Knowledge distillation parameter optimization method and system for battery prediction and storage medium

Abstract

The present application relates to a kind of knowledge distillation parameter optimization method, system and storage medium for battery prediction.The method proposed in the present application reconstructs the knowledge distillation process as the energy evolution process of thermodynamic system, balances the exploration and utilization mechanism through the variational free energy formula, effectively prevents the model from falling into local optimum;Using Maxwell daemon agent to accurately adjust the physical control valve, the dynamic adaptability of knowledge transfer is enhanced;Based on the feedback mechanism of energy dissipation rate, more intensive physical guidance signal is provided, which significantly improves the convergence efficiency and generalization performance of reinforcement learning in battery performance prediction task.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method, system and storage medium for optimizing knowledge distillation parameters for battery prediction. Background Technology

[0002] In the context of the rapid development of smart manufacturing, renewable energy, and electric vehicles, the importance of battery performance prediction as a core energy component is self-evident. Accurate prediction of battery state of health (SOH), remaining useful life (RUL), and capacitance trends plays a crucial role in ensuring reliable system operation and extending battery life. In recent years, data-driven deep learning models, such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), One-Dimensional Convolutional Neural Networks (1D-CNN), Transformers, and Temporal Convolutional Networks (TCN), have been widely applied in multi-dimensional time-series battery data modeling. These models have demonstrated strong capabilities in time-series modeling and feature representation, but with increasing model complexity, a series of challenges have arisen. For example, the significant increase in the number of parameters leads to high computational costs, making efficient deployment on resource-constrained devices difficult, increasing hardware costs, and affecting the system's real-time response performance. Simultaneously, with the widespread application of machine learning and deep learning in various tasks, model performance and efficiency have gradually become core issues. While complex models can achieve excellent performance, they face the challenges of high computational costs and resource consumption.

[0003] Against this backdrop, model compression techniques have emerged, and knowledge distillation (KD), as an effective method, has received widespread attention. This method, by transferring knowledge from a complex teacher model to a simple student model, can reduce model complexity while maintaining good performance, thus becoming an important method for compressing battery performance prediction models. Existing knowledge distillation methods can be categorized as follows:

[0004] 1. Fixed weights, i.e., using fixed loss function weights. This method is simple and direct, but it cannot be dynamically adjusted according to the learning progress of the student model, data characteristics, and task complexity, resulting in poor adaptability in different scenarios; Second, improve the loss function, which involves optimizing the knowledge distillation process by adjusting the form of the loss function. For example, introduce additional regularization terms into the loss function to constrain the relationship between the student model and the teacher model. A common approach is to add constraints on the feature distributions of the student and teacher models, prompting the student model to learn feature representations more similar to the teacher model, thus enhancing knowledge transfer. However, this method still doesn't fundamentally solve the problem of dynamic hyperparameter adjustment. Under different tasks and data distributions, fixed hyperparameter settings are unlikely to achieve optimal results.

[0005] Third, meta-learning optimization aims to enable the model to quickly adapt to the hyperparameter settings of new tasks by training on multiple tasks. In teacher knowledge distillation, meta-learning algorithms can automatically find appropriate weights based on different datasets and tasks. However, meta-learning methods are usually computationally complex, requiring significant computational resources and time for training. Furthermore, due to differences between different tasks, the hyperparameters obtained through meta-learning have poor generality on new tasks and may not be well adapted to the needs of various classification and regression tasks. Summary of the Invention

[0006] This invention provides a method, system, and storage medium for optimizing knowledge distillation parameters for battery prediction, aiming to solve the technical problems of insufficient hyperparameter dynamism and low transfer efficiency in the knowledge distillation process of existing methods.

