Battery SOC estimation method based on xLSTM model

Abstract

The invention relates to the technical field of lithium batteries, in particular to a battery SOC estimation method based on an xLSTM model. A battery SOC estimation method based on an xLSTM model comprises a data acquisition module, a data preprocessing module, an xLSTM-based voltage prediction module, a single voltage prediction module and an SOC prediction module. Compared with the prior art, the xLSTM model-based battery SOC estimation method provided by the invention has the advantages that the xLSTM model is introduced, so that a long dependency relationship in time sequence data can be better captured, and the precision of time sequence prediction tasks such as the battery SOC and the like is improved; by combining the steps of data preprocessing, model construction, training, verification and the like, the xLSTM can realize relatively high prediction performance in practical application; and an Adaboost algorithm and Kalman filtering of integrated learning are combined to realize accurate prediction of the SOC of the battery. The method enhances the robustness and prediction precision of the system, and is suitable for the practical application of the battery management system.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a battery SOC estimation method based on an xLSTM model. Background Art

[0002] SOC is usually defined as the amount of available charge remaining in the battery, and is an important indicator for evaluating the remaining capacity of the battery and predicting the battery life. In the fields of electric vehicles, energy storage systems, smart grids, and consumer electronics, accurate SOC estimation is crucial for optimizing energy management, extending battery life, and improving user experience.

[0003] The ampere-hour integration method estimates the SOC by accumulating the charge and discharge of the battery. However, this method will produce cumulative errors, is highly dependent on the initial value of SOC, and cannot solve the problem of battery self-discharge.

[0004] The open circuit voltage method is another commonly used SOC estimation method, which estimates the SOC by measuring the open circuit voltage (OCV) of the battery. This method requires a long period of static state, is affected by the hysteresis voltage, and has high requirements for the accuracy of the SOC-OCV curve. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides a battery SOC estimation method based on the xLSTM model. By introducing the xLSTM model, the long dependencies in time series data can be better captured, and the accuracy of time series prediction tasks such as battery SOC can be improved.

[0006] To achieve the above purpose, a battery SOC estimation method based on the xLSTM model is designed, including a data acquisition module, a data preprocessing module, a voltage prediction module based on xLSTM, a single cell voltage prediction module, and a SOC prediction module, which is characterized by:

[0007] (1) Data acquisition module: including a sensor interface, a data acquisition unit, and a data storage unit. Each sensor is connected to the data acquisition unit through the sensor interface, and the data acquisition unit transmits the data of each sensor to the data storage unit through the communication interface;

[0008] (2) Data preprocessing module: including a data cleaning unit, a data normalization unit, and a feature extraction unit. The sensor data are transmitted from the data acquisition module to the data cleaning unit, and the data is cleaned by the data cleaning unit. The cleaned data is transmitted to the data normalization unit, and the data is normalized by the data normalization unit. The processed data is transmitted to the feature extraction unit.

[0009] (3) Voltage prediction module based on xLSTM: including a model selection unit, a model training unit, a model verification unit, and a model prediction unit. The data processed by the data preprocessing module is transmitted to the model training unit. The data model is trained by the model training unit. The trained model is transmitted to the model verification unit. The parameters are adjusted by the model verification unit. The model after parameter adjustment is transmitted to the model prediction unit for real-time voltage prediction.

[0010] (4) Single-cell voltage prediction module: This is a single-cell voltage prediction module with adaptive weight adjustment based on the Adaboost algorithm of ensemble learning; it includes an initialization weight unit, a training base learner, a weight adjustment unit, and an iterative training unit. The data processed by the data preprocessing module is transmitted to the initialization weight unit, the weight of the base learner is initialized by the initialization weight unit, the data after the weight of the base learner is initialized is transmitted to the training base learner, the training is performed by the training base learner, the training result is transmitted to the weight adjustment unit, the weight is updated by the weight adjustment unit, and the updated weight is returned to the training base learner for the next iterative training until the preset error range is reached;

[0011] (5) SOC prediction module: It includes a Kalman filter unit and a SOC estimation unit. The voltage prediction data obtained by the xLSTM-based voltage prediction module and the cell voltage prediction module are transmitted to the Kalman filter unit, optimized by the Kalman filter unit, and the optimized data are transmitted to the SOC estimation unit, which estimates the battery SOC.

