A method and system for evaluating the health state of an energy storage system

Abstract

The application discloses a kind of energy storage system health state evaluation method and system, mainly related to health state evaluation technical field, to solve the problems that prior art cannot identify key degradation signal on frequency domain, cannot reflect the entropy change law of non-stationary change, the insufficient perception of degradation area of deep neural network.For solving the problems that prior art cannot identify key degradation signal on frequency domain, cannot reflect the entropy change law of non-stationary change, the insufficient perception of degradation area of deep neural network, the application includes: real-time acquisition of the original monitoring data of energy storage system, calculate the normalized data after time domain and frequency domain processing, calculate the characteristics of enhanced monitoring data;Residual attention deep neural network is constructed, and the current adaptive learning rate is adjusted autonomously;Dynamic focal point loss is calculated, and total loss function is calculated;The trained residual attention deep neural network is obtained;Current monitoring data is obtained, the characteristics of enhanced current monitoring data are calculated, and then the trained residual attention deep neural network is used to evaluate the health state of energy storage system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of health state evaluation, in particular to a method and system for evaluating the health state of an energy storage system. BACKGROUND

[0002] With the wide application of new energy technologies, energy storage systems, especially lithium-ion battery energy storage devices, have played a key role in grid peak shaving, renewable energy grid connection, electric transportation, etc. However, during long-term operation, energy storage systems often exhibit performance degradation phenomena such as capacity attenuation, internal resistance increase, and temperature abnormalities.

[0003] The existing solution is to collect the working parameters of each working energy storage battery pack in real time, collect the environmental parameters of the environment where each working energy storage battery pack is located in real time, calculate the fault diagnosis coefficient of each working energy storage battery pack according to the working parameters and environmental parameters of each working energy storage battery pack, and perform fault diagnosis on each working energy storage battery pack according to the fault diagnosis coefficient of each working energy storage battery pack. For the working energy storage battery pack that appears to be faulty, a standby energy storage battery pack is selected based on the rated capacity, charging efficiency and discharging efficiency of the energy storage battery pack to replace it.

[0004] However, the above technical solution still has the following problems that need to be further solved: conventional Min-Max or Z-score normalization cannot identify key degradation signals in the frequency domain, especially for weak amplitude frequency changes, which makes the degradation state indistinguishable. Traditional sliding window statistical features cannot reflect the entropy change law of non-stationary changes, and have insufficient ability to identify abnormal fluctuations or early degradation trends, which has obvious lag. Conventional deep neural networks have insufficient perception of degradation areas and cannot dynamically allocate feature weights, which can easily cause feature redundancy propagation or key information to be submerged. SUMMARY

[0005] In view of the above deficiencies of the prior art, the present application provides a method and system for evaluating the health state of an energy storage system to solve the above technical problems.

[0006] In a first aspect, the present application provides a method for evaluating the health state of an energy storage system, the method comprising:

[0007] real-time acquisition of raw monitoring data of the energy storage system, storage of the raw monitoring data according to a preset time window;

[0008] time scale unified processing of the raw monitoring data in the preset time window to obtain time normalization results, frequency domain processing of the raw monitoring data to obtain frequency compression results; and calculation of normalized data after time domain and frequency domain processing through the time normalization results and the frequency compression results;

[0009] Based on the normalized data processed in time domain and frequency domain, the fluctuation characteristics of the time series are modeled by combining the fluctuation entropy mechanism, and the features of the enhanced monitoring data are calculated;

[0010] A residual attention deep neural network is constructed by stacking an input layer, a plurality of residual attention modules, a hybrid regularization layer, and a state output layer in sequence; the current adaptive learning rate is adjusted autonomously by monitoring the variance change of the parameter gradient of each round in the training process of the residual attention deep neural network;

[0011] The residual attention deep neural network outputs the health state probability of each health state category; the dynamic focus loss is calculated based on the health state probability of each health state category and the preset loss weight corresponding to each category;

[0012] The dynamic focus loss, the time series continuity loss, and the attention sparsity regularization loss are obtained, and the total loss function is calculated; the trainable parameters are updated based on the total loss function and the adaptive learning rate; when the preset stopping iteration condition is met, the trained residual attention deep neural network is obtained;

[0013] The current monitoring data is obtained, the features of the enhanced current monitoring data are calculated, and then the trained residual attention deep neural network is used for energy storage system health state evaluation.

