output. The risk of thermal damage stems from the unpredictability of temperature change trends, while power output limitation affects key vehicle performance indicators, including power continuity, electric drive system reliability, and driving smoothness. Furthermore, the essence of this problem lies in the lack of forward-looking assessment capabilities in the thermal management strategy, making it unable to dynamically adapt to changes in thermal loads under complex operating conditions.

[0021] For example, in a high-temperature mountainous continuous uphill scenario, the target vehicle carries a high load and maintains a high speed. At this time, the ambient temperature is high, the vehicle status parameter set shows that the load rate exceeds a threshold, and the torque demand in the motor operating parameter set continues to increase. Because existing strategies only respond based on real-time temperature values, the electric drive system only activates power reduction protection when the motor temperature rises sharply due to heat accumulation and exceeds a fixed threshold, but thermal damage may have already occurred. Simultaneously, in subsequent downhill conditions, the load level decreases, but the electric drive system, due to conservative settings, still imposes unnecessary restrictions on output power, leading to power output interruption and a deterioration in driving experience. Specifically, the rapid fluctuations in thermal load in this scenario make the fixed threshold mechanism unable to distinguish between instantaneous overheating risk and sustainable operating conditions, causing a disconnect between protection actions and actual thermal conditions.

[0022] If the above problems are not resolved, the adaptability of the electric drive system under complex operating conditions will be limited. The continued presence of the risk of thermal damage may lead to accelerated aging of the insulation materials of the drive motor and a decrease in the reliability of the electric drive system; excessive power limitation will cause discontinuous power output, affecting the stability of the overall vehicle performance; thus, the electric drive system cannot meet the dual requirements of thermal safety and power performance in actual operation, thereby restricting the overall efficiency of electric vehicles.

[0023] The implementation environment of the embodiments of this application is described below. See also... Figure 1 The implementation environment of the output power control method provided in this application embodiment includes a vehicle controller 101 and an electric drive system 102.

[0024] The vehicle controller 101 is a data processing component installed on the target vehicle. The vehicle controller 101 can acquire and process relevant information. In this embodiment, the vehicle controller 101 can acquire a set of vehicle status parameters, a set of motor operating parameters, and a set of environmental parameters, and process these parameters to obtain a power limiting factor for power limiting. The vehicle controller 101 is communicatively connected to the electric drive system 102. The vehicle controller 101 can send the processed power limiting factor to the electric drive system 102 so that the electric drive system 102 can use the power limiting factor to limit the power of the drive motor.

[0025] The electric drive system 102 is used to control the drive motor, such as controlling the output power, speed and torque of the drive motor. In this embodiment, it is particularly used to limit the output power of the drive motor.

[0026] After introducing the implementation environment of the embodiments of this application, the application scenarios of the technical solutions provided by the embodiments of this application will be introduced below.

[0027] The technical solution provided in this application can be applied to any vehicle equipped with a drive motor and driven by the drive motor. By using the technical solution provided in this application, the power limiting factor of the vehicle can be determined using multi-dimensional parameters, and the output power of the drive motor can be limited using the power limiting factor to prevent thermal failure of the drive motor as much as possible.

[0028] After introducing the implementation environment and application scenarios of the embodiments of this application, the technical solutions provided by the embodiments of this application are described below. (See also...) Figure 2 Taking the vehicle controller as the executing entity as an example, the method includes the following steps.

[0029] 201. When the target vehicle is in motion, the vehicle controller determines the predicted temperature change trend of the drive motor and the load level of the target vehicle based on the vehicle status parameter set of the target vehicle, the motor operation parameter set of the drive motor, and the environmental parameter set of the environment.

[0030] The target vehicle refers to the vehicle to be regulated, which performs power control decisions. The target vehicle is equipped with an electric drive system and requires thermal risk management; it can be an electric or hybrid vehicle. The driving state refers to the operating mode where the target vehicle's drive motor is in torque output condition, at which point the target vehicle is driven by the drive motor. The vehicle state parameter set includes time-series data describing the vehicle's dynamic behavior; the motor operating parameter set includes physical quantity monitoring data during drive motor operation; and the environmental parameter set includes physical quantities of the vehicle's external environment. Predicting temperature change trends refers to the estimation process of future temperature changes in the drive motor, primarily to achieve a forward-looking assessment of the motor's thermal state and avoid passive responses relying solely on real-time temperature monitoring. In this embodiment, the predicted temperature change trend can be represented by a future temperature sequence. The load level refers to a quantitative indicator of the target vehicle's current workload, primarily to differentiate power demand characteristics under different operating conditions and reduce unnecessary power limitations.

[0031] 202. Based on the predicted temperature change trend, the load level, and the subset of motor state parameters in the set of motor operating parameters, the vehicle controller determines the power limiting factor of the target vehicle, which is used to limit the output power of the drive motor.

[0032] The power limiting factor is a coefficient used to adjust the upper limit of the drive motor's output power. The motor state parameter subset is a part of the motor operating parameter set, including parameters that reflect the operating state of the drive motor.

[0033] 203. The vehicle controller limits the output power of the drive motor based on the power limiting factor and the initial power limit of the drive motor.

[0034] The initial power limit refers to the theoretical maximum allowable power value dynamically calculated by the electric drive system based on the current operating conditions.

[0035] In steps 201-203 above, by integrating the determination mechanism of predicted temperature change trend, load level and power limiting factor, the transformation from fixed threshold response to dynamic predictive control is realized. This allows for early identification of thermal risks and optimization of power allocation under the target vehicle's driving state, avoiding intervention lag problems under sudden temperature rise conditions. At the same time, the protection margin is adjusted according to load characteristics to prevent excessive limitation of power output, thereby improving the adaptability of the electric drive system under complex operating conditions.

[0036] Furthermore, when the target vehicle is in motion, the predicted temperature change trend of the drive motor and the load level of the target vehicle are determined based on a set of vehicle state parameters, a set of motor operating parameters, and a set of environmental parameters. That is, the determination of the predicted temperature change trend utilizes multi-source data fusion technology to achieve a forward-looking assessment of the drive motor's thermal state, thus avoiding a passive response mechanism that relies solely on real-time temperature monitoring. The determination of the load level incorporates parameters such as vehicle speed and load, enabling the system to distinguish between high and low load conditions, thereby providing a basis for subsequent power limiting. The power limiting factor is determined based on the predicted temperature change trend, load level, and a subset of motor state parameters. This power limiting factor is used to limit the output power of the drive motor. The process of determining the power limiting factor deeply correlates the predicted thermal risk with the real-time state of the motor, thereby reflecting the current thermal load state and avoiding the problems of over-limiting due to a deteriorating predicted trend or insufficient output when the trend is stable.

[0037] The technical solution provided by the embodiments of this application solves the problem of thermal damage that cannot be prevented due to delayed intervention in related technologies. By predicting the temperature change trend in advance, active thermal management is achieved, which can dynamically adjust the power limit before the risk of a sudden temperature rise occurs. At the same time, it avoids excessive power output restriction caused by conservative protection margin, reduces unnecessary power restriction under low load conditions, thereby improving the adaptability of the electric drive system under complex conditions and balancing thermal safety and power performance.

[0038] It should be noted that steps 201-203 above are a simplified explanation of the output power control method provided in the embodiments of this application. The output power control method provided in the embodiments of this application will be described in more detail below with some examples. See [link to relevant documentation]. Figure 3 Taking the vehicle controller as the executing entity as an example, the method includes the following steps.

[0039] 301. When the target vehicle is in motion, the vehicle controller acquires the vehicle status parameter set, the motor operation parameter set of the drive motor, and the environmental parameter set of the surrounding environment.

[0040] The target vehicle refers to the vehicle to be regulated, which performs power control decisions. The target vehicle is equipped with an electric drive system and requires thermal risk management; it is either an electric vehicle or a hybrid vehicle. The driving state refers to the operating mode where the target vehicle's drive motor is in torque output condition, at which point the target vehicle is driven by the drive motor. The vehicle state parameter set includes time-series data describing the vehicle's dynamic behavior; the motor operating parameter set includes physical quantity monitoring data during drive motor operation; and the environmental parameter set includes physical quantities of the vehicle's external environment. The vehicle state parameter set includes vehicle speed, vehicle load, and driver operation information, including accelerator pedal opening and brake pedal opening. The motor operating parameter set includes motor temperature, motor torque, motor speed, and motor electrical parameters. The environmental parameter set includes altitude and ambient temperature. The subset of motor state parameters includes motor speed and motor electrical parameters, including motor current and motor voltage. Predicting temperature change trends refers to the estimation process of future temperature changes in the drive motor, primarily to achieve a forward-looking assessment of the motor's thermal state, avoiding passive responses relying solely on real-time temperature monitoring. In this embodiment, the predicted temperature change trend can be represented by a future temperature sequence. Load level refers to a quantitative indicator of the current working load of the target vehicle. It is mainly used to distinguish the power demand characteristics under different operating conditions and reduce unnecessary power limitations.

