Energy storage battery SOC estimation method and system and electronic device

Abstract

The invention provides an energy storage battery SOC estimation method and system and an electronic device, and the method comprises the steps: obtaining the voltage, current and temperature data of an energy storage battery, and carrying out the processing of the data; performing feature data extraction on the processed data by using a two-dimensional convolutional neural network to obtain first feature data; introducing an attention mechanism to enable a neural network to pay attention to important information in the first feature data to obtain second feature data; and inputting the second feature data into the QRLSTM for regression prediction, thereby realizing SOC point and interval prediction of the energy storage battery. The system is used for realizing the SOC estimation method. The electronic device includes a memory, a processor, and a computing program stored in the memory and executable on the processor. According to the method, the prediction precision is improved, a more comprehensive and deep data analysis tool is provided for an energy storage battery management system, and the safety and efficiency of energy storage battery use are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery state prediction technology, specifically to an energy storage battery SOC estimation method, system, and electronic device. Background Technology

[0002] New energy storage is a crucial technology and fundamental equipment for building new power systems. Currently, countries worldwide are striving towards this goal, with clean energy sources, represented by energy storage batteries, gaining popularity. Energy storage batteries, due to their high energy density, long lifespan, low self-discharge rate, and environmental friendliness, are widely used in small communication devices, new energy vehicles, and new energy storage power stations. Electricity storage is inseparable from energy storage batteries, and a Battery Management System (BMS) is indispensable in the use of energy storage batteries. The BMS can monitor battery parameters such as voltage, current, temperature, and SOC in real time, managing and controlling the charging and discharging process of the energy storage battery to prevent overcharging, over-discharging, and thermal runaway, ensuring the long-term safe and stable operation of the energy storage battery. SOC estimation is a crucial component of the BMS, affecting the performance and lifespan of the energy storage battery. High-precision SOC estimation contributes to the development of electricity storage and management. Therefore, this paper proposes a method, system, and electronic device for more accurately predicting and solving the problem of high-precision estimation of energy storage batteries and the challenges of constructing energy storage equipment. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the purpose of this invention is to solve one or more problems existing in the prior art. For example, one objective of this invention is to solve the problem of high-precision estimation of energy storage batteries and the difficulty of constructing energy storage battery equipment, while generating the uncertainty range of energy storage battery prediction.

[0004] To achieve the above objectives, the present invention provides a method for estimating the State of Charge (SOC) of an energy storage battery. The method may include the following steps: acquiring voltage, current, and temperature data of the energy storage battery and processing the data; using a two-dimensional convolutional neural network to extract feature data from the processed data to obtain first feature data; introducing an attention mechanism to enable the neural network to automatically learn and selectively focus on important information in the first feature data to obtain second feature data information; and inputting the second feature data into a QRLSTM for regression prediction to achieve prediction of the SOC point and range of the energy storage battery.

[0005] According to one or more exemplary embodiments of the present invention, the processing may include a normalization process, wherein the normalization process employs a formula that may include the following formula 1:

[0006] Formula 1:

[0007] Where x represents the original data and y represents the normalized data, x min x is the minimum value of the original data. ma y represents the maximum value of the original data. min y is the minimum value of the normalized data. max This represents the maximum value of the data after normalization.

[0008] According to one or more exemplary embodiments of one aspect of the present invention, before the feature data extraction is performed on the processed data using a two-dimensional convolutional neural network, the input time series data can be folded into two-dimensional data so that it can be input into the convolutional layer for processing.

[0009] According to one or more exemplary embodiments of one aspect of the present invention, the feature data extraction of the processed data using a two-dimensional convolutional neural network may include: using a two-dimensional convolutional neural network to extract features from the voltage, current and temperature data of the energy storage battery, retaining the most important information, while using the same weights on the same convolutional kernels throughout the input data.

[0010] According to one or more exemplary embodiments of one aspect of the present invention, the formula used for the two-dimensional convolution may include the following formula 2:

[0011] Equation 2: G[i,j]=∑ m ∑ n F[m,n]H[im,jn]

[0012] Where G[i,j] is an element in the output feature map; F[m,n] is the input data or the feature map of the previous layer; H[i,j] is the convolution kernel; and m,n are the dimension indices of the convolution kernel.

[0013] According to one or more exemplary embodiments of the present invention, the two-dimensional convolutional neural network may employ max pooling and the ReLU activation function to discard linearly uncorrelated features, wherein the formula used for max pooling and the ReLU activation function may respectively include the following formula 3 and the following formula 4:

[0014] Formula 3:

[0015] Equation 4: f(x) = max(0,x)

[0016] Among them, S l (j) is the output of the j-th pooling region in the l-th layer; w represents the width of the pooling region; X l-1 f(t) is the pooling region; x is the input; f(x) is the output.

[0017] According to one or more exemplary embodiments of one aspect of the present invention, the expression of the quantile loss function may include the following equation 5:

[0018] Formula 5:

[0019] Where β(τ) is the regression coefficient vector at the τ quantile, τ∈(0,1); X is the explanatory variable, Y is the response variable; Q Y (τ|X) represents the response variable Y with respect to the explanatory variables; ρ τ (u) is the sloping absolute value function.