[0014] In an implementation manner of the present application, the original monitoring data of the energy storage system is collected in real time, specifically including:

[0015] The original monitoring data is collected in real time by a sensor network deployed on the preset key components of the energy storage system;

[0016] The data sources of the original monitoring data include voltage sensors, current sensors, and temperature sensors of battery monomers / modules, and charge and discharge state information recorded by the battery management system;

[0017] The original monitoring data is collected by synchronous high-frequency sampling of 1 kHz or more.

[0018] In an implementation manner of the present application, the original monitoring data in a preset time window is processed in a time scale to obtain a time normalization result, and the original monitoring data is processed in a frequency domain to obtain a frequency compression result; the normalized data processed in time domain and frequency domain is calculated based on the time normalization result and the frequency compression result, specifically including:

[0019] The mean and standard deviation of the original monitoring data in the preset time window are calculated in the time dimension, and the data at the current time is centered and scaled based on this to unify the time scale, and the time normalization result is obtained;

[0020] The time normalization result is mapped to a frequency domain space by a frequency domain transformation method to obtain a frequency spectrum amplitude; and the frequency spectrum amplitude is compressed to a preset amplitude dynamic range by using a logarithmic function;

[0021] The time normalization result is multiplied by the frequency domain compression result element by element to obtain normalized data processed in the time domain and the frequency domain.

[0022] In an implementation manner of the present application, based on the normalized data processed in the time domain and the frequency domain, a fluctuation characteristic of the time sequence is modeled in combination with a fluctuation entropy mechanism, and enhanced monitoring data features are calculated, specifically including:

[0023] Extraction data of the normalized data at different time points is extracted by a sliding time window, and distances between the extraction data and a plurality of center vectors generated in advance are calculated;

[0024] Probabilities that the normalized data belongs to each center are constructed by using exponential mapping according to the distances, and a fluctuation entropy vector corresponding to the normalized data is calculated according to the probabilities of each center;

[0025] The normalized data and the corresponding fluctuation entropy vector are spliced to obtain the enhanced monitoring data features.

[0026] In an implementation manner of the present application, a residual attention deep neural network composed of an input layer, a plurality of residual attention modules, a hybrid regularization layer and a state output layer is stacked in sequence, specifically including:

[0027] The input layer receives the features of the enhanced monitoring data, and maps the features of the enhanced monitoring data to an initial high-dimensional feature space of the network;

[0028] The residual path in the residual attention module extracts a preset general feature from the features of the enhanced monitoring data in the high-dimensional feature space, and the attention path dynamically adjusts a feature weight; by the preset general feature, the features of the enhanced monitoring data, the self-attention weight and a Sigmoid activation function, a feature processed by the residual attention is obtained;

[0029] The hybrid regularization layer randomly discards a preset part of the output of the residual attention module with a preset probability to obtain a first feature, adopts a spectral normalization mechanism to perform spectral normalization processing on the output of the residual attention module to obtain a second feature, and obtains a feature processed by the hybrid regularization of the first feature and the second feature;

[0030] The state output layer obtains a basic probability of each health state category by a conventional classification function;

[0031] Based on the number difference between each health state category, a state transition matrix is constructed;

[0032] According to the base probability, the feature after the mixed regularization processing, and the state transition matrix, a health state probability is calculated.

[0033] In an implementation of the present application, by monitoring the variance change of the parameter gradient of each round in the residual attention deep neural network training process, the current adaptive learning rate is automatically adjusted, specifically including:

[0034] The gradient variance of the current round is calculated by ratio with the gradient variance of the last round; wherein the current round is greater than 1;

[0035] If the current gradient variance is large, the adaptive learning rate is relaxed; otherwise, the adaptive learning rate is gradually reduced; wherein the relaxation / reduction change range of the learning rate is realized by an exponential decay mechanism.

