Battery SOC prediction method based on SMoE model and expansion stress characteristics

Through the battery SOC prediction method based on SMoE model and expansion stress characteristics, the problem of SOC data inaccuracy caused by BMS failure is solved, and more accurate and robust battery SOC prediction is achieved, providing a new perspective on the internal state of the battery and reducing the Calculation and prediction error.

CN118938025AInactive Publication Date: 2024-11-12NANTONG LE CHUANGXIN ENERGY CO LTD

Patent Information

Application Number
CN202411404707.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-10
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

BMS failure or error causes inaccurate battery SOC data, and it is difficult for the prior art to make effective SOC prediction in this case.

Method used

Using the battery SOC prediction method based on SMoE model and expansion stress characteristics, a multi-expert model and a gated network model are constructed by testing and collecting battery operation data, and the Gumbel-Softmax algorithm is used to achieve sparse activation of the expert sub-model and pass through linear regression layer. Output battery SOC.

Benefits of technology

Improves the accuracy of SOC prediction and the robustness of the model, reduces the risk of overfitting, and enables accurate SOC prediction in case of BMS failure or error, provides a new perspective on the internal state of the battery, and reduces the calculation amount and prediction error. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118938025A_ABST
    Figure CN118938025A_ABST
Patent Text Reader

Abstract

The invention discloses a battery SOC prediction method based on an SMoE model and expansion stress characteristics, and the method comprises the steps: testing and collecting battery operation data and SOC data, and the battery operation data comprises expansion stress, current, voltage and temperature; constructing an SMoE model, wherein the SMoE model comprises a multi-expert model and a gating network model; the SMoE model is trained and verified; and the SOC of the battery is predicted, and after the training of the SMoE model is completed, on the premise that the model effect is good, the model is deployed, and the SOC of the battery is predicted by using the model. According to the method, the SOC of the battery is predicted through the SMoE model, the expression ability and prediction precision of the model are improved, the robustness and generalization ability of the whole model are improved, and the over-fitting risk is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a battery SOC prediction method, in particular to a battery SOC prediction method based on a SMoE model and expansion stress characteristics, belonging to the technical field of battery detection. Background Art

[0002] Battery SOC (State of Charge) is a measure of the remaining charge in a battery, which reflects the actual current state of charge of the battery. SOC is very important for the battery management system (BMS) because it is one of the key parameters for monitoring and controlling the battery charging and discharging process. Here are some important applications of SOC: 1. User interface display: In mobile phones, electric vehicles (EVs), and other portable electronic devices, SOC is used to display the remaining charge to the user. 2. Charging control: The battery charging system uses SOC to decide when to start and stop charging. 3. Discharge management: When the battery is discharged, SOC helps the system avoid over-discharge, which may damage the battery or reduce its life. 4. Battery health monitoring: Long-term monitoring of SOC helps assess the health and aging of the battery. 5. Range estimation: In electric vehicles, SOC is combined with other factors (such as driving habits, road conditions, etc.) to estimate the remaining range.

[0003] Accurate measurement and estimation of SOC is critical to the performance and life management of batteries. Since the characteristics of batteries change over time, usage conditions and environmental factors, SOC measurement usually requires complex algorithms and models to ensure accuracy.

[0004] During the operation of the BMS, failures or errors may occur due to various reasons, causing the BMS to stop operating or generate obviously erroneous battery SOC data. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a battery SOC prediction method based on the SMoE model and expansion stress characteristics, and obtain the current battery SOC data through model prediction when a BMS fails or an error occurs.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A battery SOC prediction method based on the SMoE model and expansion stress characteristics, characterized by comprising the following steps: S1. Test and collect battery operation data and SOC data. The battery operation data includes expansion stress, current, voltage and temperature. S2. Build the SMoE model, which includes a multi-expert model and a gating network model; S3, training and verifying the SMoE model; S4. Predict the battery SOC. After the SMoE model is trained, if the model works well, deploy the model and use it to predict the battery SOC.

[0007] Furthermore, the collection of battery operation data in step S1 is specifically as follows: Expansion stress collection: Install a micro pressure sensor inside or outside the battery to collect expansion stress, and connect the signal of the micro pressure sensor to a data logger to record real-time data of expansion stress; Current acquisition, measuring the current flowing through the battery through a current sensor, which uses a Hall effect sensor or a shunt resistor, connecting the current sensor to a data acquisition system and recording the current data; Voltage acquisition: a high-precision voltmeter or a voltage measurement circuit built into the battery management system is used to collect the terminal voltage of the battery. The voltage data is recorded by the data acquisition system. Temperature collection: A temperature sensor is used to collect the temperature of the battery. The temperature sensor uses a thermocouple, thermistor or infrared sensor. The temperature sensor is connected to the data acquisition system and records the temperature data.