[0041] In one possible implementation, when the target vehicle is in motion, the vehicle controller acquires the target vehicle's speed, vehicle load, motor temperature, motor torque, motor speed, motor current, motor voltage, altitude, and ambient temperature.

[0042] The motor temperature includes stator winding temperature, rotor temperature, motor housing temperature, and IGBT (Insulated Gate Bipolar Transistor) module temperature, comprehensively reflecting the temperature status of the drive motor. Vehicle speed is a quantified indicator of vehicle speed, reflecting the impact of driving conditions on motor cooling efficiency; vehicle load is a dynamic representation of the vehicle's total mass, directly correlated with actual load size to avoid misjudgment of load level; motor temperature is a direct feedback parameter of the drive motor's thermal state, providing real-time data on thermal status; motor torque is a physical quantity representing the motor's output rotational force, characterizing the dynamic relationship between mechanical load and heat generation rate; motor speed is a quantified value of the motor's rotational speed, correlated with changes in mechanical load; motor electrical parameters are a set of parameters related to the motor's electrical characteristics, used to quantify electrical heat loss; altitude refers to the vertical height of the vehicle's geographical location, considering the impact of air density changes on heat dissipation efficiency; ambient temperature is a measured value of the surrounding air temperature, directly correlated with external heat dissipation conditions; motor current is the current flowing through the motor windings, used to quantify heat loss based on the Joule heating effect; and motor voltage is the potential difference applied to the motor terminals, supporting the completeness of power calculations.

[0043] For example, the vehicle controller determines the target vehicle's speed by real-time acquisition of wheel speed using magnetoelectric wheel speed sensors mounted on the drive wheels. The vehicle controller estimates the vehicle's load based on the filtered output signal from the suspension airbag pressure sensor. The vehicle controller acquires the accelerator pedal and brake pedal openings using pedal sensors. The vehicle controller monitors the motor temperature using a temperature sensor on the drive motor. The vehicle controller calculates the motor torque using a Kalman filter algorithm built into the motor controller combined with the motor current. The vehicle controller obtains the motor speed by decoding the sine and cosine signals output from the resolver. The vehicle controller measures the motor current and voltage using an isolated Hall effect current sensor and a resistive voltage divider network, respectively. The vehicle controller outputs altitude data after temperature compensation from a barometric pressure sensor. The vehicle controller measures the ambient temperature using a temperature sensor located outside the vehicle body.

[0044] 302. The vehicle controller determines the predicted temperature change trend of the drive motor and the load level of the target vehicle based on the vehicle status parameter set of the target vehicle, the motor operation parameter set of the drive motor, and the environmental parameter set of the environment.

[0045] The power limiting factor is a coefficient used to adjust the upper limit of the drive motor's output power. The motor state parameter subset is a part of the motor operating parameter set, including parameters that reflect the operating state of the drive motor.

[0046] In one possible implementation, the vehicle controller extracts features from the vehicle state parameter set, the motor operating parameter set, and the environmental parameter set using a first prediction model to obtain the vehicle state features of the target vehicle. The first prediction model is obtained through multi-task joint training. The vehicle controller then uses this first prediction model to regress the vehicle state features, obtaining multiple predicted temperatures for the drive motor. These multiple predicted temperatures represent the predicted temperature change trend of the drive motor. Finally, the vehicle controller uses the first prediction model to classify the vehicle based on the vehicle state features to obtain the load level of the target vehicle.

[0047] The first prediction model refers to a machine learning framework used to simultaneously perform feature extraction, temperature trend prediction, and load level determination. It can be implemented as a neural network, eliminating data silos in multi-task processing through a unified model architecture. In this embodiment, the first preset model is obtained through multi-task joint training, where multi-tasks refer to the temperature trend prediction task and the load level determination task. Multi-task joint training can be understood as simultaneously minimizing the temperature prediction regression loss and the load classification cross-entropy loss during model optimization, enabling the underlying feature representation to adapt to both dynamic temperature modeling and load state identification, thereby enhancing the model's generalization ability to complex operating conditions. Vehicle state features refer to a comprehensive representation vector that integrates vehicle dynamic behavior, motor operating status, and environmental influences. This can be implemented using high-dimensional embedding, aiming to construct a cross-domain parameter correlation expression to support subsequent prediction tasks. Multiple predicted temperatures can be understood as discrete sequence values ​​representing the evolution of drive motor temperature over time. By quantifying the temperature change rate and inflection point features, the temperature change trend can be reflected, thus providing a dynamic basis for thermal risk prediction. Load level refers to the discretized classification result that reflects the current load intensity of the target vehicle. It can be implemented using a multi-level classification system. The purpose is to transform the continuous load state into an operable decision input to match the execution requirements of the power limiting strategy.

[0048] In the above implementation, a first prediction model with a unified prediction framework is constructed through a multi-task joint training mechanism. This model integrates the dispersed sets of vehicle state parameters, motor operating parameters, and environmental parameters into vehicle state features. These vehicle state features serve as the shared input to the first prediction model, simultaneously driving temperature trend prediction and load level determination. During the feature extraction stage, the first prediction model performs cross-domain correlation processing on multi-source heterogeneous parameters to form shared features (vehicle state features) that provide a complete contextual representation of the vehicle's operating conditions. In the prediction stage, the shared features are respectively input into the regression branch to generate a temperature change trend sequence and the classification branch to output the load level. Knowledge transfer is achieved between tasks through gradient constraints during joint training, ensuring that the predicted temperature change trend reflects the thermal effect of load changes, and that the load level determination also considers the dynamic impact of temperature evolution.

[0049] The above scheme can effectively improve the timing accuracy of temperature change trend prediction and the working condition adaptability of load level judgment, so that the subsequently determined power limiting factor can match the actual thermal risk evolution process, avoid excessive suppression of power output or protection lag due to prediction deviation, and thus maintain the stability of vehicle power performance while ensuring the thermal safety of drive motor.

[0050] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0051] In the first part, the vehicle controller extracts features from the set of vehicle state parameters, the set of motor operating parameters, and the set of environmental parameters using a first prediction model to obtain the vehicle state features of the target vehicle.

[0052] In one possible implementation, the vehicle controller performs time alignment and preprocessing on the parameters in the vehicle state parameter set, the parameters in the motor operating parameter set, and the parameters in the environmental parameter set to obtain a first input parameter sequence. The vehicle controller inputs this first input parameter sequence into the first prediction model, and through the first prediction model, performs bidirectional temporal encoding on the first input parameter sequence to obtain the vehicle state features of the target vehicle. Alternatively, the vehicle controller uses the first prediction model to perform multiple convolutions or multiple fully connected operations on the first input parameter sequence to obtain the vehicle state features of the target vehicle.

[0053] Among these, time alignment refers to dynamic synchronization based on the timestamp information of each parameter to avoid data misalignment caused by differences in sensor sampling frequencies. Preprocessing refers to normalization and denoising based on parameter distribution characteristics to eliminate dimensional differences and random interference. Bidirectional temporal coding refers to feature fusion based on the forward and backward temporal dependencies of the sequence, which can be implemented using bidirectional gated recurrent unit networks, bidirectional LSTM, or bidirectional Transformer encoders to comprehensively capture the dynamic evolution of vehicle state parameters during driving. Convolution refers to hierarchical feature extraction based on local parameter patterns, which can be implemented using dilated convolution or depthwise separable convolution operations. Fully connected refers to deep modeling based on global nonlinear relationships, which can be implemented using multilayer perceptrons with residual connections to adapt to complex interaction scenarios and enhance feature representation capabilities.

[0054] In the above implementation, the parameters of the vehicle state parameter set, motor operating parameter set, and environmental parameter set are first time-aligned and preprocessed to form a first input parameter sequence, ensuring strict alignment of parameters in the time dimension and reducing noise levels. Subsequently, this first input parameter sequence is input into a first prediction model, and features are extracted through a bidirectional temporal coding mechanism or convolution / fully connected operations. Bidirectional temporal coding utilizes historical state information for bidirectional correlation, while convolution focuses on local pattern recognition, and fully connected operations model global nonlinear relationships. This process, by standardizing input data processing and feature extraction mechanisms, enables vehicle state features to reflect key thermal risk information, thereby providing highly reliable input for temperature prediction and load level classification.

[0055] Under the above implementation, time alignment and preprocessing avoid input data timing misalignment and noise interference, improve the accuracy and robustness of vehicle state characteristics, and make the subsequent prediction of temperature change trends and load level determination more reliable, thereby ensuring that the power limiting strategy can respond to thermal risks under complex operating conditions in a timely manner.