[0036] In an implementation of the present application, according to the health state probability of each health state category and the preset loss weight corresponding to each category, a dynamic focus loss is calculated, specifically including:

[0037] The formula is:

[0038] The dynamic focus parameter of the c-th category is calculated ; wherein, represents the number of samples of the c-th category, c∈[1,C], C represents the total number of health state categories, represents the median of the number of samples of each category;

[0039] The formula is:

[0040] The loss weight of the c-th category is calculated ; wherein, represents the total number of samples, represents a preset constant;

[0041] The formula is:

[0042] The dynamic focus loss is calculated ; wherein, represents the health state probability of the c-th category;

[0043] In an implementation of the present application, the dynamic focus loss, the time series continuity loss, and the attention sparsity regularization loss are obtained, and a total loss function is calculated, specifically including:

[0044] The weight values corresponding to the dynamic focus loss, the time series continuity loss, and the attention sparsity regularization loss are obtained, and then the total loss function is obtained by weighting.

[0045] ​In an implementation form of the application, the trainable parameters of the iteration are updated based on the total loss function and the adaptive learning rate, specifically comprising:

[0046] The gradient of the trainable parameters is calculated through the total loss function and the trainable parameters of the current round;

[0047] The gradient of the trainable parameters is clipped in a gradient clipping manner;

[0048] The momentum term is updated using the clipped gradient and the adaptive learning rate, and a preset weight decay term is added to suppress overfitting;

[0049] The trainable parameters of the next iteration are generated based on the current momentum term, the updated momentum term, the adaptive learning rate, the trainable parameters of the current round, and the gradient of the clipped trainable parameters.

[0050] In a second aspect, the application provides a health state evaluation system for an energy storage system, which comprises:

[0051] The acquisition module is configured to acquire original monitoring data of the energy storage system in real time and store the original monitoring data according to a preset time window;

[0052] The normalization calculation module is configured to perform time scale uniform processing on the original monitoring data in the preset time window to obtain a time normalization result, perform frequency domain processing on the original monitoring data to obtain a frequency compression result, and calculate normalized data processed in the time domain and the frequency domain based on the time normalization result and the frequency compression result;

[0053] The enhancement calculation module is configured to model the fluctuation characteristics of the time series based on the normalized data processed in the time domain and the frequency domain and in combination with a fluctuation entropy mechanism, and calculate the features of the enhanced monitoring data;

[0054] The learning rate adjustment module is configured to construct a residual attention deep neural network composed of an input layer, a plurality of residual attention modules, a hybrid regularization layer, and a state output layer stacked in sequence, and autonomously adjust the current adaptive learning rate by monitoring the variance of the parameter gradient of each round in the training process of the residual attention deep neural network;

[0055] The loss calculation module is configured to output the health state probability of each health state category through the residual attention deep neural network, calculate a dynamic focal loss based on the health state probability of each health state category and the preset loss weight corresponding to each category, and obtain the total loss function based on the dynamic focal loss, the time series continuity loss, and the attention sparsity regularization loss;

[0056] The iterative training module is configured to update the trainable parameters of the iteration based on the total loss function and the adaptive learning rate, and obtain the trained residual attention deep neural network when a preset stopping iteration condition is met.

[0057] The real-time evaluation module is used to obtain current monitoring data, calculate the enhanced features of the current monitoring data, and then use the trained residual attention deep neural network to evaluate the health status of the energy storage system.

[0058] Those skilled in the art will appreciate that this application has at least the following beneficial effects:

[0059] 1. Combine time normalization with frequency domain logarithmic compression, and multiply the time normalization result with the frequency domain compression result element by element to obtain normalized data after time domain and frequency domain processing, so as to achieve synchronous normalization of time domain and frequency domain, eliminate non-stationary noise interference, retain key frequency characteristics, achieve suppression of non-stationary noise and enhancement of frequency degradation characteristics, and break through the expression limitations of single domain normalization methods.

[0060] 2. Construct probability distribution through dynamic clustering and exponential mapping, calculate the fluctuation entropy vector and splice it with the original normalized features to enhance the modeling ability of uncertain fluctuation patterns in the degradation process.

[0061] 3. Utilizing the residual structure and attention mechanism of the residual attention module, through skip connections and dynamic weight adjustment, we can strengthen degradation-related features, suppress redundant interference, and improve the model's ability to recognize degradation-sensitive areas.

[0062] 4. In the output layer, the state transfer matrix is ​​combined to simulate the evolution law of the health state, and the focal loss is adjusted by the number of category samples to ensure that the classification results have temporal continuity and category balance. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 This is a flow chart of a method for evaluating the health status of an energy storage system provided in an embodiment of the present application.

