Power battery pack charge state estimation method

Abstract

The invention discloses a method for estimating the state of charge of a power battery pack, and belongs to the field of battery management. The method comprises the steps of firstly collecting data such as current and voltage of a battery pack and performing normalization processing, then extracting local time sequence features through a convolutional neural network, and dynamically enhancing key features by using an attention mechanism. And then capturing a time sequence dependency relationship by adopting a bidirectional long-short-term memory network, outputting a state of charge estimation value through a full connection layer, and finally dynamically optimizing a prediction result in combination with Kalman filtering. The system comprises a data acquisition module, a preprocessing module, a hybrid neural network modeling module and a result optimization module. The advantages of the convolutional neural network, the bidirectional long-short-term memory network, the attention mechanism and the Kalman filtering are fused, the precision and robustness of charge state estimation are remarkably improved, and the method is suitable for application scenes such as electric vehicles and energy storage systems.

Description

Technology Neighborhood

[0001] The present invention relates to the field of power battery management technology, and in particular to a power battery pack state of charge (SOC) estimation method based on a hybrid neural network model, which is applicable to scenarios such as electric vehicles and energy storage systems. Background Art

[0002] With the rapid development of the global new energy vehicle industry, power batteries, as core energy storage components, have become crucial for battery management systems (BMS) in terms of accurate state of charge (SOC) estimation. As a key parameter measuring the remaining available energy in a battery, SOC directly impacts the formulation of battery charge and discharge strategies, fault diagnosis, and lifespan prediction. However, power batteries exhibit strong nonlinear characteristics during the charge and discharge process and are susceptible to multiple factors, including temperature, cycle life, and current fluctuations. This poses numerous challenges to SOC estimation, including insufficient accuracy and poor robustness.

[0003] Traditional SOC estimation methods, such as the open-circuit voltage method, rely on static voltage-SOC curves and are unable to adapt to complex dynamic operating conditions. The ampere-hour integration method suffers from the problem of initial value error accumulation, which increases the estimation error over time. Model-driven methods, such as the extended Kalman filter and particle filter, while theoretically capable of SOC estimation, require a precise battery equivalent circuit model, resulting in complex parameter identification and high computational cost. In recent years, data-driven deep learning methods have become a hot topic due to their powerful nonlinear fitting capabilities. For example, LSTM networks can capture long-term dependencies in time series, but their unidirectional structure limits their ability to utilize future information. CNNs can extract local signal features, and hybrid models combined with LSTMs (such as CNN-LSTM) have demonstrated some performance in battery state estimation. However, existing methods still suffer from insufficient attention to key features, leading to feature redundancy under complex operating conditions, a lack of dynamic optimization of prediction results, and difficulty in suppressing noise and sudden interference. Therefore, a new SOC estimation method is urgently needed to improve estimation accuracy and robustness. Summary of the Invention

[0004] Therefore, in order to solve the problems existing in the above-mentioned traditional SOC estimation method, the present invention proposes a power battery pack SOC estimation method based on a hybrid neural network model. The present invention adopts the following technical solutions:

[0005] In a first aspect, the present invention provides a method for estimating the SOC of a power battery pack based on a hybrid neural network model, comprising:

[0006] Data preprocessing: Collect the current, voltage and time data of the power battery pack and perform normalization. Through normalization, the data is mapped to a specific interval, eliminating the dimensional differences between the data, improving the model training effect and convergence speed, and making it easier for the model to learn data features. For example, using The function normalizes the data to the interval [0,1];

[0007] Feature extraction: The preprocessed data is fed into a CNN to extract local temporal features. In this method, the CNN consists of two one-dimensional convolutional layers and a ReLU activation function. The one-dimensional convolutional layers automatically extract local features from the voltage and current signals through convolution operations, such as voltage fluctuation patterns and current trends within a specific time period. The ReLU activation function enhances the network's nonlinear expression capabilities, enabling it to learn more complex feature relationships and effectively capture the spatial feature information in the power battery data.

[0008] Attention Mechanism: The attention mechanism dynamically assigns weights to extracted features, enhancing the weights of key features. The attention mechanism generates dynamic weights through global average pooling and fully connected layers, and then performs dot multiplication with the features output by the convolutional layer. The global average pooling layer compresses feature dimensions, while the fully connected layer learns feature weights. Dynamic weights are obtained through a Sigmoid activation function and then dot multiplied with the convolutional layer features to highlight key features and improve the stability and reliability of SOC estimation.