[0056] For example, the first prediction model can be specifically a bidirectional long short-term memory network (Bi-LSTM) that performs bidirectional temporal encoding on the first input parameter sequence to obtain the vehicle state features of the target vehicle. Alternatively, the first prediction model can be specifically a convolutional neural network that extracts features from the first input parameter sequence through three stacked convolutional layers to obtain the vehicle state features of the target vehicle, where the convolutional kernel size is dynamically adjusted according to the parameter type. Alternatively, the first prediction model can be specifically a fully connected network that extracts features from the first input parameter sequence through three fully connected layers to obtain the vehicle state features of the target vehicle.

[0057] The second part involves the vehicle controller using the first prediction model to regress the vehicle's state characteristics and obtain multiple predicted temperatures for the drive motor.

[0058] In one possible implementation, the vehicle controller performs multiple full connections on the vehicle state characteristics through the regression head of the first prediction model to obtain the multiple predicted temperatures.

[0059] The regression head refers to the output structure in a neural network specifically designed for temperature prediction tasks. It can be implemented by stacking multiple fully connected layers and using a nonlinear activation function. The purpose is to extract complex correlations between features layer by layer, thereby generating a continuous temperature change prediction sequence.

[0060] In the above implementation, the vehicle state features are first input into the regression head of the first prediction model. The features are then processed layer by layer through multiple fully connected operations to output the predicted temperatures at multiple time points, thus forming a temperature change trend.

[0061] For example, the first prediction model includes a regression head, which comprises two fully connected layers, each using the Sigmoid activation function. After the vehicle controller inputs the vehicle state features into the regression head, the regression head performs two fully connected layers on the vehicle state features, and after each fully connected layer, performs activation processing using the Sigmoid activation function to obtain the predicted temperature for multiple future time points. In this embodiment, "multiple future time points" refers to five consecutive future time points.

[0062] The third part involves the vehicle controller classifying the target vehicle based on its state characteristics using the first prediction model to obtain the load level of the target vehicle.

[0063] In one possible implementation, the vehicle controller performs a fully connected and normalized process on the vehicle state features using the classification head of the first prediction model to obtain a probability set. This probability set includes multiple probabilities, each probability corresponding to a candidate load level. The vehicle controller then determines the candidate load level corresponding to the highest probability in this probability set as the load level of the target vehicle.

[0064] The classification head refers to the output structure in the neural network specifically used for load level determination. It can be implemented using a fully connected layer followed by a normalization function, aiming to transform feature information into a normalized probability distribution. The probability set is the confidence vector of each candidate load level, which is normalized to make the probabilities comparable. The highest probability determination mechanism selects the category corresponding to the maximum probability based on the Argmax operation, providing a high-confidence load level determination result. In some embodiments, the multiple candidate load levels include four load levels: light load, medium load, heavy load, and extreme load. The determined load level will subsequently serve as an important auxiliary variable for the power adjustment factor.

[0065] In the above implementation, firstly, vehicle state features are input into the classification head, and after fully connected processing and normalization, probability values ​​for each candidate load level are generated. Then, the candidate load level with the highest probability in the probability set is selected as the final determination result. This allows the load level determination to rely on the peak characteristics of the probability distribution, reducing the risk of misjudgment and providing an accurate basis for subsequent power limiting.

[0066] For example, the first prediction model also includes a classification head, which consists of a fully connected layer and a normalization layer using the Softmax function to generate probability distributions for four load levels: light load, medium load, heavy load, and extreme load. When vehicle state features are input, the classification head outputs normalized probability values, and the vehicle controller can determine the load level based on these normalized probability values.

[0067] The technical solution described in the above three steps enables temperature change trend prediction to fully capture dynamic change trends, avoids the limitations of a single prediction point, and makes load level determination reliable, reducing the risk of misjudgment caused by feature ambiguity, thereby improving the accuracy of subsequent power limiting and solving the problems of response lag and over-limitation in thermal management.

[0068] The training method for the first prediction model will be explained below.

[0069] In one possible implementation, the electronic device collects sample data. All collected sample data is aligned according to a unified timestamp format, and a sliding window is used to construct a sample sequence. The window length is fixed at 30 seconds, and the step size is set to 1 second, ensuring that each sample data contains sufficient historical thermal accumulation characteristics and has timely response to changes. The sample data includes a set of vehicle state parameters, a set of motor operating parameters, and a set of environmental parameters. The raw temperature data in the sample data is filtered by median filtering and upper / lower limit clipping to eliminate extreme values ​​introduced by hardware errors and sensor faults. Electrical signal data such as speed, current, and voltage are smoothed by a first-order low-pass filter to smooth oscillations. Torque and load variables are normalized to a fixed range for consistency of multi-model inputs. The collected sample data is labeled according to the motor temperature rise rate. A temperature rise exceeding 3 degrees Celsius per minute is defined as a high-temperature sensitive area and labeled with power protection-related samples, serving as important supervisory input for subsequent multi-task models.

[0070] The output sample format consists of two parts: one is a time-series input vector, a multivariate matrix of length 30 seconds, which serves as the raw input for the subsequent first prediction model; the other part is supervision labels, including the predicted motor temperature for the next 5 seconds as the regression target, and the vehicle load level classification (light load, medium load, heavy load, extreme load) at the current moment as the classification target. The sampling frequency is uniformly set to 10Hz, ensuring the stability of the time-dimensional samples in the network structure and adapting to the input requirements of subsequent multi-task models. The sample data is finally packaged and stored as dedicated training data and inference input data for the model, forming a data closed loop.

[0071] The sampling mechanism and labeling system of the aforementioned sample data are compatible with multiple vehicle models and multi-platform electric drive architectures, possessing excellent horizontal scalability. It also provides a structured, controllable, and traceable input foundation for multi-task learning models and gating control models. The entire data acquisition and preprocessing system fully meets the modeling requirements of automotive-grade temperature control systems, demonstrates practical deployment feasibility, and can be cross-validated with thermal field simulation results in subsequent engineering verification.

[0072] During the training phase of the first prediction model, a joint loss function is employed. The temperature prediction branch uses mean absolute error as the loss function, while the load classification branch uses cross-entropy loss. To maintain training stability and task weight balance, a fixed proportional coefficient is introduced to weight the two sub-task losses. The weight values ​​are obtained through pre-experiment optimization and can be customized according to the weight configuration of the car company samples in actual engineering. The training data is fed into the first prediction model in a shuffled batch processing method. An adaptive learning rate algorithm is used for weight updates, such as the ADAM optimizer, to maximize training efficiency and convergence stability.

[0073] The first prediction model boasts engineering advantages such as unified structure, high computational efficiency, and simple deployment, enabling rapid response and high-accuracy recognition in a vehicle control environment. The shared encoder structure reduces model size and storage requirements while effectively improving the coupling modeling capability between temperature and load features, providing stable input for downstream power regulation strategies. With sufficient sample size, the first prediction model exhibits strong generalization ability, adapts to the data feature distribution of different vehicle platforms, and can be repeatedly deployed as an algorithm asset in multiple projects for automotive enterprise platforms.

[0074] 303. The vehicle controller determines the power limiting factor of the target vehicle based on the predicted temperature change trend, the load level, and the subset of motor state parameters in the set of motor operating parameters. The power limiting factor is used to limit the output power of the drive motor.

[0075] The power limiting factor is a coefficient used to adjust the upper limit of the drive motor's output power. The motor state parameter subset is a part of the motor operating parameter set, including parameters reflecting the operating state of the drive motor. In the embodiments of this application, the core of the temperature control protection strategy lies in intelligently limiting the peak power output of the drive motor based on the current thermal state (represented by the predicted temperature change trend) and load risk (represented by the load level), rather than using a fixed threshold for coarse-grained control.

[0076] In one possible implementation, the vehicle controller determines the thermal risk zone to which the drive motor belongs and its location within that zone based on the subset of motor state parameters. The thermal risk zone represents the thermal load state of the drive motor. The vehicle controller then concatenates the predicted temperature change trend, the load level, the thermal risk zone, and the subset of motor state parameters into a second input parameter sequence. This second input parameter sequence is then input into a second prediction model, which processes the sequence to obtain the power limiting factor for the target vehicle. This second prediction model is a model constructed based on a gating mechanism.

[0077] The thermal risk zone refers to the regional division used to characterize the thermal load state of the drive motor, reflecting the dynamic distribution characteristics of the motor's thermal state through hierarchical regional division. The second input parameter sequence is a multi-dimensional feature vector integrating predicted temperature change trends, load levels, thermal risk zones, locations within the thermal risk zones, and subsets of motor state parameters. It can be constructed using vector concatenation to provide comprehensive input features that simultaneously include temporal trends and categorical states. The second prediction model is a neural network model built based on a gating mechanism, which can be implemented using gated recurrent units or long short-term memory networks. It effectively filters key thermal state features and suppresses noise interference by dynamically adjusting information flow weights.