[0065] Figure 2 This is a diagram showing the influence of different normalization methods on the discrimination of health status features provided in an embodiment of the present application.

[0066] Figure 3 This is a diagram showing the impact of feature enhancement on early degradation detection capability provided by an embodiment of the present application.

[0067] Figure 4This is a training effect diagram of an adaptive learning rate optimization strategy provided in an embodiment of the present application.

[0068] Figure 5 This is a schematic diagram of the classification accuracy of different models under various health conditions provided in an embodiment of the present application.

[0069] Figure 6 This is an internal structure diagram of an energy storage system health status assessment system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0070] It should be understood by those skilled in the art that the embodiments described below are merely preferred embodiments of the present disclosure and do not imply that the present disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely intended to explain the technical principles of the present disclosure and are not intended to limit the scope of protection of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present disclosure.

[0071] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0072] The technical solutions proposed in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0073] The embodiment provides a method for evaluating the health status of an energy storage system. Figure 1 As shown, the method provided in the embodiment of the present application mainly includes the following steps:

[0074] Step 110: Collect original monitoring data of the energy storage system in real time, and store the original monitoring data according to a preset time window.

[0075] It should be noted that the raw monitoring data is collected in real time through the sensor network deployed on the key components of the energy storage system;

[0076] Data sources include the voltage sensor, current sensor, and temperature sensor of the battery cell / module, as well as the charge and discharge status information recorded by the battery management system. The current sensor can be a Hall effect sensor, and the temperature sensor can be a thermocouple sensor.

[0077] The acquisition mode adopts synchronous high-frequency sampling above 1 kHz to ensure the capture of transient fluctuation characteristics of current and other parameters.

[0078] In one embodiment, taking current data as an example, the real-time current sequence during the charging and discharging process is continuously collected by a current sensor connected in series in the battery circuit, including steady-state working current, transient impact current and abnormal current.

[0079] Meanwhile, the collected data is classified and labeled, and according to the degradation indicators such as battery capacity attenuation and internal resistance growth, the health state is divided into “normal”, “mild degradation”, “moderate degradation”, “severe degradation” and “failure” by artificial labeling. The labeling information comes from the calibration results of laboratory accelerated aging test and the fault diagnosis records in actual operation and maintenance;

[0080] All original monitoring data are stored according to a preset time window to provide basic input for subsequent time-frequency normalization processing.

[0081] In step 120, the original monitoring data in the preset time window are processed in time scale to obtain time normalization results, the original monitoring data are processed in frequency domain to obtain frequency compression results, and the normalized data after time domain and frequency domain processing are calculated and obtained through the time normalization results and the frequency compression results.

[0082] Specifically, for the original sensor data, first, the mean and standard deviation in the preset time window are calculated in the time dimension, and the data at the current time is centralized and scaled based on this, the time scale is unified, then the original data is mapped to the frequency domain space through the frequency domain transformation method to obtain the spectrum amplitude, and then the amplitude dynamic range is compressed using the logarithmic function, then the time normalization results and the frequency compression results are multiplied element by element to obtain the normalized data after time domain and frequency domain processing, to realize the synchronous normalization of time domain and frequency domain, which not only eliminates the non-stationary noise interference, but also retains the key frequency characteristics, which is expressed as:

[0083] ,

[0084] In the formula, is the normalized data after time domain and frequency domain processing at the t time;

[0085] is the original sensor data vector at the t time, such as the current data vector collected by the original current sensor at the t time;

[0086] is the time domain mean, and the calculation method is expressed as ;

[0087] is the size of the data time window, representing the time length of the data;

[0088] is the original sensor data vector at the time point within the data time window;

[0089] is the time domain standard deviation, and the calculation method is represented as ;

[0090] is the logarithmic function, and the default base is 10;

[0091] represents the fast Fourier transform, which converts the time domain signal into the frequency domain;

[0092] represents the Hadamard product.

[0093] It should be noted that the energy storage sensor data is affected by the working condition fluctuation, and the time domain statistics need to be unified in the dynamic range to suppress the transient noise. The time domain mean and the time domain standard deviation eliminate the time domain amplitude offset and the scale difference by calculating the mean and standard deviation of the data in the sliding window.