[0009] Time Series Modeling: The weighted features are fed into a bidirectional long short-term memory (Bi-LSTM) network to capture forward and backward dependencies in the time series. The Bi-LSTM consists of a forward LSTM layer and a backward LSTM layer. The forward LSTM layer processes the sequence in chronological order, while the backward LSTM layer processes it in reverse order. The outputs of the two layers are concatenated, allowing the model to consider both past and future information at each time step, accurately mining temporal dependencies in the data and improving prediction accuracy.

[0010] Preliminary prediction: Outputs a preliminary SOC estimate through the fully connected layer. The fully connected layer maps the features output by the Bi-LSTM to the final prediction space to obtain a preliminary SOC estimate.

[0011] Dynamic optimization: Kalman filtering is used to dynamically modify the initial prediction results, and the process noise covariance (Q) and measurement noise covariance (R) are adjusted through a sliding window to suppress noise interference and error accumulation. Specifically, it includes:

[0012] Prediction step: Based on the state estimate at the previous moment and the process noise covariance, calculate the prior estimate at the current moment. The formula is , ,in, is the prior estimated state, is the prior covariance matrix, A is the state transfer matrix, and Q is the process noise covariance;

[0013] Update step: Combine the initial prediction value and the measurement noise covariance, and correct the prior estimate through the Kalman gain to obtain the optimal SOC estimate. The formula is , , ,in, is the Kalman gain, which dynamically weighs the confidence of prediction and observation, is the observation matrix, is the measurement noise covariance;

[0014] Dynamic adjustment: Dynamically adjust the values ​​of Q and R based on the prediction error within the sliding window.

[0015] As when hour:

[0016] ;

[0017] ;

[0018] ,

[0019] Adaptively adjust the noise covariance accordingly to improve estimation accuracy;

[0020] Output result: Output the optimized SOC estimate, providing accurate SOC information for the battery management system, supporting functions such as charge and discharge strategy formulation, balancing control, and safety warning.

[0021] Under the dynamic stress test (DST) conditions, the SOC estimation mean absolute error (MAE) of this method is less than 0.6%, and the root mean square error (RMSE) is less than 0.65%, showing good estimation performance.

[0022] In a second aspect, the present invention provides a method for estimating the SOC of a power battery pack, comprising:

[0023] Data acquisition module: used to collect current, voltage and time data of the power battery pack in real time, providing raw data support for subsequent analysis;

[0024] Preprocessing module: normalizes the collected data to ensure that the data meets the model input requirements and improves the model training effect;

[0025] Hybrid Neural Network Module: This module includes a CNN, an attention mechanism, a Bi-LSTM, and a fully connected layer. The CNN extracts local features, the attention mechanism assigns weights, the Bi-LSTM models temporal relationships, and the fully connected layer outputs preliminary SOC predictions. These layers work together to extract deep features and provide preliminary predictions. This module is trained using the Adam optimization algorithm. By properly setting training parameters, the model converges to an optimal solution more quickly.

[0026] Kalman filter module: Dynamically optimizes the preliminary prediction value, uses the Kalman filter algorithm, and combines it with dynamically adjusted noise covariance to output the final SOC estimation value to improve estimation accuracy and stability;

[0027] Result display module: used to display SOC estimation results and error analysis, intuitively present model performance, facilitate users to understand the estimation situation, and assist in decision-making.

[0028] The principles and beneficial effects of the present invention are described below:

[0029] The present invention relates to a method for estimating the state of charge (SOC) of a power battery pack based on a hybrid neural network model. The method comprises: using a CNN to extract local features from signals such as the voltage and current of the power battery pack; combining an attention mechanism to weight the extracted features and enhance the influence of key features on SOC estimation; utilizing a Bi-LSTM to capture forward and backward dependencies in time series, further improving the temporal nature of feature expression; and finally, dynamically optimizing the initially predicted SOC value through a Kalman filter (KF) to suppress noise interference and error accumulation, ultimately outputting a high-precision SOC estimation result. Specifically, the method first preprocesses the collected battery data, then utilizes a CNN to extract local features, employs an attention mechanism to weight features, and employs a Bi-LSTM to capture temporal dependencies. Finally, dynamic optimization is achieved through a Kalman filter to achieve high-precision and robust SOC estimation of the power battery pack.