[0020] According to one or more exemplary embodiments of one aspect of the present invention, the p τ The expression for (u) can include the following formula 6:

[0021] Formula 6:

[0022] Where, ρ τ (u) is the sloping absolute value function.

[0023] According to one or more exemplary embodiments of one aspect of the present invention, the expression of the LSTM may include the following equations 7, 8, 9, 10, 11 and 12:

[0024] Formula 7:

[0025] Formula 8:

[0026] Equation 9: H t =O t ⊙tanh(C t )

[0027] Equation 10: I t =σ(x t W xi +H t-1 W hi +b i )

[0028] Equation 11: F t =σ(x t W xf +H t-1 W hf +b f )

[0029] Formula 12: O t =σ(x t W xo +H t-1 W ho +b o )

[0030] in, For memory units; Ct Memory for the input of the next time step; H t For the hidden state of the next time step; I t For input gate; F t Forgotten Gate; O t For output gate; W xi W xf W xo W hi W hf and W ho Indicates the weight parameter; b i b f and b o H represents the bias parameter. t-1 σ is the hidden state of the previous time step; σ is the sigmoid activation function; ⊙ is the Hadamard product operator.

[0031] According to one or more exemplary embodiments of one aspect of the present invention, the expression for introducing the attention mechanism may include the following equations 13, 14 and 15:

[0032] Equation 13: u t =v T tanh(W atten x t +b atten )

[0033] Formula 14:

[0034] Formula 15:

[0035] Among them, u t is the attention score; v is the training parameter vector; W atten b is the parameter matrix for learning. atten For bias terms; α t x represents the attention probability distribution; softmax is the normalized exponential function; t Let t be the t-th input data, where t is 1, 2, 3, 4, ...; This is the attention value, which ranges from [0, 1].

[0036] In another aspect, this invention provides a SOC estimation system for energy storage batteries. This system can be used to implement the SOC estimation method for energy storage batteries as described above. The estimation system may include an acquisition and processing module and a training and prediction module. The acquisition and processing module acquires the voltage, current, and temperature of the energy storage battery to form a dataset and processes the data. The training and prediction module extracts feature data using a two-dimensional convolutional neural network, introduces an attention mechanism, and enables the neural network to automatically learn and selectively focus on important information in the voltage, current, and temperature of the energy storage battery. A quantile loss function is introduced into the LSTM to achieve SOC point and interval prediction for the energy storage battery.

[0037] In another aspect, the present invention provides an electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the estimation method described above.

[0038] Compared with the prior art, the beneficial effects of the present invention include at least one of the following:

[0039] (1) This invention captures spatial features in energy storage battery data through CNN (convolutional neural network) layers, enhancing the expressive power of features. The introduction of QRLSTM (long short-term memory quantile regression neural network) layers enables the model to not only provide accurate point predictions, but also generate uncertainty intervals for predictions, so that the energy storage battery management system can more effectively assess potential risks and performance changes.

[0040] (2) By introducing an attention mechanism, this invention further enhances the attention to key time series features, thereby improving the accuracy and reliability of prediction.

[0041] (3) The present invention can still maintain a high level of prediction accuracy under complex temperature and operating conditions, demonstrating its superiority over traditional methods, and also providing new research directions and application possibilities for future energy storage battery management systems and SOC estimation technology.

[0042] (4) This invention is applicable to the estimation of SOC of energy storage batteries. It not only improves the prediction accuracy, but also provides a more comprehensive and in-depth data analysis tool for energy storage battery management systems. It helps to improve the safety and efficiency of energy storage battery use. It has important practical significance and broad application prospects for related application fields such as the construction of energy storage power stations and energy storage battery system equipment. Attached Figure Description

[0043] The above and other objects and / or features of the present invention will become clearer from the following description taken in conjunction with the accompanying drawings, in which:

[0044] Figure 1 A schematic diagram of the CNN-QRLSTM-Attention structure of the present invention is shown;

[0045] Figure 2 A schematic diagram of the QRLSTM structure of the present invention is shown;

[0046] Figure 3 A schematic diagram of the compressed and stimulated attention of the present invention is shown;

[0047] Figure 4a The diagram shows the SOC prediction results of different models under condition A at 10℃ in Example 1.

[0048] Figure 4b The diagram shows the SOC prediction results of different models under condition A at 25℃ in Example 1.

[0049] Figure 4c The diagram shows the SOC prediction results of different models under condition A at 35℃ in Example 1.

[0050] Figure 4d The diagram shows the algorithm error of different models under condition A at 10℃ in Example 1;

[0051] Figure 4e The diagram shows the algorithm error of different models under condition A at 25℃ in Example 1;

[0052] Figure 4f The diagram shows the algorithm error of different models under condition A at 35℃ in Example 1;

[0053] Figure 5a The following diagram illustrates the performance indicators of different models under condition A at 10℃ in Example 1.

[0054] Figure 5b The following diagram illustrates the performance indicators of different models under condition A at 25°C in Example 1.

[0055] Figure 5c The following diagram illustrates the performance indicators of different models under condition A at 35℃ in Example 1.