[0094] It should also be noted that the battery aging causes the amplitude of the characteristic frequency to attenuate, such as the internal resistance characteristic frequency band. The present application amplifies the weak amplitude signal by the logarithmic function term , takes the logarithm of the frequency spectrum amplitude after the Fourier transform, compresses the frequency energy of the high dynamic range, and retains the key frequency domain features related to degradation.

[0095] It should also be noted that in the calculation process of the normalized data after the time domain and frequency domain processing at the t time, the conventional simple splicing method is not used, but the Hadamard product calculation is used to realize the coupling of the time domain and frequency domain features, suppress the time domain non-stationary noise, and enhance the frequency domain attenuation features, solving the problem of ignoring the frequency energy distribution in the conventional normalization.

[0096] In one embodiment, as Figure 2The box plot results show that the traditional normalization method has significant overlap in the "mild degradation" to "severe degradation" state interval, the feature value distribution range is wide and the boundary is fuzzy, indicating that it is difficult to effectively distinguish adjacent degradation states. The feature value box of the method of the application is compact under each health state, the median position rises in steps, and an obvious gap appears between the "severe degradation" and "failure" states, indicating that the time-frequency dual-domain processing mechanism of the application eliminates transient noise interference in the time domain, enhances weak degradation characteristics in the frequency domain, and realizes time-frequency feature collaborative enhancement through Hadamard product fusion, so that the feature distribution of different degradation states is essentially separated, laying a foundation for subsequent accurate classification.

[0097] Step 130, based on the normalized data processed in the time domain and the frequency domain, the fluctuation characteristics of the time series are modeled by combining the fluctuation entropy mechanism, and the features of the enhanced monitoring data are calculated.

[0098] This step can be specifically: based on the normalized data processed in the time domain and the frequency domain, the fluctuation characteristics of the time series are modeled by combining the fluctuation entropy mechanism, specifically, first extract the data at different time points through a sliding time window, and calculate the distance between the data and a plurality of center vectors generated in advance, then, according to the distances, the probability distribution of the data belonging to each center is constructed by exponential mapping, then, according to the probability distribution, the entropy value is calculated, which represents the uncertainty intensity of the fluctuation, then, the original normalized data and the corresponding fluctuation entropy vector are spliced to obtain the enhanced features, so as to enhance the difference expression of the degradation trend related in the data, represented as:

[0099] ,

[0100] ,

[0101] ,

[0102] In the formula, is the enhanced feature (the feature of the enhanced monitoring data), which is enhanced by combining the original data and the fluctuation entropy;

[0103] is the normalized data processed in the time domain and the frequency domain;

[0104] is a splicing operation;

[0105] is the fluctuation entropy vector, quantifying fluctuation uncertainty based on entropy, and entropy increase indicating abnormal fluctuation, realizing direct correlation with health state degradation;

[0106] is the fluctuation entropy of the th class;

[0107] is the length of the time window;

[0108] is the probability that the data belongs to the th class;

[0109] denotes the logarithm function with base of natural constant;

[0110] denotes the exponential function with base of natural constant;

[0111] is the fluctuation sensitivity factor, controlling the sensitivity of probability distribution to distance, e.g., set to 0.5;

[0112] is the cluster center of the th class, generated by DBSCAN clustering, representing typical fluctuation pattern, avoiding bias caused by preset fixed center;

[0113] is the number of dynamic cluster centers, determined by DBSCAN clustering method.

[0114] It should be noted that, denotes the fluctuation entropy of the th class, which is the entropy value calculated based on the distance between input data and dynamic cluster center, while is the entire fluctuation entropy vector, containing fluctuation entropy values of all classes.

[0115] It should be noted that, in the calculation process of the enhanced feature , instead of using the conventional simple splicing method, the entropy vector is spliced as a new feature with the original data, and the entropy vector provides global statistical features of degradation trend, which is complementary to local time series features, and together enhances the sensitivity of the model to early weak degradation.

