SOH prediction and active management method and device based on battery in modular multilevel converter

Abstract

The invention relates to an SOH prediction and active management method and device based on a battery in a modular multilevel converter, and belongs to the technical field of battery energy storage systems and power electronics. The method comprises the following steps: synchronously acquiring voltage, current, temperature and high-frequency impedance spectrum of a sub-module battery, identifying equivalent circuit model parameters on line by using an adaptive forgetting factor least square method, and constructing a model error spectrum on a time domain and an incremental capacity curve on a frequency domain; a high-dimensional joint aging feature vector is formed by combining the average temperature and the number of cycles, feature importance is weighted through quantum attention coding and then mapped to a hidden space quantum state of a quantum variational auto-encoder, a quantum graph convolutional neural network is used for conducting graph structure learning and reasoning on quantum state features, SOH estimated values of sub-modules are output, and the SOH estimated values of the sub-modules are calculated. And finally, embedding the SOH estimated value as a weight factor into a model control cost function of the MMC, dynamically optimizing the switching sequence of the sub-modules, and realizing active balance management of the aging difference of the battery.

Description

Technical Field

[0001] This invention belongs to the field of battery energy storage systems and power electronics technology, specifically relating to a method and device for predicting and actively managing the state of energy (SOH) of batteries in a modular multilevel converter. Background Technology

[0002] Energy storage batteries are widely used in MMC battery energy storage systems, such as renewable energy integration, electric vehicles, and smart grids. However, batteries experience aging differences during cycling, leading to an imbalance in the State of Health (SOH) of each submodule, which in turn causes a decrease in system efficiency, a shortened lifespan, or even failure. Traditional SOH estimation methods mainly rely on empirical formulas, ampere-hour integration methods, or equivalent circuit models, but these methods have limitations: empirical formulas have low accuracy, ampere-hour integration methods are prone to error accumulation, and equivalent circuit models rely on a large number of parameters and are difficult to handle nonlinear aging characteristics. In addition, existing hardware solutions, such as additional sensors or equalizers, increase cost and complexity. Therefore, this invention addresses the above problems by proposing a data-driven and quantum machine learning-based method. This method synchronously collects battery data, constructs multi-dimensional feature vectors, and uses a quantum graph convolutional neural network (QGCNN) for SOH estimation. Finally, the results are embedded into the MMC control strategy to achieve proactive equalization of aging differences. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of traditional methods and provide a method and device for SOH prediction and active management of batteries in modular multilevel converters. This method improves the accuracy of SOH estimation by extracting multi-dimensional features (such as model error spectrum and incremental capacity curve), uses quantum graph convolutional neural network (QGCNN) to reduce and map high-dimensional nonlinear features to improve computational efficiency, and embeds the SOH estimate into the model prediction control cost function to dynamically optimize the sub-module switching order, thereby achieving active balancing of aging differences and achieving rapid online estimation and dynamic balancing of battery SOH.

[0004] To achieve the above objectives, the technical solution of this invention is: a method for predicting and actively managing the state of aging (SOH) of batteries in a modular multilevel converter (MMC). This method involves synchronously collecting the voltage, current, temperature, and high-frequency impedance spectrum of each MMC submodule battery, using the adaptive forgetting factor least squares (FF-LS) method to identify equivalent circuit model parameters online, constructing a model error spectrum in the time domain and an incremental capacity curve in the frequency domain, and combining the average temperature and cycle number to form a high-dimensional joint aging feature vector. After weighting the feature importance using quantum attention encoding (QAE), the vector is mapped to the hidden space quantum state of a quantum variational autoencoder (QVAE), achieving a low-dimensional quantum representation of the high-dimensional features. A quantum graph convolutional neural network (QGCNN) is used to learn and infer the graph structure of the quantum state features, outputting the SOH estimate of the MMC submodule. Finally, the SOH estimate is embedded as a weighting factor into the model control cost function of the MMC, dynamically optimizing the switching order of the MMC submodules, and achieving active balanced management of battery aging differences.

[0005] Furthermore, the method includes the following steps:

[0006] Step (1): For each MMC submodule battery, simultaneously acquire voltage, current and temperature timing data for at least one complete charge-discharge cycle, and simultaneously acquire high-frequency impedance spectrum.

[0007] Step (2): Identify the parameters of the equivalent circuit model online using the adaptive forgetting factor least squares method (FF-LS), update the model in real time, and calculate the theoretical terminal voltage. The eigenvalues ​​of the dynamic model error spectrum are obtained by performing a difference and integration with the measured values. Multi-scale wavelet packet decomposition was used to extract the peak height of the incremental capacity curve at different scales for the capacity-voltage curve. With peak position Peak width and sub-band energy entropy For high-frequency impedance spectra, extract the impedance magnitudes at characteristic frequency points. Phase angle The impedance spectrum fitting residuals were calculated. ;

[0008] Step (3) and average temperature and number of cycles After the multidimensional feature vector is weighted by the quantum attention encoding QAE, it is mapped to the hidden space quantum state of the quantum variational autoencoder QVAE, realizing data compression and low-dimensional quantum representation of high-dimensional features;

[0009] Step (4): Construct the battery aging relationship graph of the sub-module, input the features of the quantum state representation into the pre-trained quantum graph convolutional neural network QGCNN, capture the graph structure through the quantum convolutional layer, compress the quantum information through the pooling layer, and finally output the SOH estimate of the MMC sub-module;

[0010] Step (5): The SOH estimate of the MMC submodule is sent to the MMC modulation and equalization control module. The reciprocal of SOH is used as the penalty weight to embed the control cost function to optimize the switching function of the MMC submodule, so as to realize the active equalization of battery aging differences by the MMC converter.