[0007] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for optimizing knowledge distillation parameters for battery prediction, comprising the following steps: S101. Collect battery operation data and construct a teacher model and a student model for battery performance prediction, wherein the teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function; S102. Using the battery operation data as the initial training set, the student model is pre-trained using a preset fixed hyperparameter in the teacher knowledge distillation loss function to obtain a pre-trained student model. S103. Randomly shuffle the initial training set to obtain a cyclic training set, and start cyclic training of the pre-trained student model. S104. Based on the input samples from the cyclic training set input to the pre-trained student model, construct thermodynamic comprehensive state information including internal energy index, system entropy, work progress and characteristic similar potential energy. S105. By using a preset intelligent agent to act as Maxwell's demon, feature extraction is performed on the thermodynamic integrated state information to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function. The dynamic physical control parameters include dynamic distillation temperature and free energy coupling strength. S106. Update the teacher knowledge distillation loss function according to the dynamic physical control parameters, and perform reverse update on the pre-trained student model to obtain the pre-output student model; S107. Calculate reward information to characterize energy dissipation efficiency based on the changes in the internal energy index before and after updating the parameters of the teacher knowledge distillation loss function, and update the parameters of the preset agent based on the reward information. S108. Determine whether the pre-output student model satisfies the preset convergence condition, wherein: If not, use the pre-output student model as the new pre-trained student model and return to step S103; If so, the pre-output student model is used as the battery performance prediction model, and step S109 is executed; S109. Input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction result.

[0008] Furthermore, the teacher knowledge distillation loss function in step S101 is constructed based on the Helmholtz free energy principle, and the teacher knowledge distillation loss function is defined as follows: It satisfies the following relationship: ; in, For predicting battery performance loss, Let be the variational free energy distillation term, and: ; To achieve dynamic distillation temperature The predicted distribution of the teacher model described below The predicted distribution of the student model The divergence between them represents the internal energy term of the system; The information entropy of the predicted distribution of the student model represents the system's thermal energy term; The free energy coupling strength is denoted as .

[0009] Furthermore, in step S104, the thermodynamic integrated state information is defined as follows: It satisfies the following relationship: ; in, The system entropy of the current student model uncertainty This represents the feature similarity potential between the feature layers of the student model and the teacher model.

[0010] Furthermore, in step S105, the dynamic distillation temperature and free energy coupling strength in the dynamic physical control parameters are defined as follows: It satisfies the following relationship: ; ; in, Basic distillation temperature, and The components of the output vector for the preset intelligent agent. It is the hyperbolic tangent function. This is the activation function.

[0011] Furthermore, in step S107, the reward information is defined as... It satisfies the following relationship: ; in, and These are the internal energy indices before and after the parameter update of the teacher knowledge distillation loss function, respectively. This is the positive incentive coefficient. This is an indicator function used to characterize whether the prediction accuracy has improved. , These represent the prediction accuracy before and after the parameter update, respectively.

[0012] Furthermore, the characteristic similarity potential energy Based on the feature layer of the student model With the feature layer of the teacher model cosine similarity The structure satisfies the following relationship: ; in .

[0013] Furthermore, the preset convergence condition includes at least one of the following: The rate of change of the value of the teacher knowledge distillation loss function during cyclic training is less than a preset change threshold. The difference between the battery performance prediction results output by the pre-output student model and the actual data is less than a preset performance threshold; The number of training rounds reaches the preset iteration threshold.

[0014] Secondly, the present invention also provides a knowledge distillation parameter optimization system for battery prediction, comprising: The data collection module is used to collect battery operation data and construct a teacher model and a student model for battery performance prediction. The teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function. The pre-training module is used to pre-train the student model using the battery operation data as the initial training set and using preset fixed hyperparameters in the teacher knowledge distillation loss function to obtain a pre-trained student model. The data processing module is used to randomly shuffle the initial training set to obtain a cyclic training set, and to start cyclic training on the pre-trained student model. The state characterization module is used to construct comprehensive thermodynamic state information, including internal energy index, system entropy, work progress, and characteristic similar potential energy, based on the input samples from the cyclic training set input to the pre-trained student model. The hyperparameter adjustment module is used to extract features from the thermodynamic integrated state information by using a preset intelligent agent as Maxwell's demon to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function. The dynamic physical control parameters include dynamic distillation temperature and free energy coupling strength. The knowledge distillation module is used to update the teacher knowledge distillation loss function according to the dynamic physical control parameters, and to perform reverse update on the pre-trained student model to obtain the pre-output student model. The agent update module is used to calculate reward information to characterize energy dissipation efficiency based on the changes in the internal energy index before and after the parameter update of the teacher knowledge distillation loss function, and to update the parameters of the preset agent based on the reward information. The loop module is used to determine whether the pre-output student model meets the preset convergence condition. If not, the pre-output student model is used as a new pre-trained student model and returned to the data processing module. If yes, the pre-output student model is output as a battery performance prediction model and the analysis and prediction module is executed. The analysis and prediction module is used to input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction results.