[0116] In one embodiment, as Figure 3As shown, in order to analyze the promotion effect of fluctuation entropy characteristics on early degradation detection capability, through simulation of accelerated attenuation experiment, it is assumed that the battery reaches the unusable degree after 1000 cycles, and through the comparison curve of battery cycle number and detection degradation degree, it is shown that in the early degradation stage, the detection value of the fluctuation entropy characteristic enhancement method is obviously lower than the actual degradation curve, and there is a serious lag phenomenon, while the detection trajectory of the method is closely fitted to the actual degradation process. In the serious degradation stage, the difference between the two methods is reduced, and the test results show that by calculating the probability distribution entropy value of the data points and the dynamic clustering center, the fluctuation uncertainty of the sensor data is quantified, the entropy vector is spliced and fused with the original features, so that the model has local detail perception and global fluctuation trend judgment ability, and the sensitivity of the model to the initial weak degradation signal is significantly improved.

[0117] Step 140, constructing a residual attention deep neural network composed of an input layer, a plurality of residual attention modules, a hybrid regularization layer, and a state output layer stacked in turn; by monitoring the variance change of each round of parameter gradient in the residual attention deep neural network training process, the current adaptive learning rate is adjusted automatically.

[0118] Among them, the residual attention deep neural network composed of an input layer, a plurality of residual attention modules, a hybrid regularization layer, and a state output layer stacked in turn, specifically includes:

[0119] The input layer receives the features of the enhanced monitoring data, and maps the features of the enhanced monitoring data to the initial high-dimensional feature space of the network;

[0120] The residual path in the residual attention module extracts a preset general feature from the features of the enhanced monitoring data in the high-dimensional feature space, and the attention path dynamically adjusts the feature weight; through the preset general feature, the enhanced monitoring data feature, the self-attention weight, and the Sigmoid activation function, the feature processed by the residual attention is obtained;

[0121] The hybrid regularization layer randomly discards a preset portion of the residual attention module output with a preset probability to obtain a first feature, adopts a spectral normalization mechanism to perform spectral normalization processing on the residual attention module output to obtain a second feature; and obtains the feature after hybrid regularization processing of the first feature and the second feature;

[0122] The state output layer obtains the basic probability of each health state category through a conventional classification function;

[0123] Based on the number difference between each health state category, a state transition matrix is constructed;

[0124] According to the basic probability, the feature after hybrid regularization processing, and the state transition matrix, the health state probability is calculated.

[0125] To further illustrate, conventional deep neural networks are insufficient in modeling long-term dependencies and tend to ignore differences in sensor degradation contributions. This application uses an attention residual module in the deep neural network structure. For the input features of each layer, its feature representation is first extracted through a residual path, which includes two layers of fully connected networks and activation functions. At the same time, a set of attention weights is used to calculate the importance of different parts of the input features, and nonlinear mapping is performed through the activation function. Then, the attention-weighted features are element-wise multiplied by the residual path results and jump-connected with the original input to obtain features processed by residual attention, so as to enhance the propagation effect of important features and suppress redundant interference of irrelevant features, which is expressed as:

[0126] ;

[0127] ;

[0128] Where, is the feature after residual attention processing;

[0129] is the input feature of the residual attention deep neural network. For the first layer input of the residual attention deep neural network, That is ;

[0130] is the residual path, containing two fully connected layers;

[0131] is the Sigmoid activation function;

[0132] is the self-attention weight;

[0133] is the ReLU activation function;

[0134] is the first weight matrix, which is a trainable parameter;

[0135] is the first bias term, which is a trainable parameter;

[0136] is the second bias term, which is a trainable parameter;

[0137] is the second weight matrix, which is a trainable parameter.

[0138] Since small sample training is prone to overfitting, the Dropout of the conventional deep neural network is easy to destroy the time sequence correlation; in the deep network training process, a mixed method is adopted to suppress the overfitting risk, specifically, first, a part of the residual path output is randomly discarded with a preset probability, avoiding the model from relying on fixed features, thereby enhancing the generalization ability, then, a spectral normalization mechanism is adopted to constrain the amplitude of the weight parameters in the residual module, preventing some weights from abnormally growing due to overfitting in the training process, and finally the features after mixed regularization processing are obtained by adding the two parts, which takes into account the feature stability and training robustness, and is expressed as:

[0139] ;

[0140] ;

[0141] In the formula, is the feature after mixed regularization processing;

[0142] indicates that the residual path is randomly discarded with a preset probability, for example, the preset probability is 5%;