[0078] In the above implementation, the thermal risk zone and location of the drive motor are first determined based on a subset of motor state parameters, achieving a refined hierarchical characterization of the thermal load state. Then, the predicted temperature change trend, load level, thermal risk zone, location within the thermal risk zone, and the subset of motor state parameters are concatenated into a second input parameter sequence, thereby integrating temperature evolution trends, operating load information, and thermal risk distribution characteristics. Finally, a second prediction model based on a gating mechanism is used to perform time-series modeling of the sequence. The gating mechanism adaptively highlights the sensitive characteristics during thermal risk abrupt changes, outputting a power limiting factor that matches the current thermal state.

[0079] Through the above implementation methods, it is possible to achieve a refined hierarchical characterization of the thermal risk state of the drive motor, output the corresponding power limiting factor under the condition of rapid fluctuation of thermal load, effectively prevent thermal damage while avoiding excessive limitation of power output under low-risk conditions, thereby improving the driving experience and system reliability.

[0080] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0081] Part 1: The vehicle controller determines the thermal risk zone to which the drive motor belongs and its location within the thermal risk zone based on the subset of motor state parameters.

[0082] In one possible implementation, the vehicle controller performs random nearest neighbor embedding based on a preset distribution on the subset of motor state parameters to obtain motor state features of the subset of motor state parameters. The vehicle controller determines the degree of matching between the motor state features and multiple candidate thermal risk areas. Based on the degree of matching between the motor state features and the multiple candidate thermal risk areas, the vehicle controller determines the thermal risk area to which the drive motor belongs from the multiple candidate thermal risk areas, and determines the position of the drive motor in the thermal risk area based on the degree of matching between the motor state features and the thermal risk area to which it belongs. This thermal risk area is the candidate thermal risk area with the highest degree of matching among the multiple candidate thermal risk areas.

[0083] Among them, random nearest neighbor embedding based on a preset distribution refers to a nonlinear dimensionality reduction technique, which can be implemented using manifold learning algorithms such as T-SNE (t-Distributed Stochastic Neighbor Embedding), UMAP (Uniform Manifold Approximation and Projection), or Isomap (Isometric Feature Mapping). The aim is to map high-dimensional motor state parameters to a low-dimensional feature space to preserve the local similarity structure between parameters and avoid feature distortion caused by the curse of dimensionality. The matching degree is determined by calculating the cosine similarity or Mahalanobis distance between feature vectors to quantify the correlation strength between state features and candidate thermal risk zones. This transforms abstract thermal states into quantifiable evaluation indicators, overcoming the limitation of static rules in adapting to dynamic parameter changes. Thermal risk zones are thermal load state regions formed by clustering historical data, providing a multi-dimensional thermal risk classification benchmark to ensure that risk identification is based on current state characteristics rather than preset fixed boundaries. In some embodiments, multiple candidate thermal risk zones include normal operation zones, preheating overload zones, rapid heating zones, and high-temperature runaway critical zones.

[0084] In the above implementation, a subset of motor state parameters is first randomly embedded to convert high-dimensional parameters into low-dimensional motor state features. This process preserves the nonlinear correlation between multi-parameter combinations and thermal states. Then, the matching degree between these motor state features and each candidate thermal risk zone is determined. The thermal risk zone to which the drive motor belongs is determined by the highest matching degree, allowing thermal risk classification to dynamically adapt to the current operating conditions. Finally, the center, middle, or edge positions are divided based on the matching degree threshold to achieve continuous risk level assessment. For example, strict power limiting is triggered when the matching degree is high to address high thermal risk, while restrictions are relaxed when the matching degree is low to avoid excessive intervention, thus forming a complete dynamic thermal risk identification chain. Furthermore, when the matching degree is greater than or equal to the first matching degree threshold, the vehicle controller identifies it as the center position. This accurately identifies high-risk states and avoids excessive power limiting or protection lag due to ambiguous location. When the matching degree is between the second and first matching degree thresholds, the vehicle controller identifies it as the middle position, reflecting that the motor is in a transitional zone of the risk zone. This helps to implement appropriate power limiting and prevents loss of power output due to conservative strategies under fluctuating operating conditions. When the matching degree is less than the second matching degree threshold, it is identified as an edge position, indicating that the motor is approaching the boundary of the risk zone. This allows for timely release of power limitation margin, maintaining the driving experience while avoiding unnecessary performance sacrifices. This discretization mechanism transforms continuous risk assessment into discrete decision points, significantly improving the reliability and adaptability of thermal risk location determination.

[0085] Through the above implementation methods, the thermal risk status of the motor can be dynamically identified, avoiding the risk identification delay or misjudgment caused by static rules. Thus, while ensuring the safe operation of the drive motor, the accuracy and adaptability of the power limiting strategy can be optimized, and the continuity of power output and driving experience of the vehicle under complex working conditions can be improved.

[0086] For example, the vehicle controller uses the T-SNE algorithm to embed the subset of motor state parameters to obtain the motor state features of the subset. The vehicle controller determines the feature similarity between the motor state features and the cluster centers of the corresponding clusters of each candidate thermal risk area. The vehicle controller determines this feature similarity as the degree of matching between the motor state features and multiple candidate thermal risk areas. The vehicle controller determines the candidate thermal risk area with the highest degree of matching among the multiple candidate thermal risk areas as the thermal risk area to which the drive motor belongs. The vehicle controller performs interval judgment on the calculated matching degree according to a preset first matching degree threshold and a second matching degree threshold. When the matching degree is higher than the higher first matching degree threshold, the position is marked as the center. When it is between the first and second matching degree thresholds, the position is marked as the middle. When it is lower than the second matching degree threshold, the position is marked as the edge.

[0087] T-SNE is a nonlinear dimensionality reduction technique based on probability distribution, used to preserve the local structural relationships between original data points and avoid feature distortion caused by general embedding methods. In other words, as a nonlinear dimensionality reduction method, T-SNE projects the global structure to two- or three-dimensional space while maintaining the point-to-point local distance relationships in high-dimensional space, clearly delineating different feature regions in complex nested distributions. Feature similarity refers to an index that quantifies the correlation strength between feature vectors, which can be achieved using cosine similarity, Euclidean distance, or Mahalanobis distance, aiming to objectively assess the closeness between motor state characteristics and the core characteristics of thermal risk zones. Matching degree refers to a parameter characterizing the strength of thermal risk zone attribution, which can be understood as a direct numerical mapping of feature similarity, used to establish repeatable quantitative judgment criteria. The first matching degree threshold and the second matching degree threshold refer to preset threshold boundaries used to discretize the matching degree, which can be determined based on statistical analysis of historical thermal fault data or optimized through machine learning models, aiming to divide continuous matching degrees into clear risk level intervals. In practical applications, center position, middle position, and edge position refer to the relative position categories of the drive motor within its respective thermal risk zone. They can be understood as the discretized division of regions with different risk intensities within the thermal risk zone, with the aim of providing accurate decision-making basis for power limiting strategies.

[0088] In the example above, the T-SNE algorithm maps a subset of motor state parameters to a low-dimensional feature space, preserving the local topological characteristics of the thermal state data. This allows the dimensionality-reduced motor state features to truly reflect the intrinsic correlations of the motor's thermal behavior. Based on this, feature similarity is calculated by focusing on the cluster centers of candidate thermal risk zones rather than the original data points. This simplifies the matching calculation process and enhances noise resistance through the core regional characteristics of the cluster centers. Using feature similarity directly as the matching mechanism ensures that the thermal risk zone determination process follows objective quantitative standards, thus providing a reliable basis for the dynamic optimization of the power limiting factor. Furthermore, distinguishing the different location categories of the drive motor within the thermal risk zone avoids overly conservative or insufficiently responsive power limiting strategies in thermal risk assessment, thereby improving the thermal management efficiency and driving performance stability of the electric drive system under complex operating conditions.

[0089] The technical solutions described above improve the accuracy of motor state feature extraction and the precision of thermal risk state differentiation, making the thermal risk zone determination process repeatable and engineering applicable. This makes the generation of power limiting factors more reliable and avoids the risk of excessive power limiting or thermal damage caused by thermal risk identification errors.

[0090] The following describes another method for determining thermal risk zones and their locations within those zones.

[0091] In one possible implementation, the vehicle controller uses the subset of motor state parameters to query the target relation table to obtain the thermal risk zone to which the drive motor belongs and its location within the thermal risk zone.

[0092] The target relation table refers to a pre-built mapping data structure, which can be implemented using a lookup table, hash table, or database table. It stores the correspondence between a subset of motor state parameters and thermal risk zones and their locations, allowing for predefined mappings to replace real-time calculations and avoid the computational overhead of high-dimensional data processing. The query operation can be understood as the process of retrieving information from the target relation table based on input parameters, used to quickly obtain predefined thermal risk zone information.