[0011] Furthermore, the specific steps for obtaining the parameters in step (2) are as follows:

[0012] Step (2.1): Initialize the parameters of the adaptive forgetting factor least squares method (FF-LS) and set the initial estimate of the parameter vector. Initial values ​​of the covariance matrix Initial forgetting factor Lower limit of forgetting factor noise variance estimate ;

[0013] For each submodule, the battery terminal voltage is synchronously acquired at time k. and current I(k), combined with open-circuit voltage Constructing regression vectors :

[0014]

[0015] Calculate the error between the model prediction and the measured value. The formula is as follows:

[0016]

[0017] in , This is the parameter vector at time k-1, i.e., the previous time.

[0018] In the calculation of the equivalent circuit model using FF-LS, an adaptively adjusted forgetting factor is employed. When model error When the value exceeds the preset value, it indicates that the battery characteristics may have changed, and the old model is no longer applicable. The value will be automatically reduced. When the error is less than the preset value, Close to 1; The calculation formula is as follows:

[0019]

[0020] The recursive formula for the noise variance is as follows:

[0021]

[0022] in It is a smoothing factor;

[0023] In the calculation of the equivalent circuit model using FF-LS, the weights of the new observations on the parameter corrections are applied using the gain matrix. To express, The specific calculation formula is as follows:

[0024]

[0025] in The covariance matrix is ​​calculated using the following formula:

[0026]

[0027] Where I is the identity matrix;

[0028] The prior error is obtained through calculation. The optimal estimated parameters for the current time are calculated after determining the gain matrix K(k) and covariance matrix P(k). The final theoretical model voltage calculation equation is as follows:

[0029]

[0030] in The parameter vector to be identified;

[0031] Step (2.2), Model error spectrum eigenvalues The calculation formula is:

[0032]

[0033] Where N is the number of integration windows in one charge-discharge cycle; This is the measured voltage. This is the theoretical terminal voltage; and These represent the end time and start time of the k-th integration window, respectively, used for discrete integration calculation;

[0034] Step (2.3), Incremental capacity curve sampling calculation formula:

[0035]

[0036] The original IC curve data obtained from sampling is subjected to a five-point cubic smoothing process to obtain a smoothed IC curve. Then, Gaussian fitting is applied to the smoothed IC curve to obtain the fitted Gaussian function form as follows:

[0037]

[0038] in Extract peak height as peak width parameter. With peak voltage position ;

[0039] Peak width Obtained by Gaussian fitting The formula was then calculated as follows:

[0040]

[0041] Subband energy entropy is obtained from multi-scale wavelet packet decomposition. The formula is as follows:

[0042]

[0043] in The percentage of each sub-energy band, where M is the number of the last layer in the multi-scale wavelet packet decomposition;

[0044] Step (2.4): Extract the impedance modulus at characteristic frequency points from the high-frequency impedance spectrum. Phase angle The impedance spectrum fitting residuals are obtained by analyzing the impedance spectrum. ,in The calculation formula is as follows:

[0045]

[0046] in To measure the obtained impedance, The complex impedance of the model is calculated from the best fitted parameters. This represents the number of frequency points in the impedance spectrum.

[0047] Furthermore, in step (3), Average temperature and number of cycles Composition of multidimensional feature vectors Multidimensional feature vectors The specific manifestations are as follows:

[0048]

[0049] By using the parameterized quantum circuit of quantum attention encoding (QAE) to weight the feature importance in the multidimensional feature vector, and then mapping it to the hidden space quantum state of quantum variational autoencoder (QVAE), a low-dimensional quantum representation of high-dimensional features is realized.

[0050] In the process of mapping to the quantum state, each eigenvector is first normalized and its length is adjusted to 16, that is, the last 6 positions are padded with zeros to form a 16-dimensional vector. and ensure its mold length The value is 1; then a parameterized quantum circuit is constructed; 4 qubits are used to encode 10 battery health characteristics, with the specific mapping relationship as follows:

[0051] Quantum bits Encoding features ; Quantum bits Encoding features ; Quantum bits Encoding features ; Quantum bits Encoding features ;

[0052] In the encoding, the added underscore in the subscript indicates the corresponding data after normalization;

[0053] Designing a parameterized quantum circuit involves three key steps for each layer of the parameterized quantum circuit design:

[0054] Feature importance adjustment: By applying parameterized rotation gate operations sequentially to each qubit, using The gate adjusts the phase component of the quantum state to capture the periodic patterns of the feature, using The probability amplitude of the gate-adjusted quantum state is used to adjust the eigenvalue weights.

[0055] Feature interaction modeling: CNOT gates are used to simulate the nonlinear interactions and dependencies between different features, and entanglement is established between adjacent qubits;

[0056] Fine-tuning: The RZ rotating gate is used for phase fine-tuning to complete the single-layer transformation;

[0057] The entire parameterized quantum circuit can be represented as a unitary transformation sequence:

[0058]

[0059] in parameters It is adjustable; It is a Z-shaped revolving door. For controlled NOT gate, is the Y-rotation gate, and ∘ is the composition operator, representing the sequential composition of multiple quantum gate operations;

[0060] Encoded quantum state Through parameterized circuits The transformed quantum state is obtained. The data is passed down to the downstream quantum variational autoencoder (QVAE).

[0061] The quantum variational autoencoder (QVAE) receives the transformed quantum state. It performs low-dimensional compression and dimensionality reduction on high-dimensional features through its encoding part, and outputs compressed quantum state data. .