[0008] Furthermore, the SOC data collection in step S1 is specifically as follows: the SOC is calculated by a battery management system.

[0009] Furthermore, the multi-expert model in step S2 includes three expert sub-models, each of which is a multi-layer perceptron model, and each of which includes an expert sub-model input layer, an expert sub-model first intermediate layer, an expert sub-model second intermediate layer, and an expert sub-model output layer. The expert sub-model input layer receives input battery operation data, the dimension of the expert sub-model first intermediate layer is 64, the dimension of the expert sub-model second intermediate layer is 32, and the expert sub-model output layer outputs a 16-dimensional vector.

[0010] Furthermore, in step S2, the gated network model adopts a multi-layer perceptron model, and the gated network model includes a gated network model input layer, a gated network model middle layer and a gated network model output layer. The gated network model input layer receives the input battery operation data, the dimension of the gated network model middle layer is 32 dimensions, and the activation function is the relu function. The gated network model output layer outputs a 3-dimensional vector, and the activation function is softmax. The 3-dimensional vector corresponds to the weights of the three expert sub-models respectively and the sum of the weights of the three expert sub-models is 1. The Gumbel-Softmax algorithm is used to further assign weights to the output of the gated network to achieve sparse activation of the expert sub-model.

[0011] Furthermore, the Gumbel-Softmax algorithm is used to further weight the output of the gating network to achieve sparse activation of the expert sub-model: Initial logits: Assume that we have a logits vector z containing N elements, which represents the pre-activation score for each expert sub-model, that is, the weight vector of the expert sub-model of the output layer of the gating network model; Add Gumbel noise: Add a noise g sampled from a Gumbel (0, 1) distribution to each logits: z′=z+g Where g=−log(−log(u)), u is a vector sampled from the uniform distribution U(0,1), z is the logits before adding Gumbel noise, and z′ is the logits after adding Gumbel noise; Softmax mapping: Apply the Softmax function to the logits z′ with Gumbel noise added to obtain a probability distribution π:

[0012] Among them, π i represents the i-th element in the probability distribution π, that is, the probability of the i-th category; z ’ i represents the i-th element of the logits vector after adding Gumbel noise; z ’ j represents the jth element of the logits vector after adding Gumbel noise; τ is the temperature parameter, which controls the sharpness of the softmax distribution; when τ approaches 0, the distribution becomes very sharp, close to a one-hot vector, and when τ approaches infinity, the distribution approaches a uniform distribution; The 3D vector output by the output layer of the gated network model is used as the input of the Gumbel-Softmax algorithm. The Gumbel-Softmax algorithm re-outputs a 3D vector as the weight of the expert sub-model by adding Gumbel noise and Softmax mapping. Then the expert sub-model weights are weighted and summed with the outputs of the corresponding expert sub-models to obtain a comprehensive feature representation; Afterwards, the comprehensive feature representation is connected to a fully connected layer; Finally, the battery SOC is output through a linear regression layer.

[0013] Furthermore, the step S3 is specifically as follows: the SMoE model code is implemented and trained on the Keras platform, the number of training epochs=1000, the number of batch samples batch_size=1024; the optimizer for model training is the adam algorithm, and the loss function is the mean square error; the training data is divided into a training set and a test set in a ratio of 8:2, the model is trained on the training set, and the model is verified on the test set.

[0014] Compared with the prior art, the present invention has the following advantages and effects: 1. The present invention predicts the battery SOC through the SMoE model, which improves the expression ability and prediction accuracy of the model, improves the robustness and generalization ability of the overall model, and reduces the risk of overfitting. In the case of a failure or error in the BMS, the battery SOC can be accurately predicted through the software model to replace the BMS or multiple BMS error data for correction, thereby ensuring the accuracy of the battery SOC data; 2. The present invention adds expansion stress for battery SOC prediction. Expansion stress is a unique indicator of the internal state of the battery, reflecting the internal pressure changes caused by chemical reactions during the charge and discharge process of the battery. Using this feature can provide a new perspective on the battery state, which may help to more accurately predict the SOC. Expansion stress can supplement traditional current, voltage and temperature signals to provide more comprehensive battery state information, thereby improving the accuracy of SOC prediction. Expansion stress is directly related to the internal chemical reaction of the battery, and it can reveal changes in the internal state of the battery, which helps to reduce the prediction error that relies only on external measurement parameters. 3. The present invention adopts the SMoE model, which allows different experts to learn specific aspects of the data. For example, some experts may focus on current changes, while other experts may focus on voltage patterns. This multi-expert model can improve the accuracy of predictions. At the same time, through the sparse selection mechanism, only part of the expert network is activated, which significantly reduces the amount of calculation and improves the training and prediction efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a battery SOC prediction method based on the SMoE model and expansion stress characteristics of the present invention.