[0012] The sensors include voltage sensors, current sensors, and temperature sensors.

[0013] The data normalization processing in the data preprocessing module is to process the collected voltage, current and temperature data through the normalization formula, and the normalization formula is: Among them, x is the original data, x' is the normalized data, min(x) is the minimum value in the data set, and max(x) is the maximum value in the data set.

[0014] In the xLSTM-based voltage prediction module, xLSTM is an extended version of the LSTM model. During the training process, the xLSTM input gate formula, xLSTM forget gate formula, xLSTM output gate formula, xLSTM unit state update formula, and xLSTM hidden state update formula are used; the xLSTM input gate formula is: i t =σ(W i ·[h t-1 , x t , C t-1 ]+bi ); xLSTM forget gate formula: f t =σ(W f ·[h t-1 , x t , C t-1 ]+b f ); xLSTM output gate formula: t =σ(W O ·[h t-1 , x t , C t ]+b O ); xLSTM unit state update formula: xLSTM hidden state update formula: h t =O t tanh(C t ), where i t is the input gate activation value, f t is the activation value of the forget gate, O t is the output gate activation value, is the new candidate memory cell state, C t is the current memory cell state, h t is the current hidden state, W i , W f , W O , W C is the weight matrix, b i , b f , b O , b C is the bias, σ is the Sigmoid activation function, and tanh is the Tanh activation function.

[0015] In the LSTM model, the LSTM input gate formula, LSTM forget gate formula, LSTM output gate formula, LSTM unit state update formula, and LSTM hidden state update formula are used during the training process; the LSTM input gate formula: i t =σ(W i ·[h t-1 , x t ]+b i ); LSTM forget gate formula: f t =σ(W f ·[h t-1 , x t ]+b f );LSTM output gate formula: O t =σ(W O ·[h t-1 , x t ]+b O ); LSTM unit state update formula: LSTM hidden state update formula: h t =O t tanh(C t ), where i t is the input gate activation value, f t is the activation value of the forget gate, O t is the output gate activation value, is the new candidate memory cell state, C t is the current memory cell state, h t is the current hidden state, W i , W f , W O , W C is the weight matrix, b i , b f , b O , b C is the bias, σ is the Sigmoid activation function, and tanh is the Tanh activation function.

[0016] In the monomer voltage prediction module, after initializing the weight, the prediction ability of the training base learner is optimized through multiple iterative training and weight adjustment. The initialization weight formula, error rate calculation formula, weight update formula, and weight normalization formula are used in the process. The initialization weight formula is: Error rate calculation formula: Weight normalization formula: Among them, w i is the weight of the sample; N is the total number of samples; ∈ is the error rate of the classifier; II is an exponential function, which is 1 when the prediction is wrong and 0 otherwise; α is the learning rate.

[0017] In the SOC prediction module, the battery health status is calculated using the SOH estimation formula through real-time monitoring of the battery capacity, providing reference information for the remaining battery life. The SOH estimation formula is: Among them, SOH(t) is the health state of the battery at time t, the current capacity is the actual capacity of the battery at the current moment, and the rated capacity is the initial rated capacity of the battery.

[0018] In the SOC prediction module, Kalman filtering is used to optimize the prediction of the battery voltage, and the battery state estimation is continuously corrected and optimized through the prediction step formula and the update step formula. The prediction step formula is: P k|k-1 =AP k-1|k-1 A T +Q, update step formula: K k =P k|k-1 H T(HP k|k-1 H T +R) -1 , P k|k =(IK k H)P k|k-1 ,in, is the predicted state at time k, A is the state transfer matrix, B is the control matrix, u k is the control input, P k|k-1 is the prediction covariance, Q is the process noise covariance, K k is the Kalman gain, H is the measurement matrix, R is the observation noise covariance, z k is the observed value.