[0062] Furthermore, in step (4), the working steps of the quantum graph convolutional neural network, QGCNN, are as follows:

[0063] Step (4.1): Define the node working state vector , Let the average temperature of the k-th submodule be . Let be the mean absolute value of the current in the sampling window corresponding to the k-th submodule. The current SOC is obtained by looking up a table after using the FF-LS algorithm to identify the equivalent circuit model parameters in real time and calculate the open-circuit voltage based on the current measured value;

[0064] Step (4.2): Use weighted Euclidean distance to process the distances between working vector data.

[0065]

[0066] Where W1, W2, and W3 are the weighting coefficients. , , These represent the average temperature, mean absolute current value, and state of charge of the i-th submodule, respectively.

[0067] By setting a distance threshold Generate a binary, undirected adjacency matrix A;

[0068] Step (4.3): Perform quantum encoding of the graph structure, and project the graph structure represented by the adjacency matrix A into a quantum Hamiltonian. superior:

[0069]

[0070] This indicates the similarity of the operating conditions between submodules i and j. When the working conditions are similar, The operating conditions are not similar. , It acts on the i-th qubit, that is, represents the Pauli operator on the i-th cell, and so on. Let J be the Pauli operator for the j-th battery;

[0071] Application by Generated parameterized quantum gates , For learning parameters; the output of the quantum variational autoencoder (QVAE) The resulting initial quantum state representing the characteristics of all nodes Sent by The evolution occurs within the participating quantum circuits, resulting in the quantum state of each cell. No longer independent, it becomes a quantum superposition state containing its own original information and information from all its connected neighboring batteries with similar operating conditions, through training to learn parameters. In order to find the optimal way to aggregate information;

[0072] Step (4.4): Feature compression and refinement are performed through quantum pooling to reduce the number of qubits; after multiple layers and pooling, the remaining qubits are measured to obtain a set of classical expected values. These expected values ​​are input into a small fully connected neural network to perform a simple nonlinear transformation, which maps the quantum measurement values ​​to the final SOH value, thus obtaining the SOH estimate of each original submodule battery.

[0073] Furthermore, in step (5), the SOH weighted cost function in the MMC modulation and equalization control module is:

[0074]

[0075] Where α, β, and γ are weighting coefficients. For theoretical reference current, To actually measure the current, This is the theoretical reference voltage for the battery. For the actual measured voltage of the battery, The health status penalty weight of the k-th submodule is given in the summation symbol. For the MMC submodule in the bridge arm, k is the module number;

[0076] In each control cycle, the J values ​​are sorted in ascending order, and the top M submodules with the lowest J values ​​are selected to be put into operation, while the remaining submodules are bypassed. This makes the system tend to put more submodules with high SOH into operation while tracking current and voltage commands, and to prioritize bypassing or reduce the operation time of submodules with low SOH.

[0077] For submodules with abnormally high J values ​​(J>1), a forced bypass is implemented and a maintenance alarm is issued. Meanwhile, during normal operation, the system dynamically adjusts the power distribution among the submodules to achieve proactive balanced management of battery aging differences, thereby delaying overall system aging and extending lifespan.

[0078] The present invention also provides a SOH prediction and active management device for a battery in a modular multilevel converter, comprising:

[0079] The data acquisition module is used to acquire the voltage, current, temperature time-series data and high-frequency impedance spectrum information of the batteries in each MCC submodule;

[0080] The multi-dimensional feature processing unit is used to perform online identification of equivalent circuit model parameters and output model error spectrum feature values. It is also used to calculate and extract characteristic parameters of the incremental capacity curve IC: peak height. Peak position Peak width and sub-band energy entropy Simultaneously, the impedance modulus at characteristic frequency points is extracted from the high-frequency impedance spectrum. Phase angle The impedance spectrum fitting residuals were calculated. ;

[0081] The quantum convolutional neural network inference module is used to process the features and average temperature obtained from the multi-dimensional feature processing unit. and number of cycles A multidimensional feature vector is formed, which is mapped to a low-dimensional quantum state through quantum attention encoding (QAE) and quantum variational autoencoder (QVAE); a pre-trained QGCNN model is loaded, which receives the quantum state features, performs calculations, and outputs the SOH estimate of each sub-module;

[0082] The MMC modulation and equalization control module is used to adjust the switching order of sub-modules based on the SOH estimate to achieve active aging equalization.

[0083] Furthermore, the multi-dimensional feature processing unit identifies the parameters of the equivalent circuit model online based on the adaptive forgetting factor least squares method (FF-LS).

[0084] The present invention also provides an electronic device, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0085] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0086] Compared with the prior art, the present invention has the following beneficial effects:

[0087] 1. This invention integrates the error spectrum features of the time-domain model with the incremental capacity curve features of the frequency domain to form a high-dimensional joint aging feature vector. Then, it uses a quantum graph convolutional neural network (QGCNN) to perform deep learning and dimensionality reduction on the nonlinear features, significantly improving the accuracy of SOH estimation.

[0088] 2. This invention embeds the SOH estimate as a weighting factor into the cost function of model predictive control, establishing a direct correlation mechanism between health status and operating strategy. Through a dynamic optimization algorithm using the reciprocal of SOH as a penalty weight, the system intelligently adjusts the switching order of sub-modules, reducing the load on severely aged batteries, effectively delaying the overall aging of the battery pack, and improving the economy and sustainability of the energy storage system.

[0089] 3. Based on the self-learning characteristics of quantum machine learning, this invention can adapt to the health status assessment needs of different battery types and various complex working conditions. Attached Figure Description

[0090] Figure 1 System block diagram;

[0091] Figure 2 This is a diagram of a quantum graph convolutional neural network structure.