[0030] According to the above principles, it is not difficult to find the beneficial effects of the present invention:

[0031] 1. High-precision estimation: This method integrates CNN, Bi-LSTM, attention mechanism, and Kalman filter to fully leverage the strengths of each model. CNN extracts local features, Bi-LSTM captures temporal dependencies, attention mechanism focuses on key features, and Kalman filter dynamic optimization. Under DST conditions, the MAE is below 0.6% and the RMSE is below 0.65%, significantly improving estimation accuracy.

[0032] 2. Strong robustness: The attention mechanism enhances the weight of key features and reduces feature redundancy, and the Kalman filter suppresses noise and error accumulation, enabling the model to stably and accurately estimate SOC under complex working conditions and noisy environments, with good robustness.

[0033] 3. Wide adaptability: The method is applicable to different types of power battery packs, and parameters can be adjusted according to battery characteristics. It has broad application prospects in new energy vehicles, energy storage systems and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0035] Figure 1 Schematic diagram of the attention mechanism structure;

[0036] Figure 2 Schematic diagram of the bidirectional long short-term memory network structure;

[0037] Figure 3 This is the model framework flow chart;

[0038] Figure 4 This is a schematic diagram of the model framework connection;

[0039] Figure 5 shows the SOC estimation results and errors at 25°C under DST conditions: (a) SOC estimation value and SOC true value at 25°C; (b) SOC estimation error at 25°C. DETAILED DESCRIPTION

[0040] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In a practical application scenario, an electric vehicle is taken as an example to illustrate how the present invention can achieve accurate estimation of the SOC of a power battery pack.

[0041] 1. Data acquisition and preprocessing: When an electric vehicle is driving, the data acquisition module uses high-precision sensors to collect real-time current, voltage, and time data from the power battery pack. Suppose that during an experimental driving, a series of data is collected, such as the current varies between -200A (charging) and 150A (discharging) and the voltage fluctuates between 300-400V within a certain period of time. In order to eliminate the dimensional differences between the data and improve the model training effect, the normalization formula is used. , map the current and voltage data to the [0,1] interval. Here is the original data, and These are the minimum and maximum values ​​in the collected data, respectively. In this way, the data meets the model input requirements, making it easier for the subsequent model to learn the data features.

[0042] 2. Feature extraction: Input the preprocessed data into CNN. Figure 3 The two one-dimensional convolutional layers in the CNN automatically extract local temporal features from the voltage and current signals. For example, in a data segment spanning 50 time steps, the voltage exhibits a fluctuation pattern, first decreasing and then increasing. The one-dimensional convolutional layer captures this specific period of fluctuation. Furthermore, the ReLU activation function enhances the network's nonlinear representation capabilities, enabling it to learn more complex feature relationships and effectively mine spatial feature information from power battery data, such as the potential connection between current changes and voltage fluctuations.

[0043] 3. Attention mechanism: Use the attention mechanism to dynamically assign weights to the extracted features. Figure 1 As shown in the figure, the attention mechanism compresses feature dimensions through a global average pooling layer, then learns feature weights through a fully connected layer, and obtains dynamic weights through a sigmoid activation function. This dynamic weight is then fused with the features output by the convolutional layer through a dot product. For example, during the rapid discharge phase of a battery, current variation is more critical for SOC estimation. The attention mechanism automatically increases the weight of these features, highlighting their role in SOC estimation and improving the stability and reliability of the estimation.

[0044] 4. Time series modeling: Input the weighted features into Bi-LSTM. Figure 2 As shown in the Bi-LSTM structure diagram, it consists of a forward LSTM layer and a backward LSTM layer. The forward LSTM layer processes the sequence in chronological order, while the backward LSTM layer processes it in reverse order, and the outputs of the two are concatenated. When processing the time series of battery data during electric vehicle driving, the Bi-LSTM model allows the model to simultaneously consider both past and future information at each time step. For example, when predicting the current state of charge (SOC), it considers both previous charge and discharge history data and information from subsequent data, accurately mining the temporal dependencies in the data and thereby improving prediction accuracy.

[0045] 5. Preliminary prediction and dynamic optimization: The features output by Bi-LSTM are mapped to the final prediction space through the fully connected layer to obtain the preliminary SOC estimation value. Then the Kalman filter is used to dynamically correct the preliminary prediction results. In the prediction step, based on the formula (Here we assume ,Right now ) calculates the prior estimate of the current moment, where is the prior estimated state, A is the state transfer matrix, is the state estimate of the previous moment. In the update step, according to the formula Calculate Kalman gain , and then through the formula Get the optimal SOC estimate, here is the prior covariance matrix, is the observation matrix, is the measurement noise covariance, is the initial prediction value. Furthermore, Q and R are dynamically adjusted based on the prediction error within the sliding window. For example, when the prediction error is large, the values ​​of Q and R are appropriately increased to enhance the ability to suppress noise and error and improve estimation accuracy.