[0093] In the above implementation, by using a subset of motor state parameters as the query key for direct retrieval in a pre-built target relation table, high-dimensional embedding and clustering operations in real-time computation are avoided. Since the target relation table pre-stores the mapping relationships between all possible subsets of motor state parameters and thermal risk zones, the query process requires only constant time complexity, significantly reducing the delay in determining thermal risk zones. This mechanism is closely integrated with the power limiting factor generation process, enabling the thermal state assessment results to be input into the second prediction model in real time, ensuring the timeliness of power limiting decisions and effectively addressing dynamic changes in the motor's thermal state.

[0094] Through the above implementation methods, the computational complexity of the thermal risk zone determination process is greatly reduced, and the thermal state assessment delay is significantly reduced, enabling power limiting decisions to respond promptly to rapid changes in the motor's thermal state and avoiding the problem of thermal protection lag caused by assessment delay.

[0095] For example, the target relationship table can be implemented as a hash table stored in the vehicle's electronic control unit, where the key value is a combination of motor speed and motor current. When a subset of the current motor state parameters is obtained, the integrated controller quantizes the motor speed and motor current into discrete interval indices, calculates the storage address using a hash function, and directly reads the corresponding thermal risk zone identifier and location information.

[0096] To provide a clearer explanation of the above implementation methods, the method for determining the multiple candidate thermal risk zones will be described below.

[0097] In one possible implementation, the electronic device acquires multiple subsets of historical motor state parameters of the drive motor. The electronic device performs random nearest neighbor embedding based on a preset distribution on each subset of historical motor state parameters to obtain historical motor state features for each subset. The electronic device then clusters these historical motor state features from the multiple subsets of historical motor state parameters to obtain a preset number of clusters, with each cluster corresponding to a candidate thermal risk region.

[0098] The historical motor state parameter subset refers to a subset of state parameters recorded during the historical operation of the drive motor. This can be achieved by periodically collecting motor operation data from an onboard data recording system. The purpose is to provide samples of thermal behavior under real-world operating conditions, avoiding deviations caused by human pre-setting. Clustering specifically refers to the process of grouping similar data points. Algorithms such as K-means clustering, hierarchical clustering, or density clustering can be used to automatically identify groups based on thermal state similarity, ensuring that each cluster objectively represents a specific thermal load state. The preset number is the same as the number of candidate thermal risk areas; if there are four candidate thermal risk areas, the preset number is four.

[0099] In the above implementation, historical operating data of the drive motor is used as a basis, making the determination of candidate thermal risk zones dependent on real-world scenarios. Subsequently, random nearest neighbor embedding based on a preset distribution is used to extract features from the historical data, preserving the local similarity of thermally related features during dimensionality reduction and highlighting key dimensions of temperature changes. Finally, cluster analysis is performed on the extracted features to automatically form a preset number of clusters, each corresponding to a candidate thermal risk region. This data-driven approach enables the division of thermal risk regions to objectively reflect the thermal behavior patterns of the motor under different operating conditions, providing a reliable benchmark for subsequent thermal risk assessment.

[0100] Through the above implementation methods, the process of determining candidate thermal risk zones is data-driven and objective, solving the problem of a lack of scientific basis for thermal risk zone division. Specifically, cluster analysis based on historical data enables each candidate thermal risk zone to accurately reflect the thermal behavior characteristics of the motor under real operating conditions, thereby improving the accuracy of thermal risk assessment. This ensures that the calculation of the power limiting factor is more in line with actual thermal management needs, avoiding excessive power limitation or insufficient protection caused by inaccurate thermal risk assessment, and improving the adaptability of the electric drive system under complex operating conditions.

[0101] For example, electronic devices acquire a subset of historical motor state parameters. This subset is processed using the T-SNE algorithm, mapping high-dimensional parameters to a two-dimensional feature space while preserving the local structure of thermally relevant features. Subsequently, the embedded features are grouped using the K-means clustering algorithm, with a cluster size of 4, resulting in 4 clusters, each representing a candidate thermal risk state. This implementation ensures that the division of thermal risk areas is entirely based on historical operating data, avoiding biases caused by subjective settings.

[0102] The T-SNE algorithm can be encapsulated within the T-SNE module. In the dimensionality reduction results obtained using the T-SNE algorithm, the boundaries of thermal risk regions are clearly depicted, including the normal operation zone, preheating overload zone, rapid temperature rise zone, and high-temperature runaway critical zone. Using the T-SNE algorithm, the thermal state of the drive motor is mapped and judged in real time. Combined with the current state's distribution position in the mapping map, this further assists the second prediction model in fine-tuning the adjustment degree of the power limitation factor. Especially in regions near the boundaries, the adjustment step size and power limitation amplitude should be more conservative to avoid frequent switching of the temperature control strategy due to state jitter. After completing the state mapping during the training phase, the T-SNE module can store various thermal risk regions in an indexed manner. During the inference phase, the K-nearest neighbor algorithm is used to quickly match the position of the current state, achieving rapid mapping of the regions.

[0103] The T-SNE module is also used to perform cluster evaluation analysis on historical operating conditions and temperature control strategy results. For example, it can statistically analyze which state regions are more likely to trigger high-temperature overload events and which regions correspond to power losses caused by over-protection, thus providing an intuitive basis for model tuning and strategy optimization. In addition, this method can also be used to assist in verifying the response accuracy and rationality of multi-task models and gating networks in high-temperature boundary regions, serving as a model validity verification tool.

[0104] In some embodiments, the T-SNE module serves as a tool-based modeling component and does not participate in the real-time inference process. It plays a crucial role only in strategy formulation, strategy classification and calibration, offline boundary modeling, and model parameter tuning, exhibiting high interpretability and portability. The T-SNE module has an independent functional hierarchy and can be deployed separately as a visualization-assisted strategy system, or integrated into a host computer system to provide strategy analysis services, providing technical support for the accurate boundary identification of the vehicle power regulation system.

[0105] The third part involves the vehicle controller processing the second input parameter sequence using the second prediction model to obtain the power limiting factor for the target vehicle.

[0106] In one possible implementation, the vehicle controller performs a fully connected and activated operation on the second input parameter sequence using the second prediction model to obtain a power limiting feature. The vehicle controller then normalizes this power limiting feature using the second prediction model to obtain the power limiting factor.

[0107] In this context, "fully connected" and "activation" refer to the linear combination and nonlinear transformation of the input through weight matrices in a neural network. Activation can be achieved using activation functions such as ReLU, Sigmoid, or Tanh, aiming to enhance the nonlinear representation capability of the second prediction model for multi-dimensional input information, thereby capturing the complex correlation between thermal risk and dynamic load changes. Power limitation features can be understood as intermediate feature representations integrating multi-source information such as predicted temperature change trends, load levels, thermal risk zones, and subsets of motor state parameters, used as a comprehensive evaluation index reflecting the true thermal load state of the drive motor. In practical applications, normalization can specifically be a mathematical transformation operation that maps feature values ​​to a standardized interval. For example, the Softmax function or Min-Max normalization method can be used to constrain the range of power limitation factors, avoid extreme value interference, and ensure output stability.

[0108] In the above implementation, firstly, a fully connected layer is applied to the second input parameter sequence using a second prediction model. This linearly combines multi-dimensional input information such as predicted temperature change trends, load levels, thermal risk zones, and subsets of motor state parameters. An activation function is then used to introduce nonlinear transformation capabilities to fully extract the complex correlations between features, thereby generating highly representative power limitation features. Based on this, normalization maps the power limitation features to a preset standardized range, limiting the dynamic range of the power limitation factor. This ensures a smooth transition under continuous operating conditions, avoiding output range instability caused by feature fluctuations, and thus providing a reliable basis for power limitation.

[0109] The above implementation methods can suppress the drastic fluctuations of the power limiting factor under continuous operating conditions, avoid sudden changes in the output power of the drive motor, and thus improve the smoothness of the vehicle's power response and the reliability of the thermal management system.

[0110] For example, the second prediction model can employ a structure containing multiple fully connected networks. This second prediction model uses fully connected networks and ReLU activation functions to process the second input parameter sequence to generate power-limited features. Subsequently, a Softmax normalization layer transforms the features into a power-limiting factor in the form of a probability distribution. This factor is constrained within a standardized range of 0 to 1, ensuring a stable output value range that meets the subsequent power-limiting logic requirements.

[0111] The following specific example illustrates step 303 above.

[0112] The second prediction model is a lightweight feedforward network. Its core objective is to learn the nonlinear mapping relationship between the input state and the output power constraint factor. Its internal structure consists of a three-layer neural network. The input layer dimension is directly aligned with the multi-task output (the output of the first prediction model), containing a state vector composed of five future temperature predictions, load level probability distributions, current voltage, current, and motor speed. The hidden layers employ fully connected matrices and the ReLU activation function to enhance the second prediction model's ability to fit high-dimensional feature combinations. The output layer uses a Sigmoid function to compress the power constraint factor into a continuous range between 0 and 1, which is then used to multiply by the theoretical maximum power to achieve dynamic proportional adjustment. This power constraint factor does not directly replace the power setpoint but serves as a modulation reference on the control side, injected into the power loop of the vehicle's VCU or motor controller to achieve cycle-by-cycle dynamic adjustment.