[0092] Figure 3 This is a schematic diagram of an MMC converter;

[0093] Figure 4 This is a parameter comparison chart obtained after MATLAB simulation of the device of the present invention;

[0094] Figure 5 This is a comparison chart of SOH consistency. Detailed Implementation

[0095] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0096] This invention provides a method for predicting and actively managing the state of aging (SOH) of batteries in a modular multilevel converter (MMC). By synchronously collecting the voltage, current, temperature, and high-frequency impedance spectrum of each MMC submodule battery, the equivalent circuit model parameters are identified online using the adaptive forgetting factor least squares (FF-LS) method. This constructs a model error spectrum in the time domain and an incremental capacity curve in the frequency domain. A high-dimensional joint aging feature vector is then formed by combining the average temperature and cycle number. The feature importance is weighted using quantum attention encoding (QAE) and mapped to the latent quantum state of a quantum variational autoencoder (QVAE), achieving a low-dimensional quantum representation of the high-dimensional features. A quantum graph convolutional neural network (QGCNN) is used to learn and infer the graph structure of the quantum state features, outputting the SOH estimate of the MMC submodule. Finally, the SOH estimate is embedded as a weighting factor into the model control cost function of the MMC, dynamically optimizing the switching order of the MMC submodules and achieving proactive balanced management of battery aging differences.

[0097] The present invention also provides a SOH prediction and active management device for a battery in a modular multilevel converter, comprising:

[0098] The data acquisition module is used to acquire the voltage, current, temperature time-series data and high-frequency impedance spectrum information of the batteries in each MCC submodule;

[0099] The multi-dimensional feature processing unit is used to perform online identification of equivalent circuit model parameters and output model error spectrum feature values. It is also used to calculate and extract characteristic parameters of the incremental capacity curve IC: peak height. Peak position Peak width and sub-band energy entropy Simultaneously, the impedance modulus at characteristic frequency points is extracted from the high-frequency impedance spectrum. Phase angle The impedance spectrum fitting residuals were calculated. ;

[0100] The quantum convolutional neural network inference module is used to process the features and average temperature obtained from the multi-dimensional feature processing unit. and number of cycles A multidimensional feature vector is formed, which is mapped to a low-dimensional quantum state through quantum attention encoding (QAE) and quantum variational autoencoder (QVAE); a pre-trained QGCNN model is loaded, which receives the quantum state features, performs calculations, and outputs the SOH estimate of each sub-module;

[0101] The MMC modulation and equalization control module is used to adjust the switching order of sub-modules based on the SOH estimate to achieve active aging equalization.

[0102] The following is a detailed implementation process of the present invention.

[0103] like Figure 1 As shown, this invention provides a method for SOH prediction and active management of batteries in a modular multilevel converter. The method involves a device including a data acquisition module, a multi-dimensional feature processing unit, a quantum convolutional neural network inference module, and an MMC modulation and equalization control module. The data acquisition module is used to acquire voltage, current, and temperature time-series data, as well as high-frequency impedance spectrum information, of the batteries in each sub-module. The multi-dimensional feature processing unit is used to perform online identification of equivalent circuit model parameters and output model error spectrum feature values. Characteristic parameters used to calculate and extract the incremental capacity curve IC: peak height Peak position Peak width and sub-band energy entropy A quantum convolutional neural network inference module is used to map classical feature vectors to low-dimensional quantum states through quantum attention encoding (QAE) and quantum variational autoencoder (QVAE); loaded with a pre-trained QGCNN model (such as...). Figure 2 As shown in the diagram, the module receives quantum state characteristics, performs calculations, and outputs the SOH estimate for each submodule. The MMC modulation and equalization control module is used to adjust the switching order of submodules based on the SOH estimate to achieve active aging equalization. Figure 3 This is a schematic diagram of an MMC converter.

[0104] The specific implementation steps of this method are as follows:

[0105] Step 1: Using the data acquisition module, test and collect the battery voltage over at least one battery charge-discharge cycle. Current Temperature data And high-frequency impedance spectrum information, and record the number of charge-discharge cycles. Because the timestamps and spatial coordinates of data collected by different sensors differ, a unified spatiotemporal reference is required. A time alignment algorithm maps all data to a unified time axis, resulting in a synchronized data sequence that provides a precise data foundation for subsequent feature extraction.

[0106] Step 2: Identify the equivalent circuit model parameters online using the adaptive forgetting factor least squares (FF-LS) method, update the model in real time, and calculate the theoretical terminal voltage. The eigenvalues ​​of the dynamic model error spectrum are obtained by performing a difference and integration with the measured values. Multi-scale wavelet packet decomposition was used to extract the peak height of the incremental capacity curve at different scales for the capacity-voltage curve. With peak position Peak width and sub-band energy entropy For high-frequency impedance spectra, extract the impedance magnitudes at characteristic frequency points. Phase angle The impedance spectrum fitting residuals were calculated. The details are as follows:

[0107] Step (2.1) initializes the parameters of the adaptive forgetting factor least squares (FF-LS) method. The parameter vector... Perform initialization operations to make Initialize the covariance matrix ,make , Take the larger value, such as 10 6 Initial forgetting factor Lower limit of forgetting factor Noise variance estimate .

[0108] Dynamic model error spectrum eigenvalues The calculation process requires the following operation at each sampling time k: measuring the battery terminal voltage. and current And query Constructing regression vectors According to the parameters of the previous time step calculate Calculation error Update the forgetting factor Calculate the gain matrix Update the parameter matrix Calculate the latest .