[0015] Thirdly, the present invention also provides a computer device, comprising: a memory, a processor, and a knowledge distillation parameter optimization program for battery prediction stored in the memory and executable on the processor, wherein the processor, when executing the knowledge distillation parameter optimization program for battery prediction, implements the steps in the knowledge distillation parameter optimization method for battery prediction as described in any of the above embodiments.

[0016] Fourthly, the present invention also provides a storage medium storing a knowledge distillation parameter optimization program for battery prediction, wherein the knowledge distillation parameter optimization program for battery prediction, when executed by a processor, implements the steps in the knowledge distillation parameter optimization method for battery prediction as described in any of the above embodiments.

[0017] The beneficial effects achieved by this invention lie in proposing a knowledge distillation parameter optimization method for battery prediction. This method reconstructs the knowledge distillation process as the energy evolution process of a thermodynamic system, balances the exploration and utilization mechanisms through the variational free energy formula, and effectively prevents the model from getting trapped in local optima. It utilizes Maxwell's demon-like agent to precisely adjust the physical control valves, enhancing the dynamic adaptability of knowledge transfer. The feedback mechanism based on the energy dissipation rate provides denser physical guidance signals, significantly improving the convergence efficiency and generalization performance of reinforcement learning in battery performance prediction tasks. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the knowledge distillation parameter optimization method for battery prediction provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the knowledge distillation parameter optimization system for battery prediction provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a knowledge distillation parameter optimization method for battery prediction provided in an embodiment of the present invention. The knowledge distillation parameter optimization method for battery prediction includes the following steps: S101. Collect battery operation data and construct a teacher model and a student model for battery performance prediction, wherein the teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function.

[0021] The battery operation data in this embodiment of the invention refers to multi-dimensional time-series data (such as voltage curves, charge / discharge currents, cycle counts, ambient temperature, etc.) collected in real time by battery sensors (voltage, current, temperature, internal resistance sensors) in scenarios such as electric vehicles and energy storage power stations. This battery operation data is initially aggregated by the battery management system or edge computing nodes and then transmitted to the local machine for processing. Since the raw data may contain noise (such as sensor errors), missing values, or abnormal fluctuations, step S101 also includes a preprocessing step for the battery operation data to facilitate data analysis and prediction. The preprocessing methods include, but are not limited to: Noise reduction is performed using methods such as moving average and wavelet transform to filter high-frequency noise and ensure the stability of key signals such as voltage and current. Normalization and standardization scale the data to the range of [-1, 1] or [0, 1] to avoid the model training effect being affected by the difference in feature scale. Time series alignment involves aligning the timestamps of asynchronous data from different sensors to generate a unified time series dataset.

[0022] In this embodiment of the invention, both the teacher model and the student model are neural network models used to analyze and predict battery performance using the battery operating data as input. In terms of model scale, the teacher model is larger than the student model. It should be noted that this embodiment of the invention primarily focuses on the concepts of teacher model and student model to implement the optimization method for relevant parameters in knowledge distillation technology; therefore, this invention does not specifically limit the specific architecture of the teacher model and student model.

[0023] S102. Using the battery operation data as the initial training set, the student model is pre-trained using preset fixed hyperparameters in the teacher knowledge distillation loss function to obtain the pre-trained student model.

[0024] Pre-training of the student model is an initialization strategy aimed at enabling the student model to learn task-related knowledge in its initial stages. It should be noted that pre-training also uses the complete battery operating data. In the scenario of battery performance prediction, pre-training enables the student model to understand the relationship between the battery operating data and a certain battery performance parameter it predicts.