[0056] Figure 6a The diagram shows the SOC prediction results of different models under DST conditions at 10℃ in Example 1.

[0057] Figure 6b The diagram shows the SOC prediction results of different models under DST conditions at 25℃ in Example 1.

[0058] Figure 6cThe diagram shows the SOC prediction results of different models under DST conditions at 35℃ in Example 1.

[0059] Figure 6d The diagram shows the algorithm error of different models under DST conditions at 10℃ in Example 1.

[0060] Figure 6e The diagram shows the algorithm error of different models under DST conditions at 25℃ in Example 1;

[0061] Figure 6f This shows a schematic diagram of the algorithm error of different models under DST conditions at 35℃ in Example 1;

[0062] Figure 7a The following diagram illustrates the performance indicators of different models under DST conditions at 10℃ in Example 1.

[0063] Figure 7b The following diagram illustrates the performance indicators of different models under DST conditions at 25°C in Example 1.

[0064] Figure 7c The following diagram illustrates the performance indicators of different models under DST conditions at 35°C in Example 1.

[0065] Figure 8a This diagram illustrates the prediction results of the QRLSTM model in Example 1 at 10℃ under operating condition A in the 90% interval.

[0066] Figure 8b This diagram illustrates the prediction results of the QRLSTM model in Example 1 at 25°C under condition A, covering 90% of the intervals.

[0067] Figure 8c This diagram illustrates the prediction results of the QRLSTM model in Example 1 at 35℃ under condition A, covering 90% of the range.

[0068] Figure 9a This diagram illustrates the prediction results of the QRLSTM model in Example 1 at 10°C under DST conditions in the 90% interval.

[0069] Figure 9b This diagram illustrates the 90% interval prediction results of the QRLSTM model in Example 1 at 25°C under DST conditions.

[0070] Figure 9c This diagram illustrates the prediction results of the QRLSTM model in Example 1 at 35°C under DST conditions in the 90% interval.

[0071] Figure 10aThis diagram illustrates the prediction results of the CNN-QRGRU-Attention model in Example 1 at 10℃ in condition A, covering 90% of the intervals.

[0072] Figure 10b This diagram illustrates the prediction results of the CNN-QRGRU-Attention model in Example 1 at 90% interval under condition A (25℃).

[0073] Figure 10c This diagram illustrates the 90% prediction results of the CNN-QRGRU-Attention model in Example 1 at 35℃ under condition A.

[0074] Figure 11a This diagram illustrates the 90% prediction results of the CNN-QRGRU-Attention model in Example 1 at 10℃ under DST conditions.

[0075] Figure 11b This diagram illustrates the 90% prediction results of the CNN-QRGRU-Attention model in Example 1 at 25°C under DST conditions.

[0076] Figure 11c This diagram illustrates the 90% prediction results of the CNN-QRGRU-Attention model in Example 1 at 35°C under DST conditions.

[0077] Figure 12a This diagram illustrates the prediction results of the CNN-QRLSTM-Attention model in Example 1 for 90% of the intervals at 10℃ under condition A.

[0078] Figure 12b This diagram illustrates the prediction results of the CNN-QRLSTM-Attention model in Example 1 at 90% interval under condition A (25℃).

[0079] Figure 12c This diagram illustrates the 90% prediction results of the CNN-QRLSTM-Attention model in Example 1 at 35℃ under condition A.

[0080] Figure 13a This diagram illustrates the 90% prediction results of the CNN-QRLSTM-Attention model in Example 1 at 10℃ under DST conditions.

[0081] Figure 13b This diagram illustrates the 90% prediction results of the CNN-QRLSTM-Attention model in Example 1 at 25°C under DST conditions.

[0082] Figure 13c The diagram shows the prediction results of the CNN-QRLSTM-Attention model in Example 1 at 90% interval under DST conditions at 35°C. Detailed Implementation

[0083] In the following, a method, system, and electronic device for estimating the state of charge (SOC) of an energy storage battery according to the present invention will be described in detail with reference to the accompanying drawings and exemplary embodiments.

[0084] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0085] The purpose of this invention is to overcome the shortcomings of traditional neural networks in energy storage battery SOC estimation methods. It proposes a hybrid neural network prediction method for estimating SOC points and intervals in energy storage batteries, solving the problems of high-precision estimation and the challenges of energy storage battery equipment construction. Simultaneously, it can generate uncertainty intervals for energy storage battery prediction, enabling the Battery Management System (BMS) to more effectively assess potential risks and performance changes. The proposed method utilizes a hybrid model of a convolutional long short-term memory quantile regression neural network with an attention mechanism to extract voltage, current, and temperature features of the energy storage battery. An attention mechanism is introduced for focused learning, while the long short-term memory quantile regression neural network is used for SOC point and interval prediction, simultaneously advancing the construction of energy storage battery systems. A two-dimensional convolutional neural network is used to extract features from the energy storage battery voltage, current, and temperature data, retaining the most important information. The same convolutional kernel uses the same weights on all input data, reducing the number of parameters that need to be learned, helping to prevent overfitting and improving training and inference efficiency. Introducing an attention mechanism into the neural network model allows the neural network to automatically learn and selectively focus on important information in the energy storage battery voltage, current, and temperature, improving the model's performance and generalizability. This invention proposes a Long Short-Term Memory Quantile Regression Neural Network (QRLSTM), which introduces a quantile loss function (QR) into the standard LSTM (Long Short-Term Memory network) to provide more refined predictions for different quantiles, while also enabling SOC point and interval predictions for energy storage batteries. This is very useful for energy storage battery SOC prediction tasks that need to consider complex operating conditions.