[0019] Compared with the prior art, the present invention provides a battery SOC estimation method based on the xLSTM model. By introducing the xLSTM model, the long dependency relationship in time series data can be better captured, and the accuracy of time series prediction tasks such as battery SOC can be improved; combined with data preprocessing, model construction, training and verification steps, xLSTM can achieve higher prediction performance in practical applications; and combined with the Adaboost algorithm and Kalman filtering of integrated learning, accurate prediction of battery SOC can be achieved; the robustness and prediction accuracy of the system are enhanced, and it is suitable for practical applications of battery management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below with reference to the accompanying drawings.

[0022] like Figure 1 As shown, a battery SOC estimation method based on the xLSTM model includes a data acquisition module, a data preprocessing module, a voltage prediction module based on xLSTM, a single cell voltage prediction module, and a SOC prediction module.

[0023] (1) Data acquisition module: It includes a sensor interface, a data acquisition unit, and a data storage unit. Each sensor is connected to the data acquisition unit through the sensor interface, and the data acquisition unit transmits the sensor data to the data storage unit through the communication interface.

[0024] (2) Data preprocessing module: including data cleaning unit, data normalization unit and feature extraction unit. The sensor data are transmitted from the data acquisition module to the data cleaning unit, and then the data are cleaned by the data cleaning unit. The cleaned data are transmitted to the data normalization unit, and then the data are normalized by the data normalization unit. The processed data are transmitted to the feature extraction unit.

[0025] (3) Voltage prediction module based on xLSTM: including model selection unit, model training unit, model verification unit and model prediction unit. The data processed by the data preprocessing module is transmitted to the model training unit. The data model is trained by the model training unit. The trained model is transmitted to the model verification unit. The parameters are adjusted by the model verification unit. The model after parameter adjustment is transmitted to the model prediction unit for real-time voltage prediction.

[0026] (4) Single-cell voltage prediction module: This is a single-cell voltage prediction module with adaptive weight adjustment based on the Adaboost algorithm of ensemble learning; it includes an initialization weight unit, a training basic learner, a weight adjustment unit, and an iterative training unit. The data processed by the data preprocessing module is transmitted to the initialization weight unit, the weight of the basic learner is initialized by the initialization weight unit, the data after the weight of the basic learner is initialized is transmitted to the training basic learner, the training is performed by the training basic learner, the training result is transmitted to the weight adjustment unit, the weight is updated by the weight adjustment unit, and the updated weight is returned to the training basic learner for the next iterative training until the preset error range is reached.

[0027] (5) SOC prediction module: It includes a Kalman filter unit and a SOC estimation unit. The voltage prediction data obtained by the xLSTM-based voltage prediction module and the cell voltage prediction module are transmitted to the Kalman filter unit, optimized by the Kalman filter unit, and the optimized data are transmitted to the SOC estimation unit, which estimates the battery SOC.

[0028] The sensors include voltage sensors, current sensors, and temperature sensors.

[0029] The data normalization processing in the data preprocessing module is to process the collected voltage, current, and temperature data through the normalization formula. The normalization formula is: Among them, x is the original data, x' is the normalized data, min(x) is the minimum value in the data set, and max(x) is the maximum value in the data set.

[0030] In the voltage prediction module based on xLSTM, xLSTM is an extended version of the LSTM model. During the training process, the xLSTM input gate formula, xLSTM forget gate formula, xLSTM output gate formula, xLSTM unit state update formula, and xLSTM hidden state update formula are used; the xLSTM input gate formula is: i t =σ(W i ·[h t-1 , x t , Ct-1 ]+b i ); xLSTM forget gate formula: f t =σ(W f ·[h t-1 , x t , C t-1 ]+b f ); xLSTM output gate formula: O t =σ(W O ·[h t-1 , x t , C t ]+b O ); xLSTM unit state update formula: xLSTM hidden state update formula: h t =O t tanh(C t ), where i t is the input gate activation value, f t is the activation value of the forget gate, O t is the output gate activation value, is the new candidate memory cell state, C t is the current memory cell state, h t is the current hidden state, W i , W f , W O , W C is the weight matrix, b i 、b f 、b O 、b C is the bias, σ is the Sigmoid activation function, and tanh is the Tanh activation function.