[0025] Pre-training the student model can also stabilize the model gradient and initialize the weight relationship of the loss term. In this embodiment of the invention, the preset fixed hyperparameters are set to the following values: (e.g., set to 4.0) At this point, the intelligent agent does not intervene, and the model performs routine warm-up training.

[0026] Based on the relevant concepts of teacher and student models in the field of knowledge distillation technology, this invention creatively introduces the thermodynamic variational principle to construct a loss function. The teacher knowledge distillation loss function is constructed based on the Helmholtz free energy principle, and is defined as follows: It satisfies the following relationship: ; in, For predicting battery performance loss, Let be the variational free energy distillation term, and: ; To achieve dynamic distillation temperature The predicted distribution of the teacher model described below The predicted distribution of the student model The divergence between them (Kullback-Leibler) represents the internal energy of the system. The information entropy of the predicted distribution of the student model represents the system's thermal energy term; For dynamic distillation temperature, The free energy coupling strength is denoted as .

[0027] Unlike existing methods, this invention views knowledge distillation as a thermodynamic process of evolution from a high-energy state (untrained student) to a low-energy state (converging student). (Battery performance prediction task loss) Ensure the basic accuracy of battery prediction tasks (such as MSE loss in voltage prediction). Control the smoothness of the distribution (analogous to physical temperature). Control system entropy Contribution to total free energy. Teacher knowledge distillation loss function. By minimizing the free energy, the student model can learn not only the teacher's probability distribution during knowledge distillation (minimizing internal energy) but also... ), while maintaining appropriate uncertainty (maximizing entropy) This avoids overfitting.

[0028] S103. Randomly shuffle the initial training set to obtain a cyclic training set, and start cyclic training of the pre-trained student model.

[0029] Shuffling the initial training set is a variation strategy aimed at enhancing the generalization ability of the student model in different training iterations and avoiding overfitting. In this embodiment of the invention, the initial training set is defined as D, and the recurring training set is defined as B. Then, for each training iteration, shuffling the data to obtain the recurring training set B satisfies the following conditions: .

[0030] S104. Based on the input samples from the cyclic training set input to the pre-trained student model, construct thermodynamic comprehensive state information including internal energy index, system entropy, work progress, and characteristic similar potential energy.

[0031] In step S104, the status information It is no longer a simple feature aggregation, but a set of physical quantities reflecting the thermodynamic properties of the system. The comprehensive thermodynamic state information is defined as follows: It satisfies the following relationship: ; in, An internal energy index that reflects the current error potential energy. To reflect the system entropy that reflects the uncertainty of the current student model, To reflect the rate of work done in order to reflect the rate of energy dissipation, The feature similarity potential between the feature layers of the student model and the teacher model is usually represented by the cosine distance or Euclidean distance of the feature vectors.

[0032] Specifically, internal energy indicators It reflects the KL divergence between the teacher model and the student model at the current temperature under the current batch of data, and reflects the error potential of the student model in imitating the teacher. System entropy Output the information entropy of the distribution for the student model. This reflects the degree of uncertainty the model has about the current prediction; Work progress = This reflects the rate of energy dissipation during the training process, which is a physical measure of learning efficiency. Characteristic similarity potential energy Used to characterize student feature layers With teacher characteristic layer The structural difference potential energy between them ,in This represents cosine similarity. A lower value indicates that the student's feature structure is closer to the teacher's, conforming to the principle of energy minimization, and serving as an auxiliary structured constraint state. This state vector has a clear physical meaning and can completely map the thermodynamic evolution state of the current learning system.

[0033] S105. By using a preset intelligent agent to act as Maxwell's demon, feature extraction is performed on the thermodynamic integrated state information to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function. The dynamic physical control parameters include dynamic distillation temperature and free energy coupling strength.