[0086] This invention proposes a hybrid neural network model for estimating the state of charge (SOC) of energy storage batteries, which can be named CNN-QRLSTM-Attention. The network inputs are the voltage, current, and temperature of the energy storage battery, and the output is the estimated SOC value. This model can be used to estimate the SOC of energy storage batteries. The method of using this model is described in detail below.

[0087] Exemplary Example 1

[0088] This exemplary embodiment provides a method for estimating the state of charge (SOC) of an energy storage battery, which may include the following steps:

[0089] Acquire voltage, current, and temperature data of the energy storage battery and process the data;

[0090] A two-dimensional convolutional neural network is used to extract feature data from the processed data to obtain the first feature data.

[0091] An attention mechanism is introduced to enable the neural network to automatically learn and selectively focus on important information in the feature data to obtain the second feature data;

[0092] The second feature data is input into QRLSTM for regression prediction, thereby enabling prediction of the SOC point and range of the energy storage battery.

[0093] Specifically, such as Figure 1 As shown, the method for estimating the state of charge (SOC) of a storage battery using the CNN-QRLSTM-Attention model mainly includes the following steps:

[0094] S1. Feature data of input voltage, current, and temperature are extracted through convolutional layers, and an attention mechanism is used to make the neural network notice important information in the feature data, thereby improving prediction efficiency. Here, a two-dimensional convolutional neural network is used.

[0095] S2. By changing the dimension of the data through tiling layers, a quantile loss function is introduced into the LSTM to obtain QRLSTM. The data with the changed dimension is then input into the QRLSTM for regression prediction.

[0096] S3. Output the predicted SOC points and intervals through the fully connected layer. Here, 90% of the prediction intervals can be provided.

[0097] In this exemplary embodiment, an experimental platform is built to simulate the operating conditions of an energy storage battery under specific ambient temperatures. The voltage, current, and temperature of the energy storage battery are acquired to form a dataset. 60%–70% of the data is used as training input samples for the hybrid neural network, and 30%–40% is used as test input samples. Further, 70% of the data is used as training input samples for the hybrid neural network, and 30% is used as test input samples. After acquiring the dataset, it needs to be normalized. The normalization formula may include the following equation:

[0098] Formula 1:

[0099] Where x represents the original data and y represents the normalized data, xmin x is the minimum value of the original data. ma y represents the maximum value of the original data. min y is the minimum value of the normalized data. max This represents the maximum value of the data after normalization.

[0100] Here, Formula 1 maps the minimum and maximum values ​​of the input data rows to [-1, 1], making the input data evenly distributed within this range, so as to ensure that the data can better exhibit stability and consistency when training machine learning models.

[0101] In this exemplary embodiment, the parameter settings for the energy storage battery SOC estimation method based on the CNN-QRLSTM-Attention hybrid neural network are shown in Table 1:

[0102] Table 1. Parameter settings for the energy storage battery SOC estimation method based on CNN-QRLSTM-Attention hybrid neural network

[0103]

[0104] Here, a single-hidden-layer neural network can be applied to various time series prediction fields, and a QRLSTM neural network with 50 hidden layers can be used. Too many neurons will lead to network complexity and overfitting, while too few will lead to underfitting of the output.

[0105] In this exemplary embodiment, before using a two-dimensional convolutional neural network to extract feature data from the processed data, that is, before the data is input into a hybrid neural network, the input time series data can be folded into two-dimensional data so that it can be input into the convolutional layer for processing.

[0106] In this exemplary embodiment, using a two-dimensional convolutional neural network to extract feature data from the processed data may include: using a two-dimensional convolutional neural network to extract features from the voltage, current, and temperature data of the energy storage battery, retaining the most important information. Simultaneously, the same convolutional kernel uses the same weights across the entire input data, reducing the number of parameters that need to be learned, helping to prevent overfitting and improving training and inference efficiency. Compared to a one-dimensional convolutional neural network, a two-dimensional convolutional neural network can extract richer and more complex features and has better spatial relationship capture capabilities. The formula used for two-dimensional convolution may include the following equation 2:

[0107] Equation 2: G[i,j]=∑ m ∑ n F[m,n]H[im,jn]

[0108] Where G[i,j] is an element in the output feature map; F[m,n] is the input data or the feature map of the previous layer; H[i,j] is the convolution kernel; and m,n are the dimension indices of the convolution kernel.