[0031] In the LSTM model, the LSTM input gate formula, LSTM forget gate formula, LSTM output gate formula, LSTM unit state update formula, and LSTM hidden state update formula are used during the training process; the LSTM input gate formula is: t =σ(W i ·[h t-1 , x t ]+b i ); LSTM forget gate formula: f t =σ(W f ·[h t-1 , x t ]+b f );LSTM output gate formula: O t =σ(W O ·[h t-1 , x t ]+b O); LSTM unit state update formula: LSTM hidden state update formula: h t =O t tanh(C t ), where i t is the input gate activation value, f t is the activation value of the forget gate, O t is the output gate activation value, is the new candidate memory cell state, C t is the current memory cell state, h t is the current hidden state, W i , W f , W O , W C is the weight matrix, b i 、b f 、b O 、b C is the bias, σ is the Sigmoid activation function, and tanh is the Tanh activation function.

[0032] In the single cell voltage prediction module, after initializing the weights, the prediction ability of the training base learner is optimized through multiple iterative training and weight adjustment. The initialization weight formula, error rate calculation formula, weight update formula, and weight normalization formula are used in the process. The initialization weight formula is: Error rate calculation formula: Weight normalization formula: Among them, w i is the weight of the sample; N is the total number of samples; ∈ is the error rate of the classifier; II is an exponential function, which is 1 when the prediction is wrong and 0 otherwise; α is the learning rate.

[0033] In the SOC prediction module, the battery health status is calculated using the SOH estimation formula through real-time monitoring of the battery capacity, providing reference information for the remaining battery life. The SOH estimation formula is: Among them, SOH(t) is the health state of the battery at time t, the current capacity is the actual capacity of the battery at the current moment, and the rated capacity is the initial rated capacity of the battery.

[0034] In the SOC prediction module, Kalman filtering is used to optimize the prediction of battery voltage, and the battery state estimation is continuously corrected and optimized through the prediction step formula and the update step formula. The prediction step formula is: P k|k-1 =AP k-1|k-1 A T +Q, update step formula: Kk =P k|k-1 H T (HP k|k-1 H T +R) -1 , P k|k =(IK k H)P k|k-1 ,in, is the predicted state at time k, A is the state transfer matrix, B is the control matrix, u k is the control input, P k|k-1 is the prediction covariance, Q is the process noise covariance, K k is the Kalman gain, H is the measurement matrix, R is the observation noise covariance, z k is the observed value.

[0035] Implementation Cases:

[0036] 1. Data acquisition module:

[0037] (1) Initialize sensors: When the system starts, initialize the voltage sensor, current sensor, and temperature sensor installed on the battery;

[0038] (2) Collect sensor data: regularly collect voltage, current and temperature data from sensors and record the number of battery discharge and charge cycles;

[0039] (3) Data transmission: Data is transmitted to the central control unit via Ethernet;

[0040] (4) Data storage: Store data in a local database in real time and back it up to a remote server regularly.

[0041] 2. Data preprocessing module:

[0042] (1) Data cleaning: Apply a low-pass filter to remove data noise, and detect and process outliers by setting reasonable thresholds;

[0043] (2) Data normalization: Normalize the voltage, current, and temperature data to make them suitable for model input; (3) Extract features such as voltage change rate and temperature change rate from the normalized data.

[0044] 3. Voltage prediction module based on xLSTM:

[0045] (1) Model selection: Select the xLSTM model as the time series prediction model;

[0046] (2) Model training: Use historical data to train the xLSTM model and adjust hyperparameters to improve model performance;

[0047] (3) Model validation: Use the validation set to validate and adjust the model to improve the generalization ability of the model;

[0048] (4) Model prediction: Input new feature data and use the trained xLSTM model to predict the future battery voltage.