[0034] In this step, the pre-defined agent plays the role of "Maxwell's Demon" in physics, capable of observing the motion state of microscopic particles (i.e.,...). It then operates the "thermal valve (parameter control module)". The preset intelligent agent outputs action vectors. And convert them into physical control parameters. The dynamic distillation temperature and free energy coupling strength in the dynamic physical control parameters are defined as follows: It satisfies the following relationship: ; ; in, Basic distillation temperature, and The components of the output vector for the preset intelligent agent. The hyperbolic tangent function is used to... Mapped to The range makes the temperature Fluctuations around the base temperature simulate the dynamic adjustment of the hot bath environment; This is an activation function used to... Mapped to The interval is used to adaptively control the coupling strength between the entropy term (exploration capability) and the internal energy term (utilization capability).

[0035] The relationship is used to dynamically adjust the distillation temperature. When the student model gets stuck in a local optimum, the agent can increase the temperature. Smoothing the probability distribution helps the model escape the potential well; as the model tends to converge, reducing... To sharpen the focus.

[0036] The relationship is used to dynamically adjust the free energy coupling strength. In the early stages of training, a larger... Encourage exploration (high entropy); reduce entropy in the later stages. The focus is on precise imitation (low internal energy).

[0037] S106. Update the teacher knowledge distillation loss function according to the dynamic physical control parameters, and perform reverse update on the pre-trained student model to obtain the pre-output student model.

[0038] S107. Calculate reward information to characterize energy dissipation efficiency based on the changes in the internal energy index before and after the parameter update of the teacher knowledge distillation loss function, and update the parameters of the preset agent based on the reward information.

[0039] In step S107, the reward function design follows the principle of minimum action or maximizes energy dissipation efficiency. The reward information is defined as... It satisfies the following relationship: ; in, and These are the internal energy indices (i.e., KL divergence loss) before and after the parameter update of the teacher knowledge distillation loss function, respectively. A positive difference indicates a decrease in internal energy (successful dissipation). This is the positive incentive coefficient. This is an indicator function used to characterize whether the prediction accuracy has improved, specifically when the prediction accuracy on the validation set... Additional rewards will be given upon promotion. , These represent the prediction accuracy before and after the parameter update, respectively.

[0040] The reward function design mechanism is directly related to physical internal energy. The amount of decrease. If the agent chooses a temperature... and coupling This led to a significant decrease in the internal energy (error) of the student model. If the result is a positive result (i.e., a positive reward is given), then a negative result (ii.e., a penalty is given) is given. This dense physical guidance signal is more effective in guiding the convergence of reinforcement learning than the traditional sparse accuracy reward.

[0041] During implementation, the pre-defined agent utilizes policy gradient algorithms (such as PPO or REINFORCE) to... Update the parameters.

[0042] S108. Determine whether the pre-output student model satisfies the preset convergence condition, wherein: If not, use the pre-output student model as the new pre-trained student model and return to step S103; If so, the pre-output student model is used as the battery performance prediction model, and step S109 is executed.

[0043] S109. Input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction result.

[0044] The preset convergence condition includes at least one of the following: The rate of change of the value of the teacher knowledge distillation loss function during cyclic training is less than a preset change threshold. The difference between the battery performance prediction results output by the pre-output student model and the actual data is less than a preset performance threshold; The number of training rounds reaches the preset iteration threshold.

[0045] The beneficial effects achieved by this invention lie in proposing a knowledge distillation parameter optimization method for battery prediction. This method reconstructs the knowledge distillation process as the energy evolution process of a thermodynamic system, balances the exploration and utilization mechanisms through the variational free energy formula, and effectively prevents the model from getting trapped in local optima. It utilizes Maxwell's demon-like agent to precisely adjust the physical control valves, enhancing the dynamic adaptability of knowledge transfer. The feedback mechanism based on the energy dissipation rate provides denser physical guidance signals, significantly improving the convergence efficiency and generalization performance of reinforcement learning in battery performance prediction tasks.