[0016] Figure 2 It is a structural diagram of the SOC prediction model based on the SMoE model of the present invention.

[0017] Figure 3 It is the expert sub-model structure diagram of the present invention.

[0018] Figure 4 It is a structural diagram of the gated network model of the present invention.

[0019] Figure 5 It is a table of battery expansion stress and other characteristics and SOC data set samples according to an embodiment of the present invention.

[0020] Figure 6 It is a data table of prediction result samples of an embodiment of the present invention.

[0021] Figure 7 It is a bar chart of a sample of prediction results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to elaborate on the technical scheme adopted by the present invention to achieve the predetermined technical purpose, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments, and the technical means or technical features in the embodiments of the present invention can be replaced without paying creative work. The present invention will be described in detail below with reference to the drawings and in conjunction with the embodiments.

[0023] like Figure 1 As shown, a battery SOC prediction method based on the SMoE model and expansion stress characteristics of the present invention is characterized by comprising the following steps: S1. Test and collect battery operation data and SOC data. The battery operation data includes expansion stress, current, voltage and temperature.

[0024] The collection of battery operation data in step S1 is specifically as follows: Expansion stress collection: Install a micro pressure sensor inside or outside the battery to collect expansion stress, and connect the signal of the micro pressure sensor to a data logger to record real-time data of expansion stress.

[0025] Current acquisition, the current flowing through the battery is measured by a current sensor. The current sensor uses a Hall effect sensor or a shunt resistor. The current sensor is connected to the data acquisition system and the current data is recorded.

[0026] Voltage acquisition uses a high-precision voltmeter or a voltage measurement circuit built into the battery management system to collect the terminal voltage of the battery, and the voltage data is recorded through the data acquisition system.

[0027] Temperature collection: A temperature sensor is used to collect the temperature of the battery. The temperature sensor uses a thermocouple, a thermistor (NTC or PTC) or an infrared sensor. The temperature sensor is connected to the data acquisition system and records the temperature data.

[0028] The specific process of collecting the above parameters is as follows: Design the sensor layout: Make sure that the sensor is placed in a position that can accurately reflect the battery status. Install the sensor: Install the sensor on the battery or battery management system. Calibrate the sensor: Calibrate the sensor to ensure that its readings are accurate and reliable. Connect the data acquisition system: Connect all sensors to the data acquisition system. Data synchronization: Ensure that all sensor data is synchronized during collection to enable accurate correlation analysis. Data recording: Record data in real time during the battery's charging and discharging process. Data processing: Pre-process the collected data, such as filtering, normalization, etc., for subsequent analysis. Through the above steps, relevant data of the battery under normal working conditions can be effectively collected.

[0029] Expansion stress data can also be obtained during battery operation. The battery will expand and shrink in volume during the charge and discharge process, which will produce internal expansion stress. The change in internal pressure of the battery is related to the intensity of the electrochemical reaction and can indirectly reflect the change in SOC. Long-term pressure changes will affect the structural integrity and performance of the battery, which is also an indicator for monitoring the health of the battery. Adding expansion stress data can increase the accuracy of battery SOC prediction. Combining multi-dimensional data such as expansion stress, current, voltage and temperature, the battery SOC can be accurately monitored and predicted.

[0030] The SOC data collection in step S1 is specifically as follows: SOC is calculated by the battery management system, which combines current, voltage and other sensor data to estimate the battery SOC. The types of data collected by the sensors that can be used in the data collection stage are not limited to the expansion stress, current, voltage and temperature data during battery operation. More parameters can be used to obtain accurate battery SOC as a sample for the prediction model.

[0031] S2, build SMoE model, such as Figure 2 As shown in Figure 1, the SMoE model includes a multi-expert model and a gating network model.