[0049] 4. Monomer voltage prediction module with adaptive weight adjustment based on ensemble learning Adaboost algorithm:

[0050] (1) Initialization weights: assign initial weights to the base learner;

[0051] (2) Training the base learner: Use the weighted training data to train the base learner and calculate the prediction error;

[0052] (3) Adjust weights: Adjust the weights of the basic learner according to the prediction error and perform multiple iterative training until the preset error range is reached.

[0053] 5. SOC prediction module:

[0054] (1) Kalman filtering: Use Kalman filtering to optimize the prediction of battery voltage and correct the battery state estimation;

[0055] (2) SOC estimation: Based on the predicted voltage and actual current data, the Coulomb counting method is used to estimate the battery SOC and calculate the battery state of health (SOH).

[0056] After a period of actual operation and testing, the performance of the renewable energy company's energy storage system has been significantly improved. The specific results are as follows:

[0057] (1) Voltage prediction error: reduced by about 20%, making battery voltage prediction more accurate;

[0058] (2) SOC estimation accuracy: The SOC estimation error is reduced by about 12%, which improves the reliability of battery remaining power estimation;

[0059] (3) Battery management efficiency: Through more accurate SOH calculation, the battery charge and discharge management strategy is optimized, improving the overall battery efficiency and life.

Claims

1. A battery SOC estimation method based on an xLSTM model, comprising a data acquisition module, a data preprocessing module, an xLSTM-based voltage prediction module, a cell voltage prediction module, and an SOC prediction module, characterized in that: (1) Data acquisition module: including a sensor interface, a data acquisition unit, and a data storage unit. Each sensor is connected to the data acquisition unit through the sensor interface, and the data acquisition unit transmits the data of each sensor to the data storage unit through the communication interface; (2) Data preprocessing module: including a data cleaning unit, a data normalization unit, and a feature extraction unit. The sensor data are transmitted from the data acquisition module to the data cleaning unit, and the data is cleaned by the data cleaning unit. The cleaned data is transmitted to the data normalization unit, and the data is normalized by the data normalization unit. The processed data is transmitted to the feature extraction unit. (3) Voltage prediction module based on xLSTM: including a model selection unit, a model training unit, a model verification unit, and a model prediction unit. The data processed by the data preprocessing module is transmitted to the model training unit. The data model is trained by the model training unit. The trained model is transmitted to the model verification unit. The parameters are adjusted by the model verification unit. The model after parameter adjustment is transmitted to the model prediction unit for real-time voltage prediction. (4) Single-cell voltage prediction module: This is a single-cell voltage prediction module with adaptive weight adjustment based on the Adaboost algorithm of ensemble learning; it includes an initialization weight unit, a training base learner, a weight adjustment unit, and an iterative training unit. The data processed by the data preprocessing module is transmitted to the initialization weight unit, the weight of the base learner is initialized by the initialization weight unit, the data after the weight of the base learner is initialized is transmitted to the training base learner, the training is performed by the training base learner, the training result is transmitted to the weight adjustment unit, the weight is updated by the weight adjustment unit, and the updated weight is returned to the training base learner for the next iterative training until the preset error range is reached; (5) SOC prediction module: It includes a Kalman filter unit and a SOC estimation unit. The voltage prediction data obtained by the xLSTM-based voltage prediction module and the cell voltage prediction module are transmitted to the Kalman filter unit, optimized by the Kalman filter unit, and the optimized data are transmitted to the SOC estimation unit, which estimates the battery SOC.

2. A battery SOC estimation method based on the xLSTM model according to claim 1, characterized in that: The sensors include voltage sensors, current sensors, and temperature sensors.

3. A battery SOC estimation method based on the xLSTM model according to claim 1, characterized in that: The data normalization processing in the data preprocessing module is to process the collected voltage, current and temperature data through the normalization formula, and the normalization formula is: Among them, x is the original data, x′ is the normalized data, min(x) is the minimum value in the data set, and max(x) is the maximum value in the data set.