[0046] This invention also provides a knowledge distillation parameter optimization system 200 for battery prediction, please refer to... Figure 2 , Figure 2 This is a schematic diagram of the knowledge distillation parameter optimization system for battery prediction provided in an embodiment of the present invention, which includes: The data collection module 201 is used to collect battery operation data and construct a teacher model and a student model for battery performance prediction. The teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function. The pre-training module 202 is used to pre-train the student model using the battery running data as the initial training set and using preset fixed hyperparameters in the teacher knowledge distillation loss function to obtain a pre-trained student model. The data processing module 203 is used to randomly shuffle the initial training set to obtain a cyclic training set, and start cyclic training on the pre-trained student model. The state characterization module 204 is used to construct thermodynamic comprehensive state information including internal energy index, system entropy, work progress and characteristic similar potential energy based on the input samples from the cyclic training set input to the pre-trained student model. The hyperparameter adjustment module 205 is used to extract features from the thermodynamic integrated state information by using a preset intelligent agent as Maxwell's demon to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function. The dynamic physical control parameters include dynamic distillation temperature and free energy coupling strength. The knowledge distillation module 206 is used to update the teacher knowledge distillation loss function according to the dynamic physical control parameters, and to perform reverse update on the pre-trained student model to obtain the pre-output student model; The agent update module 207 is used to calculate reward information to characterize energy dissipation efficiency based on the change of the internal energy index before and after the parameter update of the teacher knowledge distillation loss function, and update the parameters of the preset agent based on the reward information. The loop module 208 is used to determine whether the pre-output student model meets the preset convergence condition. If not, the pre-output student model is used as a new pre-trained student model and the process is returned to the data processing module 203. If yes, the pre-output student model is output as a battery performance prediction model and the analysis and prediction module 209 is executed. The analysis and prediction module 209 is used to input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction result.

[0047] The knowledge distillation parameter optimization system 200 for battery prediction can implement the steps in the knowledge distillation parameter optimization method for battery prediction as described in the above embodiments, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0048] This invention also provides a computer device, please refer to... Figure 3 , Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. The computer device 300 includes: a memory 302, a processor 301, and a knowledge distillation parameter optimization program for battery prediction stored in the memory 302 and executable on the processor 301.

[0049] The processor 301 calls the knowledge distillation parameter optimization program for battery prediction stored in the memory 302, and executes the steps in the knowledge distillation parameter optimization method for battery prediction provided in this embodiment of the invention. Please refer to... Figure 1 Specifically, it includes the following steps: S101. Collect battery operation data and construct a teacher model and a student model for battery performance prediction, wherein the teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function.

[0050] The teacher knowledge distillation loss function in step S101 is constructed based on the Helmholtz free energy principle, and is defined as follows: It satisfies the following relationship: ; in, For predicting battery performance loss, Let be the variational free energy distillation term, and: ; To achieve dynamic distillation temperature The predicted distribution of the teacher model described below The predicted distribution of the student model The divergence between them represents the internal energy term of the system; The information entropy of the predicted distribution of the student model represents the system's thermal energy term; The free energy coupling strength is denoted as .

[0051] S102. Using the battery operation data as the initial training set, the student model is pre-trained using preset fixed hyperparameters in the teacher knowledge distillation loss function to obtain the pre-trained student model.

[0052] S103. Randomly shuffle the initial training set to obtain a cyclic training set, and start cyclic training of the pre-trained student model.

[0053] S104. Based on the input samples from the cyclic training set input to the pre-trained student model, construct comprehensive thermodynamic state information including internal energy index, system entropy, work progress, and characteristic similar potential energy. In step S104, the thermodynamic integrated state information is defined as follows: It satisfies the following relationship: ; in, The system entropy of the current student model uncertainty This represents the feature similarity potential between the feature layers of the student model and the teacher model.

[0054] The characteristic similar potential energy Based on the feature layer of the student model With the feature layer of the teacher model cosine similarity The structure satisfies the following relationship: ; in .

[0055] S105. By using a preset intelligent agent to act as Maxwell's demon, feature extraction is performed on the thermodynamic integrated state information to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function.

[0056] In step S105, the dynamic distillation temperature and free energy coupling strength in the dynamic physical control parameters are defined as follows: It satisfies the following relationship: ; ; in, Basic distillation temperature, and The components of the output vector for the preset intelligent agent. It is the hyperbolic tangent function. This is the activation function.

[0057] S106. Update the teacher knowledge distillation loss function according to the dynamic hyperparameters, and perform reverse update on the pre-trained student model to obtain the pre-output student model.