[0109] in and The calculation formula is as follows:

[0110]

[0111]

[0112] Step (2.2) finally sets the theoretical terminal voltage. Compared with measured values The eigenvalues ​​of the error spectrum of the dynamic model are obtained by performing difference integration. The formula is as follows:

[0113]

[0114] Where N is the number of integration windows in one charge-discharge cycle; This is the measured voltage. The theoretical terminal voltage, and These represent the end time and start time of the k-th integration window, respectively, used for discrete integration calculation;

[0115] Step (2.3) involves acquiring the unprocessed capacity curve (IC) discrete point data and smoothing it. The number of decomposition levels M for the multi-scale wavelet packet decomposition is determined to be 4 levels, and the smoothed IC curve signal is considered as the initial node. That is, the 0th layer and the 0th node. For the signal of the i-th node in the M-th layer... The signal is decomposed by low-pass filtering and downsampling, and high-pass filtering and downsampling, using a pair of orthogonal mirror filter banks to obtain two branch coefficients. This decomposition is repeated until the number of decomposition layers M is 4. Wavelet packet reconstruction is performed on the coefficients of the last layer to obtain the time-domain (voltage-domain) signal corresponding to the frequency band. The main peak in the reconstructed signal is fitted using a Gaussian function to obtain the optimal parameters.

[0116] Extracting peak height from a Gaussian function Peak voltage location Peak width parameter And obtain the peak width .

[0117] Calculate subband energy entropy First, calculate all 2s in the last layer (the Mth layer). M The energy of each subband, for the Mth subband, is:

[0118]

[0119] Let M be the coefficient length of the Mth subband. Calculate the total energy. :

[0120]

[0121] Calculate the energy percentage of each sub-band Finally, we obtained :

[0122]

[0123] Step (2.4) uses the data acquisition device to obtain the impedance magnitude at the extracted characteristic frequency point. Phase angle The data was analyzed and calculated to obtain the impedance spectrum fitting residuals. The specific value.

[0124] Several key frequencies sensitive to aging were selected as f=100Hz, 50Hz, 10Hz, and 1Hz, and the impedance magnitude at these frequencies was directly read. and phase angle The measured impedance spectrum data were compared with the established equivalent circuit model, and the overall difference between the measured and actual values ​​was calculated. .

[0125] Step 3: Put and average temperature and number of cycles The multidimensional feature vectors are weighted by the feature importance using quantum attention encoding (QAE) and then mapped to the hidden space quantum state of a quantum variational autoencoder (QVAE), achieving data compression and low-dimensional quantum representation of high-dimensional features; specifically as follows:

[0126] Step (3.1) will obtain the data By distinguishing and combining modules according to their identifiers, multidimensional classic feature vectors are obtained, such as...

[0127]

[0128] in represent Mutually, For the upper bridge arm, For the first Submodules (SM);

[0129] Step (3.2) involves weighting the feature importance in the multidimensional feature vector using the parameterized quantum circuit of quantum attention encoding (QAE) and mapping it to the hidden space quantum state of quantum variational autoencoder (QVAE) to realize the low-dimensional quantum representation of high-dimensional features.

[0130] In the process of mapping to the quantum state, each eigenvector is first normalized and its length is adjusted to 16, that is, the last 6 positions are padded with zeros to form a 16-dimensional vector. and ensure its mold length The value is 1. A parameterized quantum circuit is then constructed. Four qubits are used to encode ten battery health characteristics, with the specific mapping relationship as follows:

[0131] Quantum bits Encoding features ; Quantum bits Encoding features ; Quantum bits Encoding features ; Quantum bits Encoding features

[0132] In encoding, such as , This is the corresponding data after normalization.

[0133] Designing a parameterized quantum circuit involves three key steps for each layer of the parameterized quantum circuit design:

[0134] Feature importance adjustment. This is achieved by sequentially applying parameterized rotation gate operations to each qubit, using... The phase component of the quantum state is adjusted by a gate to capture the periodic pattern of the characteristic. Using... The probability amplitude of the gate-adjusted quantum state is used to adjust the feature importance weight.

[0135] Feature interaction modeling. CNOT gates are used to simulate the nonlinear interactions and dependencies between different features, establishing entanglement between adjacent qubits.

[0136] Fine-tuning. The RZ rotating gate is applied again for phase fine-tuning to complete the single-layer transformation.

[0137] The entire parameterized quantum circuit can be represented as a unitary transformation sequence:

[0138]

[0139] in parameters It is adjustable; It is a Z-shaped revolving door. For controlled NOT gate, It is a Y-shaped revolving door;

[0140] parameter The specific values ​​are obtained from the initial training. During the training process, the line parameters... It will self-adjust, so that the probability amplitude of quantum states corresponding to features that are more important to the task objective is amplified, while the probability amplitude of features that are not important is suppressed.

[0141] Encoded quantum state Through parameterized circuits The transformed quantum state is obtained. The data is passed down to the downstream quantum variational autoencoder (QVAE).

[0142] The quantum variational autoencoder accepts the transformed quantum state It performs low-dimensional compression and dimensionality reduction on high-dimensional features through its encoding part, and outputs compressed quantum state data. By using QAE feature weighting, important features are prioritized for preservation during compression. Then, QVAE dimensionality reduction reduces the impact of noisy features, improving overall feature quality. This process reduces feature dimensionality while preserving key information from aging features, directly serving the Quantum Graph Convolutional Neural Network (QGCNN) for SOH estimation, thus improving computational efficiency and estimation accuracy.

[0143] Step (3.3) The quantum variational autoencoder receives the transformed quantum state. It performs low-dimensional compression and dimensionality reduction on high-dimensional features through its encoding part, and outputs compressed quantum state data. By using QAE feature weighting, important features are prioritized for preservation during compression. Then, QVAE dimensionality reduction reduces the impact of noisy features, improving overall feature quality. This process reduces feature dimensionality while preserving key information from aging features, resulting in output quantum state data. It will be directly used for SOH estimation in quantum graph convolutional neural networks (QGCNN), improving computational efficiency and estimation accuracy.