[0046] 6. Results Output and Application: Under DST conditions and 25°C, Figure 5 shows that (a) both the true and estimated SOC decrease over time, with generally consistent trends, though some deviations exist. The discrepancies are clearly visible in the zoomed-in area. (b) The error between the estimated SOC and the true SOC fluctuates between -1% and 1.5%. Based on these experimental results, an optimized SOC estimate is ultimately output, providing accurate SOC information for the electric vehicle's battery management system. In practical applications, this information can be used to formulate appropriate charging and discharging strategies: when the estimated SOC is low, the system reminds the driver to charge immediately. During charging, if the SOC approaches full capacity, the charging current is adjusted to prevent overcharging. Furthermore, this information can be used for cell balancing control to ensure SOC consistency among individual cells in the battery pack, extending battery life. It can also be used for safety warnings, issuing alarms when SOC is abnormal, ensuring the safe operation of the electric vehicle.

[0047] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for estimating the state of charge (SOC) of a power battery pack based on a hybrid neural network model, characterized in that: A framework combining a neural network consisting of a convolutional neural network (CNN), an attention mechanism, and a bidirectional long short-term memory network (Bi-LSTM) with a Kalman filter (KF) is adopted. The neural network is used to extract data features, perform time series modeling, and output preliminary SOC estimates. The Kalman filter is used to dynamically correct the preliminary estimates to suppress noise interference and error accumulation.

2. The power battery pack SOC estimation method based on the hybrid neural network model according to claim 1 is characterized in that: The specific steps include: Data preprocessing: Collect current, voltage and time data of the power battery pack and perform normalization processing; Feature extraction: input the preprocessed data into CNN to extract local temporal features; Attention mechanism: Dynamically assign weights to the extracted features through the attention mechanism to enhance the weights of key features; Time series modeling: The weighted features are input into Bi-LSTM to capture the forward and backward dependencies of the time series; Preliminary prediction: Outputs a preliminary SOC estimate through the fully connected layer; Dynamic optimization: Kalman filtering is used to dynamically modify the initial prediction results. The process noise covariance (Q) and measurement noise covariance (R) are adjusted through a sliding window to suppress noise interference and error accumulation. Output result: Output the optimized SOC estimation value.

3. The method according to claim 2, characterized in that In step 2, the CNN includes two one-dimensional convolutional layers and a ReLU activation function, which are used to extract local features of voltage and current signals.

4. The method according to claim 2, characterized in that In step 3, the attention mechanism generates dynamic weights through the global average pooling layer and the fully connected layer, and performs point multiplication fusion with the features output by the convolutional layer.

5. The method according to claim 2, characterized in that In step 4, the Bi-LSTM is composed of a forward LSTM layer and a backward LSTM layer, which is used to simultaneously capture past and future time series information.

6. The method according to claim 2, characterized in that In step 6, the dynamic optimization of the Kalman filter includes the following sub-steps: Prediction step: Based on the state estimate and Q at the previous moment, calculate the prior estimate of the current moment; Update step: Combine the initial prediction value and R, and correct the prior estimate through the Kalman gain to obtain the optimal SOC estimate; Dynamic adjustment: Dynamically adjust the values ​​of Q and R based on the prediction error within the sliding window.

7. The method according to claim 2, characterized in that The proposed method outperforms traditional methods in terms of mean absolute error and root mean square error of SOC estimation under dynamic stress test (DST) conditions.

8. A system for realizing SOC estimation of a power battery pack, characterized in that: include: Data acquisition module: used to collect current, voltage and time data of power battery pack in real time; Preprocessing module: used to normalize the collected data; Hybrid neural network module: including CNN, attention mechanism, Bi-LSTM and fully connected layers, used to extract features, assign weights, model temporal relationships and output preliminary SOC prediction values; Kalman filter module: used to dynamically optimize the preliminary prediction value and output the final SOC estimation value; Result display module: used to display SOC estimation results and error analysis.

9. The method according to claim 8, characterized in that The hybrid neural network module is trained using the Adam optimization algorithm.

Images