[0058] S107. Calculate reward information to characterize energy dissipation efficiency based on the state information and the dynamic physical control parameters, and update the parameters of the preset intelligent agent based on the reward information.

[0059] In step S107, the reward information is defined as follows: It satisfies the following relationship: ; in, and These are the internal energy indices before and after the parameter update of the teacher knowledge distillation loss function, respectively. This is the positive incentive coefficient. This is an indicator function used to characterize whether the prediction accuracy has improved. , These represent the prediction accuracy before and after the parameter update, respectively.

[0060] S108. Determine whether the pre-output student model satisfies the preset convergence condition, wherein: If not, use the pre-output student model as the new pre-trained student model and return to step S103; If so, the pre-output student model is used as the battery performance prediction model, and step S109 is executed.

[0061] S109. Input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction result.

[0062] The preset convergence condition includes at least one of the following: The rate of change of the value of the teacher knowledge distillation loss function during cyclic training is less than a preset change threshold. The difference between the battery performance prediction results output by the pre-output student model and the actual data is less than a preset performance threshold; The number of training rounds reaches the preset iteration threshold.

[0063] The computer device 300 provided in this embodiment of the invention can implement the steps in the knowledge distillation parameter optimization method for battery prediction as described in the above embodiments, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0064] This invention also provides a storage medium storing a knowledge distillation parameter optimization program for battery prediction. When executed by a processor, this knowledge distillation parameter optimization program for battery prediction implements the various processes and steps in the knowledge distillation parameter optimization method for battery prediction provided in this invention, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0065] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a mobile phone, computer, server, air conditioner, or network device, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0066] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0067] The embodiments of the present invention have been described above with reference to the accompanying drawings. The disclosed embodiments are merely preferred embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many equivalent changes in form without departing from the spirit and scope of the claims of the present invention, and all such changes are within the protection scope of the present invention.

Claims

1. A knowledge distillation parameter optimization method for battery prediction, characterized in that, Includes the following steps: S101. Collect battery operation data and construct a teacher model and a student model for battery performance prediction, wherein the teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function; S102. Using the battery operation data as the initial training set, the student model is pre-trained using a preset fixed hyperparameter in the teacher knowledge distillation loss function to obtain a pre-trained student model. S103. Randomly shuffle the initial training set to obtain a cyclic training set, and start cyclic training of the pre-trained student model. S104. Based on the input samples from the cyclic training set input to the pre-trained student model, construct thermodynamic comprehensive state information including internal energy index, system entropy, work progress and characteristic similar potential energy. S105. By using a preset intelligent agent to act as Maxwell's demon, feature extraction is performed on the thermodynamic integrated state information to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function. The dynamic physical control parameters include dynamic distillation temperature and free energy coupling strength. S106. Update the teacher knowledge distillation loss function according to the dynamic physical control parameters, and perform reverse update on the pre-trained student model to obtain the pre-output student model; S107. Calculate reward information to characterize energy dissipation efficiency based on the changes in the internal energy index before and after updating the parameters of the teacher knowledge distillation loss function, and update the parameters of the preset agent based on the reward information. S108. Determine whether the pre-output student model satisfies the preset convergence condition, wherein: If not, use the pre-output student model as a new pre-trained student model and return to step S103; If so, the pre-output student model is used as the battery performance prediction model, and step S109 is executed; S109. Input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction result.

2. The knowledge distillation parameter optimization method for battery prediction according to claim 1, characterized in that, The teacher knowledge distillation loss function in step S101 is constructed based on the Helmholtz free energy principle, and is defined as follows: It satisfies the following relationship: ； in, For predicting battery performance loss, Let be the variational free energy distillation term, and: ； To achieve dynamic distillation temperature The predicted distribution of the teacher model described below The predicted distribution of the student model The divergence between them represents the internal energy term of the system; The information entropy of the predicted distribution of the student model represents the system's thermal energy term; For dynamic distillation temperature, The free energy coupling strength is denoted as .