[0144] Step 4: Construct the battery aging relationship graph for the sub-module, input the features of the quantum state representation into the pre-trained quantum graph convolutional neural network (QGCNN), capture the graph structure through quantum convolutional layers, compress quantum information through pooling layers, and finally output the SOH estimate of the sub-module, as follows:

[0145] Step (4.1) Define the node working state vector , Let the average temperature of the k-th submodule be . The average absolute value of the current in the sampling window corresponding to the k-th submodule is used. The open-circuit voltage (OCV) of the equivalent circuit model is identified in real time by the FF-LS algorithm, and the current value is obtained after looking up the table. value.

[0146] Step (4.2) uses weighted Euclidean distance to process the distances between the working vector data:

[0147]

[0148] Where W1, W2, and W3 are the weighting coefficients (W1=0.5, W2=0.3, W3=0.2). , , These represent the average temperature, mean absolute current value, and state of charge of the i-th submodule, respectively.

[0149] Next, a similarity threshold is set. This is done by setting a distance threshold. According to the threshold This generates a binary, undirected adjacency matrix A. Adjacency matrix A defines the graph structure at the current time step.

[0150] Step (4.3) involves quantum encoding of the graph structure by mapping the graph structure represented by the adjacency matrix A to a quantum Hamiltonian. superior. The calculation formula is as follows:

[0151]

[0152] This indicates the similarity of the operating conditions between submodules i and j. When the working conditions are similar, The operating conditions are not similar. , It acts on the i-th qubit, that is, represents the Pauli operator on the i-th cell, and so on. Let J be the Pauli operator for the j-th battery;

[0153] Application by Generated parameterized quantum gates It will be by The resulting initial quantum state representing the characteristics of all nodes Sent by The evolution occurs within the participating quantum circuits, resulting in the quantum state of each cell. It is no longer independent. It has become a quantum superposition state containing its original information as well as information from all its connected (i.e., similarly functioning) neighboring batteries. Parameters can be learned through training. In order to find the optimal way to aggregate information.

[0154] Step (4.4) reduces the number of qubits by performing feature compression and refinement through quantum pooling. After multiple layers and pooling, the remaining qubits are measured to obtain a set of classical expectation values. These expectation values ​​are then input into a small classical fully connected neural network. The role of this network is to perform a simple nonlinear transformation, mapping the quantum measurement values ​​to the final SOH value. This yields a high-precision SOH estimate for each original submodule battery.

[0155] Step 5: The SOH estimate is fed into the MMC modulation and equalization control module. The reciprocal of the SOH is used as a penalty weight to be embedded in the control cost function to optimize the switching function of the submodule, thereby realizing the active equalization of battery aging differences in the MMC converter. The details are as follows:

[0156] Step (5.1) In the MMC modulation and equalization control module, based on the SOH weighted cost function:

[0157]

[0158] Where α, β, and γ are weighting coefficients. For theoretical reference current, To actually measure the current, This is the theoretical reference voltage for the battery. For the actual measured voltage of the battery, The health status penalty weight of the k-th submodule is given in the summation symbol. is the MMC submodule in the bridge arm, and k is the module number.

[0159] In each control cycle, the SOH values ​​are sorted in descending order, and the top M modules with the highest SOH values ​​are selected for operation, while the remaining modules are bypassed. This makes the system, while tracking current and voltage commands, tend to put more sub-modules with high SOH (healthy) into operation, while giving priority to bypassing or reducing the operation time of sub-modules with low SOH (severe aging).

[0160] For battery submodules with abnormally high SOH values, a forced bypass is implemented and a maintenance alarm is issued. Simultaneously, during normal operation, the system dynamically adjusts the power distribution among submodules through the aforementioned optimization process, achieving proactive balanced management of battery aging differences, thereby delaying overall system aging and extending its lifespan.

[0161] The parameter comparison graph obtained after MATLAB simulation of this device is shown below. Figure 4 ,from Figure 4 The comparison chart of SOH estimation accuracy shows that, compared with the traditional method, the SOH estimation value of the present invention is closer to the true SOH value, and its estimation accuracy is improved by an average of 60%, especially in the middle stage of battery aging.

[0162] from Figure 5 As can be seen from the comparison chart of SOH consistency, the battery health estimation consistency under the control of the present invention is significantly improved compared with the traditional scheme, which is beneficial to the management of battery power and the extension of battery life.

[0163] The present invention also provides an electronic device, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0164] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0165] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for predicting and actively managing the state of harm (SOH) of a battery in a modular multilevel converter, characterized in that, By synchronously collecting the voltage, current, temperature, and high-frequency impedance spectrum of each MMC submodule battery, the equivalent circuit model parameters are identified online using the adaptive forgetting factor least squares (FF-LS) method. This constructs the model error spectrum in the time domain and the incremental capacity curve in the frequency domain. Combined with the average temperature and cycle count, a high-dimensional joint aging feature vector is formed. After weighting the feature importance using quantum attention encoding (QAE), the vector is mapped to the hidden space quantum state of a quantum variational autoencoder (QVAE), achieving a low-dimensional quantum representation of the high-dimensional features. A quantum graph convolutional neural network (QGCNN) is used to learn and infer the graph structure of the quantum state features, outputting the SOH estimate of the MMC submodule. Finally, the SOH estimate is embedded as a weighting factor into the model control cost function of the MMC, dynamically optimizing the switching order of the MMC submodules and achieving proactive balanced management of battery aging differences.