[0109] In this exemplary embodiment, the two-dimensional convolutional neural network may employ max pooling and the ReLU activation function to discard linearly irrelevant features. The formulas used for max pooling and the ReLU activation function may include the following equations 3 and 4, respectively:

[0110] Formula 3:

[0111] Equation 4: f(x) = max(0,x)

[0112] Among them, S l (j) is the output of the j-th pooling region in the l-th layer; w represents the width of the pooling region; X l-1 f(t) is the pooling region; x is the input; f(x) is the output.

[0113] In this exemplary embodiment, quantile regression is a statistical technique used to model and analyze different quantiles of data. Unlike traditional least squares regression, it focuses not only on the mean or median of the dependent variable, but can focus on any quantile of the data, making quantile regression particularly useful when dealing with data that is heteroscedastic or nonnormally distributed. Figure 2 Given the explanatory variable X and the response variable Y, the quantile regression model can be expressed as: Q Y (τ|X)=X T β(τ), where β(τ) is the regression coefficient vector at the τ quantile, τ∈(0,1); X is the explanatory variable, Y is the response variable; Q Y (τ|X) represents the response variable Y with respect to the explanatory variables. When combined with LSTM, it is used as the loss function, and the expression for the quantile loss function can include the following equation:

[0114] Formula 5:

[0115] Where β(τ) is the regression coefficient vector at the τ quantile, τ∈(0,1); X is the explanatory variable, Y is the response variable; Q Y (τ|X) represents the response variable Y with respect to the explanatory variables; ρ τ (u) is the sloping absolute value function.

[0116] In this exemplary embodiment, the tilted absolute value function ρ τ The expression for (u) can include the following formula 6:

[0117] Formula 6:

[0118] Here, when τ = 0.5, it can be used as the quantile regression output of the QRLSTM neural network point prediction, which can be expressed as: QRLSTM is a combined neural network that integrates quantile regression with LSTM. This combination leverages the sequence processing capabilities of LSTM and the sensitivity of quantile regression to asymmetric data distributions. LSTM first processes time-series data, capturing its time dependencies and dynamic features, and then quantile regression is applied to the LSTM output. Figure 2 As shown, the QRLSTM neural network structure is the same as LSTM, but the input of each quantile needs to be trained by LSTM. The essence of this method is to assign different weights to different quantiles, thereby obtaining fitting functions for different quantiles.

[0119] In this exemplary embodiment, the expression for LSTM may include the following equations: 7, 8, 9, 10, 11, and 12:

[0120] Formula 7:

[0121] Formula 8:

[0122] Equation 9: H t =O t ⊙tanh(Ct)

[0123] Equation 10: I t =σ(x t W xi +H t-1 W hi +b i )

[0124] Equation 11: F t =σ(x t W xf +H t-1 W hf +b f )

[0125] Formula 12: O t =σ(x t W xo +H t-1 W ho +b o )

[0126] in, For memory units; C t Memory for the input of the next time step; H t For the hidden state of the next time step; I t For input gate; Ft Forgotten Gate; O t For output gate; W xi W xf W xo W hi W hf and W ho Indicates the weight parameter; b i b f and b o H represents the bias parameter. t-1 σ represents the hidden state of the previous time step; σ is the sigmoid activation function; and ⊙ is the Hadamard product operator. Here, the memory unit... The core of LSTM is the input gate, which mainly maintains the long-term state. The input gate determines which new information will be stored in the cell state, the forget gate determines which information should be forgotten or discarded from the cell state, and the output gate determines the next hidden state, that is, the output of the next time step and which parts of the current cell state.

[0127] In this exemplary embodiment, as Figure 3 As shown, compression and attention activation are used for the model's focused learning. The input two-dimensional data X is processed by the convolution operation F. tr The feature map U is then obtained, and then the extrusion module F is introduced. sq and incentive module F ex The squeezing module converts the 2D data into scalar values ​​using global average pooling, resulting in a 1×1×C matrix. The activation module then operates on this 1×1×C matrix using the sigmoid activation function to obtain a weighted 1×1×C color matrix. Finally, the dot product module F... scale Multiply this colored matrix by U to obtain the weighted two-dimensional data.

[0128] Expressions that introduce attention mechanisms may include the following equations: Equation 13, Equation 14, and Equation 15:

[0129] Equation 13: u t =v T tanh(W atten x t +b atten )

[0130] Formula 14:

[0131] Formula 15:

[0132] Among them, u t is the attention score; v is the training parameter vector; W atten b is the parameter matrix for learning.atten For bias terms; α t x represents the attention probability distribution; softmax is the normalized exponential function; t Let t be the t-th input data, where t is 1, 2, 3, 4, ...; This is the attention value, which ranges from [0, 1].

[0133] Here, Formula 13 can be used to calculate the attention score, using a scoring function to calculate the current input data x. t The importance of the input within the entire two-dimensional data X is such that a higher attention score indicates the need for more attention to the current input. The attention score u is then obtained. t Then, the attention probability distribution α needs to be normalized using the softmax function. t As shown in Formula 14. Finally, for each input data x t By weighting, we obtain the attention value. As shown in Formula 15.

[0134] Exemplary Example 2

[0135] This exemplary embodiment provides an energy storage battery SOC estimation system, which can be used to implement the energy storage battery SOC estimation method as described in Exemplary Embodiment 1.