[0032] like Figure 3 As shown, the multi-expert model in step S2 includes three expert sub-models, each of which is a multi-layer perceptron model, and each of which includes an expert sub-model input layer, an expert sub-model first intermediate layer, an expert sub-model second intermediate layer, and an expert sub-model output layer. The expert sub-model input layer receives input battery operation data, the dimension of the expert sub-model first intermediate layer is 64, the dimension of the expert sub-model second intermediate layer is 32, and the expert sub-model output layer outputs a 16-dimensional vector.

[0033] like Figure 4As shown, in step S2, the gated network model adopts a multi-layer perceptron model, and the gated network model includes a gated network model input layer, a gated network model middle layer and a gated network model output layer. The gated network model input layer receives the input battery operation data, the dimension of the gated network model middle layer is 32 dimensions, and the activation function is the relu function. The gated network model output layer outputs a 3D vector, and the activation function is softmax. The 3D vectors correspond to the weights of the three expert sub-models respectively and the sum of the weights of the three expert sub-models is 1. The Gumbel-Softmax algorithm is used to further distribute the weights of the output of the gated network to achieve sparse activation of the expert sub-model.

[0034] The outputs of the three expert sub-models are output1, output2 and output3 respectively. These three outputs are all high-dimensional vectors. Taking 32 dimensions as an example, they represent the feature representation of the input data by each sub-expert model.

[0035] The Gumbel-Softmax algorithm is used to further weight the output of the gating network to achieve sparse activation of the expert sub-model: Initial logits: Assume that we have a logits vector z containing N elements, which represents the pre-activation score for each expert sub-model, that is, the weight vector of the expert sub-model of the output layer of the gating network model; Add Gumbel noise: Add a noise g sampled from a Gumbel (0, 1) distribution to each logits: z′=z+g Where g=−log(−log(u)), u is a vector sampled from the uniform distribution U(0,1), z is the logits before adding Gumbel noise, and z′ is the logits after adding Gumbel noise; Softmax mapping: Apply the Softmax function to the logits z′ with Gumbel noise added to obtain a probability distribution π:

[0036] Among them, π i represents the i-th element in the probability distribution π, that is, the probability of the i-th category; z ’ i represents the i-th element of the logits vector after adding Gumbel noise; z ’ jrepresents the jth element of the logits vector after adding Gumbel noise; τ is the temperature parameter, which controls the sharpness of the softmax distribution; when τ approaches 0, the distribution becomes very sharp, close to a one-hot vector, and when τ approaches infinity, the distribution approaches a uniform distribution; The 3D vector output by the output layer of the gated network model is used as the input of the Gumbel-Softmax algorithm. The Gumbel-Softmax algorithm re-outputs a 3D vector as the weight of the expert sub-model by adding Gumbel noise and Softmax mapping. Then the expert sub-model weights are weighted and summed with the outputs of the corresponding expert sub-models to obtain a comprehensive feature representation; Afterwards, the comprehensive feature representation is connected to a fully connected layer; Finally, the battery SOC is output through a linear regression layer.

[0037] Gumbel-Softmax is a technique for implementing differentiable sampling in neural networks, which allows us to use gradient descent to optimize discretely chosen parameters during training.

[0038] Gumbel-Softmax combines Gumbel distribution and Softmax function to make the originally non-differentiable discrete sampling process differentiable. Gumbel distribution is a continuous distribution used for sampling index or sorting, while Softmax function is used to convert real number vectors into probability distributions.

[0039] Traditional machine learning models often use a single model to fit the entire data set, which is inadequate when faced with complex and highly heterogeneous data. Such models have limited generalization capabilities, are prone to overfitting problems, and are inefficient when processing large-scale data.

[0040] The Sparse Mixture of Experts (SMoE) model is a machine learning architecture designed for complex prediction and classification problems. The core of the model is to use multiple "expert" sub-models, each of which specializes in processing a specific part of the dataset or a specific type of input. The predictions of these sub-models are weighted and aggregated through a gating network to obtain the final prediction.

[0041] The "sparse" nature of SMoE is reflected in that, at each prediction, not all expert sub-models are activated, but only a portion are selected to participate in the calculation. This selective activation mechanism greatly reduces the computational cost and improves the operational efficiency of the model.

[0042] The implementation of sparse strategy can be achieved through some algorithms such as Gumbel algorithm, which selects a few experts to be activated among multiple experts, while the majority of others remain inactive.