2. The method for SOH prediction and active management of batteries in a modular multilevel converter according to claim 1, characterized in that, The method includes the following steps: Step (1): For each MMC submodule battery, simultaneously acquire voltage, current and temperature timing data for at least one complete charge-discharge cycle, and simultaneously acquire high-frequency impedance spectrum. Step (2): Identify the parameters of the equivalent circuit model online using the adaptive forgetting factor least squares method (FF-LS), update the model in real time, and calculate the theoretical terminal voltage. The eigenvalues ​​of the dynamic model error spectrum are obtained by performing a difference and integration with the measured values. Multi-scale wavelet packet decomposition was used to extract the peak height of the incremental capacity curve at different scales for the capacity-voltage curve. Peak position Peak width and sub-band energy entropy For high-frequency impedance spectra, extract the impedance magnitudes at characteristic frequency points. Phase angle The impedance spectrum fitting residuals were calculated. ; Step (3) and average temperature and number of cycles After the multidimensional feature vector is weighted by the quantum attention encoding QAE, it is mapped to the hidden space quantum state of the quantum variational autoencoder QVAE, realizing data compression and low-dimensional quantum representation of high-dimensional features; Step (4): Construct the battery aging relationship graph of the sub-module, input the features of the quantum state representation into the pre-trained quantum graph convolutional neural network QGCNN, capture the graph structure through the quantum convolutional layer, compress the quantum information through the pooling layer, and finally output the SOH estimate of the MMC sub-module; Step (5): The SOH estimate of the MMC submodule is sent to the MMC modulation and equalization control module. The reciprocal of SOH is used as the penalty weight to embed the control cost function to optimize the switching function of the MMC submodule, so as to realize the active equalization of battery aging differences by the MMC converter.

3. The method for predicting and actively managing the state of harm (SOH) of a battery in a modular multilevel converter according to claim 2, characterized in that, The specific steps for obtaining the parameters in step (2) are as follows: Step (2.1): Initialize the parameters of the adaptive forgetting factor least squares method (FF-LS) and set the initial estimate of the parameter vector. Initial values ​​of the covariance matrix Initial forgetting factor Lower limit of forgetting factor noise variance estimate ; For each submodule, the battery terminal voltage is synchronously acquired at time k. and current I(k), combined with open-circuit voltage Constructing regression vectors : Calculate the error between the model prediction and the measured value. The formula is as follows: in , This is the parameter vector at time k-1, i.e., the previous time. In the calculation of the equivalent circuit model using FF-LS, an adaptively adjusted forgetting factor is employed. When model error When the value exceeds the preset value, it indicates that the battery characteristics may have changed, and the old model is no longer applicable. The value will be automatically reduced. ; When error When the preset value is used, Close to 1; The calculation formula is as follows: The recursive formula for the noise variance is as follows: in It is a smoothing factor; In the calculation of the equivalent circuit model using FF-LS, the weights of the new observations on the parameter corrections are applied using the gain matrix. To express, The specific calculation formula is as follows: in The covariance matrix is ​​calculated using the following formula: Where I is the identity matrix; The prior error is obtained through calculation. The optimal estimated parameters for the current time are calculated after determining the gain matrix K(k) and covariance matrix P(k). The final theoretical model voltage calculation equation is as follows: in The parameter vector to be identified; Step (2.2), Model error spectrum eigenvalues The calculation formula is: Where N is the number of integration windows in one charge-discharge cycle; This is the measured voltage. The theoretical terminal voltage, and These represent the end time and start time of the k-th integration window, respectively, used for discrete integration calculation; Step (2.3), Incremental capacity curve sampling calculation formula: The original IC curve data obtained from sampling is subjected to a five-point cubic smoothing process to obtain a smoothed IC curve. Then, Gaussian fitting is applied to the smoothed IC curve to obtain the fitted Gaussian function form as follows: in Extract peak height as peak width parameter. With peak voltage position ; Peak width Obtained by Gaussian fitting The formula was then calculated as follows: Subband energy entropy is obtained from multi-scale wavelet packet decomposition. The formula is as follows: in The percentage of each sub-energy band, where M is the number of the last layer in the multi-scale wavelet packet decomposition; Step (2.4): Extract the impedance modulus at characteristic frequency points from the high-frequency impedance spectrum. Phase angle The impedance spectrum fitting residuals are obtained by analyzing the impedance spectrum. ,in The calculation formula is as follows: in To measure the obtained impedance, The complex impedance of the model is calculated from the best fitted parameters. This represents the number of frequency points in the impedance spectrum.

4. The method for SOH prediction and active management of batteries in a modular multilevel converter according to claim 2, characterized in that, In step (3), Average temperature and number of cycles Composition of multidimensional feature vectors Multidimensional feature vectors The specific manifestations are as follows: By using the parameterized quantum circuit of quantum attention encoding (QAE) to weight the feature importance in the multidimensional feature vector, and then mapping it to the hidden space quantum state of quantum variational autoencoder (QVAE), a low-dimensional quantum representation of high-dimensional features is realized. In the process of mapping to the quantum state, each eigenvector is first normalized and its length is adjusted to 16, that is, the last 6 positions are padded with zeros to form a 16-dimensional vector. and ensure its mold length The value is 1; then a parameterized quantum circuit is constructed; 4 qubits are used to encode 10 battery health characteristics, with the specific mapping relationship as follows: Quantum bits Encoding features ; Quantum bits Encoding features ; Quantum bits Encoding features ; Quantum bits Encoding features ; In the encoding, the added underscore in the subscript indicates the corresponding data after normalization; Designing a parameterized quantum circuit involves three key steps for each layer of the parameterized quantum circuit design: Feature importance adjustment: By applying parameterized rotation gate operations sequentially to each qubit, using The gate adjusts the phase component of the quantum state to capture the periodic patterns of the feature, using The probability amplitude of the gate-adjusted quantum state is used to adjust the eigenvalue weights. Feature interaction modeling: CNOT gates are used to simulate the nonlinear interactions and dependencies between different features, and entanglement is established between adjacent qubits; Fine-tuning: The RZ rotating gate is used for phase fine-tuning to complete the single-layer transformation; The entire parameterized quantum circuit can be represented as a unitary transformation sequence: in parameters It is adjustable; It is a Z-shaped revolving door. For controlled NOT gate, is the Y-rotation gate, and ∘ is the composition operator, representing the sequential composition of multiple quantum gate operations; Encoded quantum state Through parameterized circuits The transformed quantum state is obtained. The data is passed down to the downstream quantum variational autoencoder (QVAE). The quantum variational autoencoder (QVAE) receives the transformed quantum state. It performs low-dimensional compression and dimensionality reduction on high-dimensional features through its encoding part, and outputs compressed quantum state data. .