3. The knowledge distillation parameter optimization method for battery prediction according to claim 1, characterized in that, In step S104, the thermodynamic integrated state information is defined as follows: It satisfies the following relationship: ； in, An internal energy index that reflects the current error potential energy. To reflect the system entropy that reflects the uncertainty of the current student model, To reflect the rate of work done in order to reflect the rate of energy dissipation, This represents the feature similarity potential between the feature layers of the student model and the teacher model.

4. The knowledge distillation parameter optimization method for battery prediction according to claim 1, characterized in that, In step S105, the dynamic distillation temperature and free energy coupling strength in the dynamic physical control parameters are defined as follows: It satisfies the following relationship: ； ； in, Basic distillation temperature, and The components of the output vector for the preset intelligent agent. It is the hyperbolic tangent function. This is the activation function.

5. The knowledge distillation parameter optimization method for battery prediction according to claim 3, characterized in that, In step S107, the reward information is defined as follows: It satisfies the following relationship: ； in, and These are the internal energy indices before and after the parameter update of the teacher knowledge distillation loss function, respectively. This is the positive incentive coefficient. This is an indicator function used to characterize whether the prediction accuracy has improved. , These represent the prediction accuracy before and after the parameter update, respectively.

6. The knowledge distillation parameter optimization method for battery prediction according to claim 3, characterized in that, The characteristic similar potential energy Based on the feature layer of the student model With the feature layer of the teacher model cosine similarity The structure satisfies the following relationship: ； in, .

7. The knowledge distillation parameter optimization method for battery prediction according to claim 1, characterized in that, The preset convergence condition includes at least one of the following: The rate of change of the value of the teacher knowledge distillation loss function during cyclic training is less than a preset change threshold. The difference between the battery performance prediction results output by the pre-output student model and the actual data is less than the preset performance threshold; The number of training rounds reaches the preset iteration threshold.

8. A knowledge distillation parameter optimization system for battery prediction, characterized in that, include: The data collection module is used to collect battery operation data and construct a teacher model and a student model for battery performance prediction. The teacher model guides the student model to perform knowledge distillation learning based on the teacher knowledge distillation loss function. The pre-training module is used to pre-train the student model using the battery operation data as the initial training set and using preset fixed hyperparameters in the teacher knowledge distillation loss function to obtain a pre-trained student model. The data processing module is used to randomly shuffle the initial training set to obtain a cyclic training set, and to start cyclic training on the pre-trained student model. The state representation module is used to construct thermodynamic comprehensive state information, including internal energy index, system entropy, work progress, and characteristic similar potential energy, based on the input samples from the cyclic training set input to the pre-trained student model. The hyperparameter adjustment module is used to extract features from the thermodynamic integrated state information by using a preset intelligent agent as Maxwell's demon to obtain dynamic physical control parameters for adjusting the teacher knowledge distillation loss function. The dynamic physical control parameters include dynamic distillation temperature and free energy coupling strength. The knowledge distillation module is used to update the teacher knowledge distillation loss function according to the dynamic physical control parameters, and to perform reverse update on the pre-trained student model to obtain the pre-output student model. The agent update module is used to calculate reward information to characterize energy dissipation efficiency based on the changes in the internal energy index before and after the parameter update of the teacher knowledge distillation loss function, and to update the parameters of the preset agent based on the reward information. The loop module is used to determine whether the pre-output student model meets the preset convergence condition. If not, the pre-output student model is used as a new pre-trained student model and returned to the data processing module. If yes, the pre-output student model is output as a battery performance prediction model and the analysis and prediction module is executed. The analysis and prediction module is used to input the battery operation data into the battery performance prediction model for analysis and prediction, and output the battery performance prediction results.

9. A computer device, characterized in that, include: The system includes a memory, a processor, and a knowledge distillation parameter optimization program for battery prediction stored in the memory and executable on the processor. When the processor executes the knowledge distillation parameter optimization program for battery prediction, it implements the steps of the knowledge distillation parameter optimization method for battery prediction as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a knowledge distillation parameter optimization program for battery prediction, which, when executed by a processor, implements the steps in the knowledge distillation parameter optimization method for battery prediction as described in any one of claims 1-7.

Images