[0136] The estimation system may include an acquisition and processing module and a training and prediction module.

[0137] The acquisition and processing module can acquire the voltage, current and temperature of the energy storage battery, form a dataset, and process the data.

[0138] The training and prediction module uses a two-dimensional convolutional neural network to extract feature data and introduces an attention mechanism to enable the neural network to automatically learn and selectively focus on important information in the voltage, current and temperature of the energy storage battery. A quantile loss function is introduced into the LSTM to achieve prediction of the SOC point and interval of the energy storage battery.

[0139] Exemplary Example 3

[0140] This exemplary embodiment provides an electronic device.

[0141] The electronic device may include a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the estimation method as described in Exemplary Example 1.

[0142] To better understand the above exemplary embodiments 1-3, we will provide a more detailed explanation below with reference to specific examples.

[0143] Example 1

[0144] Cycle A is a test procedure specifically designed for evaluating and testing buses operating in the XX region. Cycle A aims to simulate the actual driving conditions of buses in the XX region, taking into account the varied traffic environment, including frequent starts and stops, different speeds, and congestion. It is a test condition that conforms to common real-world usage scenarios and can effectively reflect the performance of lithium-ion batteries under actual use conditions.

[0145] Dynamic stress test (DST) is a battery test condition proposed in the USABC Electric Vehicle Battery Test Procedure Manual. It evaluates the battery's responsiveness, durability, and performance stability by simulating the battery usage of an electric vehicle under real road conditions.

[0146] 1. Perform SOC point prediction performance analysis for energy storage batteries:

[0147] In this example, voltage, current, and temperature data of lithium-ion batteries under conditions A and DST (Digital Subtraction Angle) were acquired using an experimental platform at ambient temperatures of 10℃, 25℃, and 35℃. The CNN-QRLSTM-Attention model was then tested, with 70% of the data used as training data and 30% as test data. Performance was compared with CNN-QRGRU-Attention, CNN-LSTM-Attention, QRLSTM, and CNN-LSTM models. For algorithms involving quantile regression, the 0.5 quantile was used as the prediction point. The SOC prediction results of different models under condition A at 10℃, 25℃, and 35℃ are shown below. Figure 4a , 4b and Figure 4c As shown, the algorithm errors are respectively as follows: Figure 4d , 4e and Figure 4f As shown, the performance index results are as follows: Figure 5a , 5b and Figure 5c As shown in the figure. The SOC prediction results of different models under DST conditions at three temperatures: 10℃, 25℃, and 35℃ are respectively shown in the figure. Figure 6a , 6b and Figure 6c As shown, the algorithm errors are respectively as follows: Figure 6d , 6e and Figure 6f As shown, the performance index results are as follows: Figure 7a , 7b and Figure 7c As shown.

[0148] Depend on Figures 4a-4f and Figures 6a-6fClearly, the CNN-QRLSTM-Attention model lithium-ion battery estimation method of this invention exhibits excellent prediction performance and small errors under different ambient temperatures, in both Condition A and Condition DST, and can predict SOC values ​​with high accuracy. Compared with the CNN-LSTM-Attention algorithm, quantile regression, by modeling different quantiles, better adapts to and predicts outliers. Furthermore, the introduction of quantile regression allows the LSTM model to not only focus on the central trend of the data but also to more comprehensively capture the characteristics of the entire data distribution, thereby greatly improving the overall prediction accuracy. Compared with CNN-QRLSTM, LSTM's forget gate allows the network to learn when to forget old information, and LSTM can better handle long-term dependencies, thus performing better than GRU in applications such as lithium-ion battery SOC estimation. Compared with QRLSTM, CNN, through convolution operations on the input data, obtains richer and more complex features, and can simultaneously understand complex spatial and temporal patterns, thereby improving the overall prediction accuracy. By introducing an attention mechanism, the neural network model can selectively focus on important information in the input, thereby improving the model's performance and generalizability.

[0149] Depend on Figures 5a-5c Therefore, under condition A, when the ambient temperature is 10℃, the MAE, RMSE, and R of CNN-QRLSTM-Attention are... 2 The MAE, RMSE, and R of CNN-QRLSTM-Attention were 0.424%, 0.576%, and 0.99949, respectively, at an ambient temperature of 25°C. 2 The MAE, RMSE, and R-values ​​of CNN-QRLSTM-Attention were 0.574%, 0.807%, and 0.99915, respectively, at an ambient temperature of 35°C. 2 They were 0.473%, 0.645%, and 0.99948, respectively. (By...) Figures 7a-7c It can be seen that, under DST conditions, when the ambient temperature is 10℃, the MAE, RMSE, and R of CNN-QRLSTM-Attention are... 2 The MAE, RMSE, and R-values ​​of CNN-QRLSTM-Attention were 0.289%, 0.481%, and 0.99966, respectively, at an ambient temperature of 25°C. 2 The MAE, RMSE, and R-values ​​of CNN-QRLSTM-Attention were 0.256%, 0.409%, and 0.99978, respectively, at an ambient temperature of 35°C. 2The values ​​are 0.228%, 0.366%, and 0.99983, respectively. Clearly, under different temperatures and operating conditions, the MAE and RMSE of CNN-QRLSTM-Attention are lower than other related algorithms. A smaller MAE indicates that the model's predictions are closer to the actual values, resulting in higher accuracy. A smaller RMSE indicates that the model's prediction accuracy is higher while also exhibiting less instantaneous error fluctuation. A larger R... 2 This indicates that the model has a better fit and can better explain the variation in the observed data.