5. A method for predicting and actively managing the state of harm (SOH) of a battery in a modular multilevel converter according to claim 2, characterized in that, In step (4), the working steps of the quantum graph convolutional neural network, QGCNN, are as follows: Step (4.1): Define the node working state vector , Let the average temperature of the k-th submodule be . Let be the mean absolute value of the current in the sampling window corresponding to the k-th submodule. The current SOC is obtained by looking up a table after the open-circuit voltage is calculated by using the FF-LS algorithm to identify the equivalent circuit model parameters in real time and the current measurement value. Step (4.2): Use weighted Euclidean distance to process the distances between working vector data. Where W1, W2, and W3 are the weighting coefficients. , , These represent the average temperature, mean absolute current value, and state of charge of the i-th submodule, respectively. By setting a distance threshold Generate a binary, undirected adjacency matrix A; Step (4.3): Perform quantum encoding of the graph structure, and project the graph structure represented by the adjacency matrix A into a quantum Hamiltonian. superior: This indicates the similarity of the operating conditions between submodules i and j. When the working conditions are similar, The operating conditions are not similar. , It acts on the i-th qubit, that is, represents the Pauli operator on the i-th cell, and so on. Let J be the Pauli operator for the j-th battery; Application by Generated parameterized quantum gates , For learning parameters; the output of the quantum variational autoencoder (QVAE) The resulting initial quantum state representing the characteristics of all nodes Sent by The evolution occurs within the participating quantum circuits, resulting in the quantum state of each cell. No longer independent, it becomes a quantum superposition state containing its own original information and information from all its connected neighboring batteries with similar operating conditions, through training to learn parameters. In order to find the optimal way to aggregate information; Step (4.4): Feature compression and refinement are performed through quantum pooling to reduce the number of qubits; after multiple layers and pooling, the remaining qubits are measured to obtain a set of classical expected values. These expected values ​​are input into a small fully connected neural network to perform a simple nonlinear transformation, which maps the quantum measurement values ​​to the final SOH value, thus obtaining the SOH estimate of each original submodule battery.

6. The method for SOH prediction and active management of batteries in a modular multilevel converter according to claim 2, characterized in that, In step (5), the SOH weighted cost function in the MMC modulation and equalization control module is: Where α, β, and γ are weighting coefficients. For theoretical reference current, To actually measure the current, This is the theoretical reference voltage for the battery. For the actual measured voltage of the battery, The health status penalty weight of the k-th submodule is given in the summation symbol. For the MMC submodule in the bridge arm, k is the module number; In each control cycle, the J values ​​are sorted in ascending order, and the top M submodules with the lowest J values ​​are selected to be put into operation, while the remaining submodules are bypassed. This makes the system tend to put more submodules with high SOH into operation while tracking current and voltage commands, and to prioritize bypassing or reduce the operation time of submodules with low SOH. For submodules with abnormally high J values ​​(J>1), a forced bypass is implemented and a maintenance alarm is issued. Meanwhile, during normal operation, the system dynamically adjusts the power distribution among the submodules to achieve proactive balanced management of battery aging differences, thereby delaying overall system aging and extending lifespan.

7. A device for predicting and actively managing the state of harm (SOH) of a battery in a modular multilevel converter, characterized in that, include: The data acquisition module is used to acquire the voltage, current, temperature time-series data and high-frequency impedance spectrum information of the batteries in each MCC submodule; The multi-dimensional feature processing unit is used to perform online identification of equivalent circuit model parameters and output model error spectrum feature values. It is also used to calculate and extract characteristic parameters of the incremental capacity curve IC: peak height. Peak position Peak width and sub-band energy entropy Simultaneously, the impedance modulus at characteristic frequency points is extracted from the high-frequency impedance spectrum. Phase angle The impedance spectrum fitting residuals were calculated. ; The quantum convolutional neural network inference module is used to process the features and average temperature obtained from the multi-dimensional feature processing unit. and number of cycles A multidimensional feature vector is formed, which is mapped to a low-dimensional quantum state through quantum attention encoding (QAE) and quantum variational autoencoder (QVAE); a pre-trained QGCNN model is loaded, which receives the quantum state features, performs calculations, and outputs the SOH estimate of each sub-module; The MMC modulation and equalization control module is used to adjust the switching order of sub-modules based on the SOH estimate to achieve active aging equalization.

8. The SOH prediction and active management device for a battery in a modular multilevel converter according to claim 7, characterized in that, A multi-dimensional feature processing unit is used to identify the parameters of the equivalent circuit model online based on the adaptive forgetting factor least squares method (FF-LS).

9. An electronic device, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, which, when executed by the processor, enable the implementation of the steps of the method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, It stores computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, it can implement the steps of the method as described in any one of claims 1-6.

Images