[0113] The training phase of the second prediction model uses real-world operating condition data samples. A supervisory labeling system is established for power output and temperature changes before and after the high-temperature trigger point. Based on whether the motor overheats, whether the protection mechanism is triggered, and the corresponding temperature curve shape after adjustment, a standard power limiting factor label is constructed as the supervisory target for the second prediction model. Mean squared error is used as the loss function during training, and a penalty term is introduced in conjunction with the temperature rise control target to suppress the risk of over-adjustment or under-adjustment, ensuring that the second prediction model retains the maximum usable driving force while guaranteeing thermal safety. Mini-batch optimization is employed during training, with random sampling of batch samples to avoid overfitting.

[0114] During the inference phase, the second prediction model runs once per second, receiving the latest temperature predictions, load classification outputs, and the current actual operating status. It then generates a new power limiting factor and adjusts the target power limit of the motor in real time. This mechanism is compatible with various vehicle architectures and can be integrated into the communication link between the VCU and MCU, or embedded in the motor controller firmware as a self-protection module.

[0115] This second prediction model fully leverages the high-dimensional thermal load characterization capabilities provided by the upstream first prediction model. Through a simple neural structure, it achieves rapid and continuous power regulation behavior, greatly improving the responsiveness and flexibility of traditional fixed-power protection logic. Compared to traditional temperature-threshold-based on-off protection methods, this gating mechanism supports continuous value adjustment, enabling early control of temperature rise trends, avoiding system shocks caused by critical point fluctuations, and improving the smoothness of vehicle operation and the robustness of thermal management.

[0116] 304. The vehicle controller limits the output power of the drive motor based on the power limiting factor and the initial power limit of the drive motor.

[0117] The initial power limit refers to the theoretical maximum allowable power value dynamically calculated by the electric drive system based on the current operating conditions.

[0118] In one possible implementation, the vehicle controller uses the power limiting factor to limit the initial power limit, thus obtaining a target power limit for the drive motor. The vehicle controller then limits the power of the drive motor based on this target power limit.

[0119] Among them, limiting the power of the drive motor with the target power limit means inputting the calculated target power limit as a control command to the electric drive system. In other words, the output of the drive motor is constrained by updating the power limit in real time to prevent thermal failure of the drive motor.

[0120] In the above implementation, a dynamically adjusted target power limit is generated by calculating the power limiting factor with the initial power limit. This target power limit directly affects the drive motor control. The power limiting factor, as a dynamic parameter integrating predicted temperature change trends, load levels, and motor state parameters, is combined with the initial power limit to output a continuously changing target power limit, which is used to limit the drive motor's output power in real time. This mechanism ensures that the power limiting process adaptively adjusts based on real-time thermal risk conditions, avoiding sudden power fluctuations caused by thermal load variations, thereby maintaining the smoothness of vehicle power output.

[0121] The above implementation method avoids sudden changes or discontinuities in the target power limit, ensures a smooth transition of the drive motor output power, and improves the smoothness of vehicle power output and driving comfort.

[0122] To provide a clearer explanation of the above implementation methods, the method for determining the target power upper limit in the above implementation methods will be described below.

[0123] In one possible implementation, the vehicle controller multiplies the power limiting factor by the initial power limit to obtain the target power limit of the drive motor.

[0124] The following explains another method for determining the target power limit.

[0125] In one possible implementation, the vehicle controller performs time filtering on the power limiting factor to obtain a filtered power limiting factor. The time filtering employs either moving average filtering or exponential decay filtering to improve the smoothness of the power limiting factor over consecutive periods. The vehicle controller then compares the filtered power limiting factor with the initial power... Time filtering refers to techniques for smoothing the power limiting factor over time. Specifically, methods such as moving average filtering or exponential decay filtering can be used to suppress random fluctuations caused by measurement noise or transient changes in operating conditions. Moving average filtering involves arithmetically averaging the power limiting factor over a fixed time window, typically based on the average of multiple historical periods, aiming to eliminate random noise. Exponential decay filtering involves weighted averaging of the power limiting factor, assigning higher weight to recent data. This can be achieved using a smoothing algorithm with an exponential decay coefficient, aiming to better track trends in thermal risk.

[0126] In the above implementation, by providing a dual mechanism of direct application of the power limiting factor and smoothing processing, when the power limiting factor fluctuates greatly, the time filtering mechanism smooths the factor value in continuous periods, making the change of the target power limit more gradual and avoiding sudden changes in output power.

[0127] As a specific implementation method, when fluctuations in the power limiting factor are detected due to instantaneous changes in operating conditions, the vehicle controller uses exponential decay filtering to process the factor, assigns higher weight to recent data, obtains a smoothed power limiting factor, and then multiplies it with the initial power limit to achieve a gradual adjustment of the target power limit.

[0128] The above technical solution effectively suppresses abrupt changes in the power limiting factor, making the target power limit change smoothly in a continuous cycle, avoiding instability in the output power of the drive motor, significantly alleviating the jerking sensation of vehicle power output, and improving driving smoothness and user experience.

[0129] The technical solutions provided in the embodiments of this application are described below in conjunction with the above-mentioned technical features.

[0130] The ultimate goal of the second predictive model is to dynamically limit and adjust the output power of the drive motor based on its current thermal state, predicted temperature change trends, and load level. A complete power control logic chain is constructed around the output value of the gating network (second predictive model), ensuring the control signals have interpretability, consistent response, and deployment adaptability throughout the vehicle system. The control logic uses the power limiting factor output by the second predictive model as the core control variable, combined with the existing power adjustment channel of the motor controller, to achieve periodic updates and flexible adjustments in the control command issuance chain.

[0131] The core control variable of the power control logic is the power limiting factor output by the gating network. This power limiting factor is a continuous value between 0 and 1, representing the proportion of currently available power to the maximum theoretical power. This value is not statically set but calculated and generated in real time, reflecting the dynamic assessment capability of the drive motor's future temperature rise risk. The vehicle controller integrates this limiting factor with the initial power limit of the current drive motor, generates a limited target power limit through a multiplication operation, and inputs the result into the power control loop. The limited target power limit serves as a reference value for the controller and participates in the internal adjustment of the PI or model predictive control module, achieving a soft-limit response to speed and torque.

[0132] To ensure the smoothness and anti-fluctuation capability of the control signal, a time filtering strategy is introduced after the power limiting factor is output. The power limiting factor output in continuous cycles is processed by using the moving average or exponential decay weight method to avoid power command jumps caused by model output fluctuations and improve the driving smoothness of the whole vehicle.

[0133] The control logic also includes a strategy state judgment module, which determines the adjustment mode of the control logic based on the T-SNE state mapping region label of the current vehicle. For example, when the state is in the stable region, the constraint factor changes slowly, and the control logic prioritizes performance; when the state is in the boundary sensitive region, the constraint factor updates more frequently, the adjustment amplitude increases, and the control strategy prioritizes thermal safety. This region label is embedded in the control logic as a judgment parameter, improving the granularity of the control strategy.

[0134] The entire power control logic interfaces with the vehicle's electronic control system via a standard interface, and is compatible with mainstream VCU communication protocols. This includes CAN frame data encapsulation, signal refresh frequency configuration, and fault diagnosis compatibility, all of which meet mass-production vehicle requirements. Parameters such as power limit, model output status, T-SNE area number, and thermal risk level can be transmitted back to the vehicle diagnostic system as runtime signals for post-processing analysis and strategy tracing, achieving transparent control management.

[0135] To ensure the fusion model has stable operation and resource adaptability in a mass-produced vehicle environment, the model structure needs to be deployed and optimized, and a complete closed-loop verification needs to be completed in the vehicle control chain. The deployment phase focuses on model structure compression, computing resource adaptation, inference response time control, and communication interface integration, while the verification phase uses a test platform to evaluate the functionality, stability, and robustness of the temperature control strategy.

[0136] The model deployment process begins with pruning and quantization optimization of the multi-task learning model (first prediction model) and the gating network (second prediction model). The Bi-LSTM encoder of the first prediction model undergoes pruning to reduce the number of neurons, retaining only the input channels sensitive to temperature rise trends, thus reducing computational load. The output dimension of the shared encoder in the first prediction model is reduced to decrease the number of parameters for subsequent classification and regression tasks. Without affecting prediction accuracy, some activation functions are replaced with less computationally complex linear or approximate functions to improve inference speed. The gating network, due to its lightweight structure, only undergoes weight quantization and matrix sparsity processing to improve the model's efficiency on low-frequency embedded controllers. The T-SNE module, used only for offline mapping and policy-aided generation, does not enter the real-time control path during deployment; it only retains the mapping table for power policy logic lookup and requires no computational resources.