[0150] 2. Perform SOC range prediction performance analysis for energy storage batteries:

[0151] Traditional regression models typically predict the conditional mean of the response variable, while quantile regression focuses on different quantiles of the conditional distribution. This means that quantile regression can provide information about the distribution of data at different levels, not just the central tendency. By predicting different quantiles, the LSTM of quantile regression can provide a range for the predicted value. For example, by calculating the 5th and 95th quantiles, a 90% prediction range can be obtained, which means that there is a 90% probability that future observations will fall within this prediction range. This method can effectively quantify the uncertainty of the prediction. In lithium-ion battery SOC prediction, the 90% range prediction has important practical significance and application value: (1) Enhanced safety: The 90% prediction range provides the range of uncertainty about the SOC estimate, which helps to identify potential safety risks, such as overcharging or over-discharging. (2) Optimized battery management: By providing a reliable SOC prediction range, the charge and discharge cycles of the battery can be managed more effectively, extending battery life and reducing maintenance costs. (3) Improved reliability: In power systems or electric vehicles, battery reliability is crucial. The 90% prediction range allows system operators to prepare for potential performance degradation while maintaining efficient energy use. (4) Better adaptability: The performance of lithium-ion batteries is affected by a variety of factors, including temperature, aging, charge and discharge rates. The 90% prediction range can take into account the uncertainties brought about by these factors, providing a more comprehensive SOC estimate. At the same time, the battery operating conditions can be very complex and variable. The range prediction takes into account possible changes and anomalies, providing a robust estimate applicable under different conditions.

[0152] In this example, the prediction results of the QRLSTM model in the 90% interval at different temperatures under condition A are as follows: Figure 8a , 8b As shown in Figure 8c, the prediction results of the QRLSTM model in the 90% interval at different temperatures under DST conditions are as follows: Figure 9a , 9bAs shown in Figure 9c, the prediction results of the CNN-QRGRU-Attention model in the 90% interval under different temperatures in condition A are as follows: Figure 10a , 10b As shown in Figure 10c, the prediction results of the CNN-QRGRU-Attention model in the 90% interval under different temperatures in DST conditions are as follows: Figure 11a , 11b As shown in Figure 11c, the prediction results of the CNN-QRLSTM-Attention model in the 90% interval under different temperatures in condition A are as follows: Figure 12a , 12b As shown in Figure 12c, the prediction results of the CNN-QRLSTM-Attention model in the 90% interval under different temperatures in DST conditions are as follows: Figure 13a , 13b As shown in 13c.

[0153] Depend on Figures 8a-8c , Figures 9a-9c , Figures 10a-10c , Figures 11a-11c , Figures 12a-12c and Figures 13a-13c As can be seen, the CNN-QRLSTM-Attention model of the present invention has a narrower prediction interval, indicating that it has strong stability while having high point prediction accuracy, so that the model can always maintain a good prediction interval width.

[0154] The interval prediction performance indicators of each model under condition A are shown in Table 2.

[0155] Table 2A shows the interval prediction performance indicators of each model under the working condition.

[0156]

[0157] Table 3 shows the interval prediction performance of each model under the DST condition.

[0158] Table 3. Interval prediction performance indicators of each model under DST conditions.

[0159]

[0160]

[0161] Table 2 shows that, under condition A, the PICP (Probability of Predicted Interval Coverage) of CNN-QRLSTM-Attention is 0.90004, 0.9077, and 0.89489 at the 90% nominal confidence level (PINC) for 10℃, 25℃, and 35℃, respectively, and 0.85642, 0.81747, and 0.85237 at the 80% nominal confidence level (PINC). Table 3 shows that, under condition DST, the PICP of CNN-QRLSTM-Attention is 0.91364, 0.93395, and 0.91361 at the 90% nominal confidence level (PINC) for 10℃, 25℃, and 35℃, respectively, and 0.74354, 0.84983, and 0.84859 at the 80% nominal confidence level (PINC). Compared to the other two algorithms that use quantile regression, CNN-QRLSTM-Attention has significantly lower PINAW (prediction interval average bandwidth) and CWC (coverage width-based standard), indicating that the method of this invention has a more compact prediction interval and higher prediction accuracy.

[0162] In summary, this invention captures spatial features in energy storage battery data through CNN layers, enhancing the expressive power of these features. The introduction of QRLSTM layers enables the model to not only provide accurate point predictions but also generate uncertainty intervals for predictions, allowing the energy storage battery management system to more effectively assess potential risks and performance changes. The use of attention mechanisms further improves the model's focus on key time-series features, enhancing the accuracy and reliability of predictions. Moreover, the model maintains a high level of prediction accuracy even under complex temperature and operating conditions, enabling the battery management system to more effectively assess potential risks and performance changes.