[0043] The sparse strategy of the present application is different from the traditional sparse strategy. In the present invention, multiple expert sub-models are all activated, and all expert sub-models output a vector as a result, rather than just outputting a result value. Then, a weight vector is calculated for the output vector of each expert sub-model through the gated network model, and the output vectors of all expert sub-models and the weight vector of the gated network model are weighted and summed to finally obtain the battery SOC result. Compared with the expert model of the traditional sparse strategy, the result obtained is more accurate.

[0044] S3. Train and verify the SMoE model.

[0045] Step S3 specifically implements the SMoE model code on the Keras platform and trains it. The number of training epochs=1000 and the number of batch samples batch_size=1024 are used during training. The optimizer for model training is the adam algorithm, and the loss function is the mean square error. The training data is divided into a training set and a test set in a ratio of 8:2. The model is trained on the training set and verified on the test set.

[0046] S4. Predict the battery SOC. After the SMoE model is trained, the model is deployed and used to predict the battery SOC if the model works well. The prediction target of the SMoE model is SOC. The loss function of the model is the mean square error (MSE); the optimizer is adam; the evaluation metric is the mean absolute error (MAE), which is the sum of the absolute values ​​of the difference between the target value and the predicted value.

[0047] In one embodiment of the present invention, the collected battery operation data is as follows: Figure 5 As shown, it is input into the trained SMoE model, and the prediction results are as follows Figure 6 and Figure 7 As shown in the figure, the mean of the SOC test set is 2.1866, the standard deviation is 1.2831, and the mean of the prediction set is 2.1961, the standard deviation is 1.2791, which shows that the data distribution of the prediction results is basically consistent with the data distribution of the real data.

[0048] The root mean square error between the prediction set and the test set of battery SOC is only 0.0192, indicating that the prediction effect of the model is good.

[0049] The present invention predicts the battery SOC through the SMoE model, improves the expression ability and prediction accuracy of the model, improves the robustness and generalization ability of the overall model, reduces the risk of overfitting, and can accurately predict the battery SOC through the software model in the event of a BMS failure or error, so as to replace the BMS or multiple BMS error data for correction, thereby ensuring the accuracy of the battery SOC data; the present invention adds expansion stress for battery SOC prediction, which is a unique indicator of the internal state of the battery, reflecting the internal pressure changes caused by chemical reactions during the charge and discharge process of the battery; using this feature can provide a new perspective on the battery state, which may help to more accurately predict SOC OC; expansion stress can supplement traditional current, voltage and temperature signals, provide more comprehensive battery state information, and thus improve the accuracy of SOC prediction; expansion stress is directly related to the internal chemical reaction of the battery, and it can reveal the changes in the internal state of the battery, which helps to reduce the prediction error that relies solely on external measurement parameters; the present invention adopts the SMoE model, which allows different experts to learn specific aspects of the data. For example, some experts may focus on current changes, while other experts may focus on voltage patterns. This multi-expert model can improve the accuracy of prediction; at the same time, through the sparse selection mechanism, only part of the expert network is activated, which significantly reduces the amount of calculation and improves the training and prediction efficiency of the model.

[0050] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A battery SOC prediction method based on the SMoE model and expansion stress characteristics, characterized in that The following steps are involved: S1. Test and collect battery operation data and SOC data. The battery operation data includes expansion stress, current, voltage and temperature. S2. Build the SMoE model, which includes a multi-expert model and a gating network model; S3, training and verifying the SMoE model; S4. Predict the battery SOC. After the SMoE model is trained, if the model works well, deploy the model and use it to predict the battery SOC.

2. The battery SOC prediction method based on the SMoE model and expansion stress characteristics according to claim 1, characterized in that: The collection of battery operation data in step S1 is specifically as follows: Expansion stress collection: Install a micro pressure sensor inside or outside the battery to collect expansion stress, and connect the signal of the micro pressure sensor to a data logger to record real-time data of expansion stress; Current acquisition, measuring the current flowing through the battery through a current sensor, which uses a Hall effect sensor or a shunt resistor, connecting the current sensor to a data acquisition system and recording the current data; Voltage acquisition: a high-precision voltmeter or a voltage measurement circuit built into the battery management system is used to collect the terminal voltage of the battery. The voltage data is recorded by the data acquisition system. Temperature collection: A temperature sensor is used to collect the temperature of the battery. The temperature sensor uses a thermocouple, thermistor or infrared sensor. The temperature sensor is connected to the data acquisition system and records the temperature data.