[0137] In this embodiment, vehicle state features (such as fused features of vehicle state parameters, motor operating parameters, and environmental parameters) extracted by the first prediction model of multi-task learning serve as the input basis for T-SNE. Multi-task learning enables these features to simultaneously incorporate information from temperature prediction and load classification, thus making the features more discriminative and robust. T-SNE utilizes these rich features for dimensionality reduction and clustering, enabling more accurate definition of thermal risk zones. For example, multi-task learning may capture the correlation between load and temperature, making T-SNE clustering more capable of distinguishing different risk levels. This improves the accuracy of thermal risk zone assessment, thereby providing a more reliable risk signal for power limiting.

[0138] The combination of multi-task learning and T-SNE makes the overall state assessment more comprehensive. Multi-task learning provides multi-perspective features, while T-SNE transforms these features into actionable risk area information, reducing the possibility of misjudgment.

[0139] Furthermore, in the combination of T-SNE and a gating network, the thermal risk zone and its location information (such as center, middle, or edge position) output by T-SNE serve as one of the inputs to the gating network. The gating network can dynamically adjust the weights of different input parameters based on this risk information. For example, if T-SNE indicates that the motor is in a high thermal risk zone, the gating network may focus more on the temperature prediction trend, thereby strengthening power limiting. The gating mechanism allows the model to adaptively fuse multi-source information, including the risk assessment of T-SNE, the prediction output of multi-task learning, and real-time motor parameters. This fusion makes the calculation of the power limiting factor more closely resemble actual operating conditions, avoiding overly conservative or risky power control.

[0140] The process of combining T-SNE with a gating network makes power control more refined and adaptive. The gating network flexibly adjusts its strategy according to the risk level of the T-SNE, improving the system's response speed and security.

[0141] Furthermore, regarding the combination of multi-task learning and gating networks, the predicted temperature change trends and load levels generated by multi-task learning are directly input into the gating network, avoiding the complexity of training multiple models separately. The gating network, as a secondary prediction model, can efficiently integrate these multi-task outputs and utilize the gating mechanism to select the most relevant features. This combination forms an end-to-end learning framework: multi-task learning handles the front-end multi-objective prediction, and the gating network handles the back-end decision-making. This reduces error accumulation and improves the overall efficiency of the system.

[0142] The combination of multi-task learning and gating networks makes the training and inference processes more efficient and the model more compact; the calculation of the power limiting factor is based on shared features and dynamic gating, thereby maintaining stable performance under complex driving conditions.

[0143] The model inference is encapsulated as an independent control logic module, deployed within the VCU or MCU in C language or ONNX format. The input port receives state vector data from the underlying acquisition module, including temperature, voltage, current, vehicle speed, load, and ambient temperature. The output port sends the adjusted power limiting factor to the motor controller's power loop input interface. The operating frequency is set to 1Hz or 2Hz to meet real-time control requirements. During operation, it consumes less than 20% of the vehicle's control processing capacity, meeting the computational requirements of traditional dual-core MCUs under time-sharing task scheduling.

[0144] Vehicle-level verification was conducted through a combination of test bench testing and real-vehicle road testing. A test matrix was built to simulate different thermal environments and driving conditions, including various typical thermal risk scenarios such as high temperature and high load, frequent start-stop, continuous uphill climbing, and emergency acceleration. The power limiting factor response output by the model and the actual motor temperature rise trajectory were monitored in real time. The key verification focused on whether the power limiting strategy could effectively reduce the power output before the temperature rises abnormally, whether it could avoid false triggering of high temperature protection, and whether the vehicle's driving performance was stable and the overall vehicle response was continuous and without abrupt changes during the adjustment process.

[0145] During the verification process, the strategy system recorded all adjustment behaviors and model output states. Post-processing analysis was performed using the T-SNE state mapping results to statistically analyze the model's response accuracy and protection success rate within each thermal state range, confirming its ability to judge critical temperature rise conditions and control the adjustment range. Comparative tests were conducted with existing vehicle high-temperature protection strategies to verify the system's effectiveness in improving the vehicle's usable driving force while ensuring safety.

[0146] The final deployment plan and verification data are solidified into a strategy configuration package, which can be directly loaded and used in mass production projects, or locally adjusted according to vehicle platform. The entire deployment optimization and verification system is a complete closed loop, ensuring the efficient operation of the model in resource-constrained environments, and forming a targeted, adjustable, and traceable power output control mechanism under high-temperature conditions, supporting automakers' continuous upgrades and differentiated competitive capabilities in electric drive thermal protection technology.

[0147] 305. When the target vehicle is under preset operating conditions, the vehicle controller stops using the power limiting factor to limit the power of the drive motor. The preset operating conditions include emergency braking and overtaking conditions.

[0148] In the preset operating conditions, the power limitation on the drive motor is stopped to ensure that the drive motor can perform at its maximum capacity and improve safety under the preset operating conditions.

[0149] The technical solution provided in this application achieves proactive thermal management by predicting temperature change trends in advance, enabling dynamic adjustment of power limits before a sudden temperature rise occurs. Simultaneously, it avoids excessive power output limitation due to conservative protection margins, reducing unnecessary power limitations under low-load conditions, thereby improving the adaptability of the electric drive system under complex operating conditions and achieving a balance between thermal safety and power performance.

[0150] Figure 4 This is a schematic diagram of the structure of an output power control device provided in an embodiment of this application. See also... Figure 4 The device includes: a trend level determination module 401, a factor determination module 402, and a power limiting module 403.

[0151] The trend level determination module 401 is used to determine the predicted temperature change trend of the drive motor and the load level of the target vehicle when the target vehicle is in motion, based on the vehicle state parameter set of the target vehicle, the motor operation parameter set of the drive motor, and the environmental parameter set of the environment.

[0152] The factor determination module 402 is used to determine the power limiting factor of the target vehicle based on the predicted temperature change trend, the load level, and a subset of motor state parameters in the set of motor operating parameters. The power limiting factor is used to limit the output power of the drive motor.

[0153] The power limiting module 403 is used to limit the output power of the drive motor based on the power limiting factor and the initial power limit of the drive motor.

[0154] In one possible implementation, the trend level determination module 401 is used to extract features from the vehicle state parameter set, the motor operating parameter set, and the environmental parameter set using a first prediction model to obtain the vehicle state features of the target vehicle. The first prediction model is a model obtained through multi-task joint training. The vehicle state features are then regressed using the first prediction model to obtain multiple predicted temperatures of the drive motor, which represent the predicted temperature change trend of the drive motor. Based on the vehicle state features, the load level of the target vehicle is determined using the first prediction model.

[0155] In one possible implementation, the trend level determination module 401 is used to perform time alignment and preprocessing on the parameters in the vehicle state parameter set, the parameters in the motor operation parameter set, and the parameters in the environmental parameter set to obtain a first input parameter sequence. This first input parameter sequence is then input into the first prediction model, which performs bidirectional temporal encoding on the first input parameter sequence to obtain the vehicle state features of the target vehicle. Alternatively, the first prediction model can perform multiple convolutions or multiple fully connected layers on the first input parameter sequence to obtain the vehicle state features of the target vehicle.

[0156] In one possible implementation, the trend level determination module 401 is used to perform multiple fully connected operations on the vehicle state features through the regression head of the first prediction model to obtain multiple predicted temperatures. Then, the vehicle state features are fully connected and normalized through the classification head of the first prediction model to obtain a probability set, which includes multiple probabilities, each probability corresponding to a candidate load level. The candidate load level corresponding to the highest probability in the probability set is determined as the load level of the target vehicle.

[0157] In one possible implementation, the factor determination module 402 is used to determine the thermal risk zone to which the drive motor belongs and its location within the thermal risk zone based on the subset of motor state parameters. The thermal risk zone represents the thermal load state of the drive motor. The predicted temperature change trend, the load level, the thermal risk zone, the location, and the subset of motor state parameters are concatenated into a second input parameter sequence. This second input parameter sequence is input into a second prediction model, which processes the sequence to obtain the power limiting factor for the target vehicle. This second prediction model is a model constructed based on a gating mechanism.

[0158] In one possible implementation, the factor determination module 402 is used to perform random nearest neighbor embedding based on a preset distribution on the subset of motor state parameters to obtain the motor state features of the subset of motor state parameters. The degree of matching between the motor state features and multiple candidate thermal risk areas is determined. Based on the degree of matching between the motor state features and the multiple candidate thermal risk areas, the thermal risk area to which the drive motor belongs is determined from the multiple candidate thermal risk areas, and the position of the drive motor in the thermal risk area is determined based on the degree of matching between the motor state features and the thermal risk area to which it belongs. This thermal risk area is the candidate thermal risk area with the highest degree of matching among the multiple candidate thermal risk areas.