4. The battery SOC estimation method based on the xLSTM model according to claim 1, characterized in that: In the xLSTM-based voltage prediction module, xLSTM is an extended version of the LSTM model. During the training process, the xLSTM input gate formula, xLSTM forget gate formula, xLSTM output gate formula, xLSTM unit state update formula, and xLSTM hidden state update formula are used; the xLSTM input gate formula is: i t =σ(W i ·[h t-1 , x t , C t-1 ]+b i ); xLSTM forget gate formula: f t =σ(W f ·[h t-1 , x t , C t-1 ]+b f ); xLSTM output gate formula: t =σ(W O ·[h t-1 , x t , C t ]+b O ); xLSTM unit state update formula: xLSTM hidden state update formula: h t =O t tanh(C t ), where i t is the input gate activation value, f t is the activation value of the forget gate, O t is the output gate activation value, is the new candidate memory cell state, C t is the current memory cell state, h t is the current hidden state, W i , W f , W O , W C is the weight matrix, b i 、b f 、b O 、b C is the bias, σ is the Sigmoid activation function, and tanh is the Tanh activation function.

5. A battery SOC estimation method based on the xLSTM model according to claim 4, characterized in that: In the LSTM model, the LSTM input gate formula, LSTM forget gate formula, LSTM output gate formula, LSTM unit state update formula, and LSTM hidden state update formula are used during the training process; the LSTM input gate formula: ix = σ(W i ·[h t-1 , x t ]+b i ); LSTM forget gate formula: f t =σ(W f ·[h t-1 , x t ]+b f );LSTM output gate formula: O t =σ(W O ·[h t-1 , x t ]+b O ); LSTM unit state update formula: LSTM hidden state update formula: h t =O t tanh(C t ), where i t is the input gate activation value, f t is the activation value of the forget gate, O t is the output gate activation value, is the new candidate memory cell state, C t is the current memory cell state, h t is the current hidden state, W i , W f , W O , W C is the weight matrix, b i , b f , b O , b C is the bias, σ is the Sigmoid activation function, and tanh is the Tanh activation function.

6. The battery SOC estimation method based on the xLSTM model according to claim 1, characterized in that: In the monomer voltage prediction module, after initializing the weight, the prediction ability of the training base learner is optimized through multiple iterative training and weight adjustment. The initialization weight formula, error rate calculation formula, weight update formula, and weight normalization formula are used in the process. The initialization weight formula is: Error rate calculation formula: w i ←w i ·exp(α·II(y i ≠h(x i ))), Weight normalization formula: Among them, w i is the weight of the sample; N is the total number of samples; ∈ is the error rate of the classifier; II is an exponential function, which is 1 when the prediction is wrong and 0 otherwise; α is the learning rate.

7. The battery SOC estimation method based on the xLSTM model according to claim 1, characterized in that: In the SOC prediction module, the battery health status is calculated using the SOH estimation formula through real-time monitoring of the battery capacity, providing reference information for the remaining battery life. The SOH estimation formula is: Among them, SOH(t) is the health state of the battery at time t, the current capacity is the actual capacity of the battery at the current moment, and the rated capacity is the initial rated capacity of the battery.

8. The battery SOC estimation method based on the xLSTM model according to claim 1, characterized in that: In the SOC prediction module, Kalman filtering is used to optimize the prediction of the battery voltage, and the battery state estimation is continuously corrected and optimized through the prediction step formula and the update step formula. The prediction step formula is: P k|k-1 =AP k-1|k-1 A T +Q, update step formula: K k =P k|k-1 H T (HP k|k-1 H T +R) -1 , P k|k =(IK k H)P k|k-1 ,in, is the predicted state at time k, A is the state transfer matrix, B is the control matrix, u k is the control input, P k|k-1 is the prediction covariance, Q is the process noise covariance, K k is the Kalman gain, H is the measurement matrix, R is the observation noise covariance, z k is the observed value.

Images