3. The battery SOC prediction method based on the SMoE model and expansion stress characteristics according to claim 1, characterized in that: The SOC data collection in step S1 is specifically as follows: the SOC is calculated by the battery management system.

4. The battery SOC prediction method based on the SMoE model and expansion stress characteristics according to claim 1 is characterized in that: In step S2, the multi-expert model includes three expert sub-models, each of which is a multi-layer perceptron model. Each expert sub-model includes an expert sub-model input layer, an expert sub-model first intermediate layer, an expert sub-model second intermediate layer and an expert sub-model output layer. The expert sub-model input layer receives input battery operation data, the dimension of the expert sub-model first intermediate layer is 64, the dimension of the expert sub-model second intermediate layer is 32, and the expert sub-model output layer outputs a 16-dimensional vector.

5. The battery SOC prediction method based on the SMoE model and expansion stress characteristics according to claim 1, characterized in that: In step S2, the gated network model adopts a multi-layer perceptron model, and the gated network model includes a gated network model input layer, a gated network model middle layer and a gated network model output layer. The gated network model input layer receives the input battery operation data, the dimension of the gated network model middle layer is 32 dimensions, and the activation function is the relu function. The gated network model output layer outputs a 3-dimensional vector, and the activation function is softmax. The 3-dimensional vector corresponds to the weights of the three expert sub-models respectively and the sum of the weights of the three expert sub-models is 1. The Gumbel-Softmax algorithm is used to further weight the output of the gated network to achieve sparse activation of the expert sub-model.

6. The battery SOC prediction method based on the SMoE model and expansion stress characteristics according to claim 5 is characterized in that: The Gumbel-Softmax algorithm is used to further weight the output of the gating network to achieve sparse activation of the expert sub-model: Initial logits: Assume that we have a logits vector z containing N elements, which represents the pre-activation score for each expert sub-model, that is, the weight vector of the expert sub-model of the output layer of the gating network model; Add Gumbel noise: Add a noise g sampled from a Gumbel (0, 1) distribution to each logits: z′=z+g Where g=−log(−log(u)), u is a vector sampled from the uniform distribution U(0,1), z is the logits before adding Gumbel noise, and z′ is the logits after adding Gumbel noise; Softmax mapping: Apply the Softmax function to the logits z′ with Gumbel noise added to obtain a probability distribution π: ; Among them, π i represents the i-th element in the probability distribution π, that is, the probability of the i-th category; z ’ i represents the i-th element of the logits vector after adding Gumbel noise; z ’ j represents the jth element of the logits vector after adding Gumbel noise; τ is the temperature parameter, which controls the sharpness of the softmax distribution; when τ approaches 0, the distribution becomes very sharp, close to a one-hot vector, and when τ approaches infinity, the distribution approaches a uniform distribution; The 3D vector output by the output layer of the gated network model is used as the input of the Gumbel-Softmax algorithm. The Gumbel-Softmax algorithm re-outputs a 3D vector as the weight of the expert sub-model by adding Gumbel noise and Softmax mapping. Then the expert sub-model weights are weighted and summed with the outputs of the corresponding expert sub-models to obtain a comprehensive feature representation; Afterwards, the comprehensive feature representation is connected to a fully connected layer; Finally, the battery SOC is output through a linear regression layer.

7. The battery SOC prediction method based on the SMoE model and expansion stress characteristics according to claim 1, characterized in that: Specifically, the step S3 is to implement the SMoE model code on the Keras platform and train it. During training, the number of training epochs=1000 and the number of batch samples batch_size=1024; the optimizer for model training is the adam algorithm, and the loss function is the mean square error; the training data is divided into a training set and a test set in a ratio of 8:2, the model is trained on the training set, and the model is verified on the test set.

Citation Information

Patent Citations

  • Method for predicting residual capacity of battery based on ELM (Extreme Learning Model) and battery impedance spectrum data

    CN115902664A

  • Precise model and lightweight model combined depression identification method and system

    CN118098289A

  • Method for estimating SOC (State of Charge) of lithium battery, battery management system, medium and electric energy equipment

    CN118191637A

  • SOH prediction method based on MOE model and current, voltage and temperature of battery

    CN118584357A

  • Method for predicting RUL based on MOE model and current resistance impedance of battery

    CN118625152A

Cited By

  • Model training method and prediction method for jointly predicting SOC, RUL and SOH of battery

    CN119936687A