[0159] In one possible implementation, the random nearest neighbor embedding of the preset distribution is T-SNE, and the factor determination module 402 is used to determine the feature similarity between the motor state feature and the cluster centers of the corresponding clusters of each candidate thermal risk area. This feature similarity is determined as the degree of matching between the motor state feature and multiple candidate thermal risk areas.

[0160] In one possible implementation, the factor determination module 402 is configured to determine the location of the thermal risk area to which the drive motor belongs as the center position when the matching degree is greater than or equal to a first matching degree threshold; determine the location of the thermal risk area to which the drive motor belongs as the middle position when the matching degree is greater than or equal to a second matching degree threshold but less than the first matching degree threshold; and determine the location of the thermal risk area to which the drive motor belongs as the edge position when the matching degree is less than the second matching degree threshold.

[0161] In one possible implementation, the device further includes a thermal risk region determination module, used to acquire multiple subsets of historical motor state parameters of the drive motor. Each subset of historical motor state parameters is then subjected to random nearest neighbor embedding based on a preset distribution to obtain historical motor state features for each subset. These historical motor state features from the multiple subsets of historical motor state parameters are then clustered to obtain a preset number of clusters, with each cluster corresponding to a candidate thermal risk region.

[0162] In one possible implementation, the factor determination module 402 is used to perform a fully connected and activated operation on the second input parameter sequence using the second prediction model to obtain power limiting features. The power limiting features are then normalized using the second prediction model to obtain the power limiting factor.

[0163] In one possible implementation, the power limiting module 403 is used to limit the initial power limit using the power limiting factor to obtain a target power limit for the drive motor. The drive motor is then power-limited based on this target power limit.

[0164] In one possible implementation, the power limiting module 403 is used to multiply the power limiting factor by the initial power upper limit to obtain the target power upper limit of the drive motor. Alternatively, the power limiting factor is subjected to time filtering to obtain a filtered power limiting factor. The time filtering uses moving average filtering or exponential decay filtering to improve the smoothness of the power limiting factor over continuous periods. The filtered power limiting factor is then multiplied by the initial power upper limit to obtain the target power upper limit of the drive motor.

[0165] In one possible implementation, the power limiting module 403 is further configured to stop using the power limiting factor to limit the power of the drive motor when the target vehicle is in a preset operating condition, the preset operating condition including emergency braking condition and overtaking condition.

[0166] It should be noted that the output power control device provided in the above embodiments is only illustrated by the division of the above functional modules when controlling the output power of the drive motor. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the output power control device and the output power control method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0167] This application also provides a vehicle. Figure 5 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application.

[0168] Typically, vehicle 500 includes one or more processors 501 and one or more memories 502.

[0169] Processor 501 may include one or more processing cores, such as a quad-core processor, a penta-core processor, etc. Processor 501 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 501 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 501 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 501 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0170] The memory 502 may include one or more computer-readable storage media, which may be non-transitory. The memory 502 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 502 are used to store at least one computer program, which is executed by the processor 501 to implement the output power control method provided in the method embodiments of this application.

[0171] Those skilled in the art will understand that Figure 5 The structure shown does not constitute a limitation on vehicle 500 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0172] In addition, the device provided in the embodiments of this application may specifically be a chip, component or module. The chip may include a connected processor and a memory. The memory is used to store instructions. When the processor calls and executes the instructions, the chip can execute the output power control method provided in the above embodiments.

[0173] This embodiment also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the output power control method provided in the above embodiment.

[0174] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to achieve a method for controlling output power provided in the above embodiment.

[0175] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0176] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0177] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0178] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of controlling output power, characterized by, The method comprises: In the case that the target vehicle is in a driving state, based on a set of vehicle state parameters of the target vehicle, a set of motor operating parameters of the driving motor and a set of environmental parameters of the environment, a predicted temperature change trend of the driving motor and a load level of the target vehicle are determined; Based on the predicted temperature change trend, the load level and a subset of motor state parameters in the set of motor operating parameters, a power limiting factor of the target vehicle is determined, the power limiting factor being used to limit the output power of the driving motor; Based on the power limiting factor and an initial upper limit of power of the driving motor, the output power of the driving motor is limited.

2. The method of claim 1, wherein, The determination of the predicted temperature change trend of the driving motor and the load level of the target vehicle based on the set of vehicle state parameters of the target vehicle, the set of motor operating parameters of the driving motor and the set of environmental parameters of the environment comprises: Feature extraction is performed on the set of vehicle state parameters, the set of motor operating parameters and the set of environmental parameters by a first prediction model to obtain vehicle state features of the target vehicle, the first prediction model being a model obtained through multi-task joint training; Regression is performed on the vehicle state features by the first prediction model to obtain a plurality of predicted temperatures of the driving motor, the plurality of predicted temperatures being used to represent the predicted temperature change trend of the driving motor; Classification is performed on the vehicle state features based on the vehicle state features by the first prediction model to obtain the load level of the target vehicle.

3. The method of claim 2, wherein, The feature extraction on the set of vehicle state parameters, the set of motor operating parameters and the set of environmental parameters by the first prediction model to obtain the vehicle state features of the target vehicle comprises: Time alignment and preprocessing are performed on the parameters in the set of vehicle state parameters, the parameters in the set of motor operating parameters and the parameters in the set of environmental parameters to obtain a first input parameter sequence; The first input parameter sequence is input into the first prediction model, bidirectional time sequence encoding is performed on the first input parameter sequence by the first prediction model to obtain the vehicle state features of the target vehicle, or multiple convolutions or multiple full connections are performed on the first input parameter sequence by the first prediction model to obtain the vehicle state features of the target vehicle.

4. The method of claim 2, wherein, The regression on the vehicle state features by the first prediction model to obtain the plurality of predicted temperatures of the driving motor comprises: Multiple full connections are performed on the vehicle state features by a regression head of the first prediction model to obtain the plurality of predicted temperatures; The classification on the vehicle state features based on the vehicle state features by the first prediction model to obtain the load level of the target vehicle comprises: Full connection and normalization are performed on the vehicle state features by a classification head of the first prediction model to obtain a probability set, the probability set comprising a plurality of probabilities, one probability corresponding to one candidate load level; the candidate load level corresponding to the highest probability in the probability set is determined as the load level of the target vehicle.

5. The method of claim 1, wherein, The method comprises the following steps: determining a thermal risk zone to which the drive motor belongs and a position in the thermal risk zone to which the drive motor belongs based on the subset of motor state parameters; concatenating the predicted temperature change trend, the load level, the thermal risk zone, the position, and the subset of motor state parameters into a second input parameter sequence; inputting the second input parameter sequence into a second prediction model, processing the second input parameter sequence through the second prediction model, and obtaining the power limitation factor of the target vehicle, wherein the second prediction model is a model constructed based on a gating mechanism.

6. The method of claim 5, wherein, The method comprises the following steps: performing random nearest neighbor embedding of the subset of motor state parameters based on a preset distribution to obtain motor state features of the subset of motor state parameters; determining a matching degree between the motor state features and a plurality of candidate thermal risk zones; determining the thermal risk zone to which the drive motor belongs from the plurality of candidate thermal risk zones based on the matching degree between the motor state features and the plurality of candidate thermal risk zones, and determining the position of the drive motor in the thermal risk zone based on the matching degree between the motor state features and the thermal risk zone, wherein the thermal risk zone is a candidate thermal risk zone with the highest matching degree among the plurality of candidate thermal risk zones.

7. The method of claim 6, wherein, The method comprises the following steps: determining a feature similarity between the motor state features and a clustering center of each candidate thermal risk zone cluster; determining the feature similarity as the matching degree between the motor state features and the plurality of candidate thermal risk zones.

8. The method of claim 6, wherein, The method comprises the following steps: in a case where the matching degree is greater than or equal to a first matching degree threshold, determining the position in the thermal risk zone to which the drive motor belongs as a center position; in a case where the matching degree is greater than or equal to a second matching degree threshold and less than the first matching degree threshold, determining the position in the thermal risk zone to which the drive motor belongs as an intermediate position; in a case where the matching degree is less than the second matching degree threshold, determining the position in the thermal risk zone to which the drive motor belongs as an edge position.

9. The method of claim 6, wherein, The method for determining the plurality of candidate thermal risk zones comprises the following steps: obtaining a plurality of historical subsets of motor state parameters of the drive motor; performing random nearest neighbor embedding of each historical subset of motor state parameters based on the preset distribution to obtain historical motor state features of each subset of motor state parameters; performing clustering on the historical motor state features of the plurality of historical subsets of motor state parameters to obtain a preset number of clusters, and one cluster corresponds to one candidate thermal risk zone.

10. A vehicle characterized by comprising: The vehicle comprises: a memory for storing executable program code; a processor for calling and running the executable program code from the memory, so that the vehicle performs the output power control method as claimed in any one of claims 1 to 9.

Images