[0163] Although the invention has been described above in conjunction with exemplary embodiments, those skilled in the art will understand that various modifications and changes can be made to the exemplary embodiments of the invention without departing from the spirit and scope defined by the claims.

Claims

1. A method for estimating the State of Charge (SOC) of an energy storage battery, characterized in that, The method comprises the following steps: acquiring voltage, current and temperature data of the energy storage battery, processing the data; using a two-dimensional convolutional neural network to extract feature data from the processed data to obtain first feature data; introducing an attention mechanism to enable the neural network to automatically learn and selectively focus on important information in the first feature data to obtain second feature data; inputting the second feature data into a QRLSTM for regression prediction to realize point and interval prediction of the SOC of the energy storage battery.

2. The energy storage battery SOC estimation method of claim 1, wherein, The processing includes normalization processing, and the formula used in the normalization processing includes the following formula 1: Formula 1: Wherein, x is the original data, y is the normalized data, x min is the minimum value of the original data, x max is the maximum value of the original data, y min is the minimum value of the normalized data, y max is the maximum value of the normalized data.

3. The method of claim 1, wherein, Before the feature data extraction from the processed data using the two-dimensional convolutional neural network, the input time series data is folded into two-dimensional data to be input into the convolutional layer for processing.

4. The energy storage battery SOC estimation method according to claim 1 or 3, characterized by, The feature data extraction from the processed data using the two-dimensional convolutional neural network comprises: using a two-dimensional convolutional neural network to extract features from the voltage, current and temperature data of the energy storage battery, retaining the most important information, while the same convolution kernel uses the same weight on the entire input data.

5. The energy storage battery SOC estimation method according to claim 1 or 4, characterized by, The formula used in the two-dimensional convolution includes the following formula 2: Formula 2: G[i, j] = ∑ m ∑ n F[m, n] H[i - m, j - n] where G[i, j] is an element in the output feature map; F[m, n] is the input data or the feature map of the previous layer; H[i, j] is the convolution kernel; m, n is the dimension index of the convolution kernel.

6. The method of claim 1, wherein, The two-dimensional convolutional neural network uses the maximum pooling technique and the RELU activation function to discard linearly irrelevant features, and the formula used in the maximum pooling and the RELU activation function respectively includes the following formula 3 and formula 4: Formula 3: Formula 4: f(x) = max(0, x) where S l (j) is the output of the jth pooling region of the lth layer; w represents the width of the pooling region; X l-1 (t) is the pooling region; x is the input; f(x) is the output.

7. The method of claim 1, wherein, The expression of the quantile loss function includes the following formula 5: Formula 5: where β(τ) is the regression coefficient vector at the τ-quantile, τ ∈ (0, 1); X is the explanatory variable, Y is the response variable; Q Y (τ|X) represents the corresponding response variable Y under the condition of the explanatory variable; p τ (u) is the tilted absolute value function.

8. The method of claim 7, wherein, The ρ τ The expression of (u) includes the following formula 6: Formula 6: where p τ (u) is a tilted absolute value function.

9. The method of claim 1, wherein, The expression of the LSTM includes the following formula 7, formula 8, formula 9, formula 10, formula 11 and formula 12: Formula 7: Formula 8: Formula 9: H t =O t ⊙tanh(C t ) Formula 10: I t = σ(x t W xi + H t-1 W hi + b i ) Formula 11: F t = σ(x t W xf + H t-1 W hf + b f ) Formula 12: O t = σ(x t W xo +H t-1 W ho +b o ) wherein, is the memory cell; C t is the input memory of the next time step; H t is the hidden state of the next time step; I t is the input gate; F t is the forget gate; O t is the output gate; W xi , W xf , W xo , W hi , W hf , and W ho denote weight parameters; b i , b f , and b o denote bias parameters; H t-1 is the hidden state of the previous time step; σ is the sigmod activation function; and is the Hadamard product operator.

10. The method of claim 1, wherein, The expression of the attention mechanism includes the following formula 13, formula 14 and formula 15: Formula 13: u t = v T tanh(W atten x t + b atten ) Formula 14: Formula 15: where u t is the attention score; v is the parameter vector of training; W atten is the parameter matrix of learning; b atten is the bias term; a t is the attention probability distribution; softmax is the normalization exponential function; x t is the tth input data, t is 1, 2, 3, 4, …; is the attention value, the value range is [0, 1].

11. An energy storage battery SOC estimation system, comprising: The system is used to realize the energy storage battery SOC estimation method in any one of claims 1 to 10, and the estimation system comprises an acquisition and processing module and a training and prediction module, wherein The acquisition and processing module can acquire the voltage, current and temperature of the energy storage battery, form a data set, and process the data; The training and prediction module extracts feature data using a two-dimensional convolutional neural network, introduces an attention mechanism to enable the neural network to automatically learn and selectively focus on important information in the voltage, current and temperature of the energy storage battery, and introduces a quantile loss function in the LSTM to realize point and interval prediction of the SOC of the energy storage battery. 12.An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein, The processor realizes the steps of the method in any one of claims 1 to 10 when executing the computer program.

Images