[0143] is the spectral constraint strength, for example, it is set to 0.01;

[0144] is the spectral normalization constraint;

[0145] is the L2 norm;

[0146] is the weight matrix of the first layer of the residual attention deep neural network;

[0147] indicates the maximum value function;

[0148] Since the conventional deep neural network adopts Softmax to output classification probability, ignoring the state transition probability, in the output end, a state transition matrix is used to model the classification result in Markov, specifically, the output of the last layer of the network obtains the basic probability of each health state through the conventional classification function, then, based on the number difference between states, a state transition matrix is constructed to simulate the transition trend between adjacent states, and the final output of the model is the product of the basic classification result and the state transition matrix, so that the prediction result has smooth continuity in the time sequence, and is expressed as:

[0149] ;

[0150] ;

[0151] ​wherein, is a Softmax function; is a weight matrix of the last layer of the network, and linearly maps to the health state category space;

[0152] is a health state probability vector after Markov output layer processing, and the classification category is determined by the category corresponding to the maximum probability value in the health state probability vector;

[0153] is a state transition matrix, and the transition probability is given based on the state number distance, and in view of the continuity of the energy storage system monitoring data, the adjacent state transition probability is high, and the jump state probability is low, and the prediction result is forced to comply with the degradation continuity;

[0154] is a transition probability from the i th health state to the j th health state;

[0155] is a transition decay factor, such as being set to 0.5;

[0156] represents the state number difference between the i th state and the j th state;

[0157] represents the state number difference between the i th state and the k th state;

[0158] C is a health state classification category of the energy storage system.

[0159] In the step, the variance of the parameter gradient of each round in the monitoring residual attention deep neural network training process is monitored, and the current adaptive learning rate is automatically adjusted, and specifically includes:

[0160] The gradient variance of the current round is calculated by ratio with the gradient variance of the last round; wherein, the current round is greater than 1;

[0161] If the current gradient variance is large, the adaptive learning rate is relaxed; otherwise, the adaptive learning rate is gradually reduced; wherein, the change range of the learning rate relaxation / reduction is realized by an exponential decay mechanism.

[0162] Specifically, conventional deep neural networks usually use a fixed learning rate, which is difficult to adapt to the staged convergence characteristics of health status classification. The present application monitors the variance changes of the parameter gradients in each round during network training and autonomously adjusts the current learning rate. Specifically, the ratio of the gradient variance of the current round to the gradient variance of the previous round is calculated. If the current gradient variance is large, the learning rate is relaxed to jump out of the local optimum. Otherwise, the learning rate is gradually reduced to promote convergence stability. The change of the learning rate is achieved through an exponential decay mechanism to control its change rate, which is expressed as:

[0163] ;

[0164] ;

[0165] Where, e is the training round;

[0166] is the adaptive learning rate of the e-th round;

[0167] is the initial learning rate, for example, set to 0.001;

[0168] is the learning rate attenuation coefficient, for example, set to 0.1;

[0169] is the parameter gradient vector of the e-th round;

[0170] For the wheel parameter gradient vector;

[0171] is the variance calculation function;

[0172] are the trainable parameters of the residual attention deep neural network;

[0173] Indicates the parameter quantity;

[0174] For the The gradient of the trainable parameters;

[0175] For the The gradient of a trainable parameter.

[0176] It should be noted that the degradation of health status occurs in stages. Characterizes the gradient variance and is used to monitor the stability of parameter updates. When the gradient variance is large, increasing the learning rate will help escape the local optimum.

[0177] In one embodiment, Figure 4As shown in the figure, the training process of the fixed learning rate and adaptive learning rate strategies is compared, and a double vertical axis line chart is used to simultaneously display the change in loss value and the learning rate adjustment trajectory. The fixed learning rate strategy falls into a plateau in the middle term, and the loss value stagnates at a high level. The adaptive strategy breaks through the platform by increasing the learning rate in the same round, and gradually converges to a lower loss level in the later stage. The learning rate curve shows that the adaptive mechanism maintains a high learning rate to accelerate convergence in the early stage of training, dynamically adjusts the learning rate according to the gradient variance in the middle term to jump out of the local optimum, and automatically decays in the final stage to ensure stable convergence. In the training of the health status assessment model, it is necessary to quickly approach the global optimal solution and avoid oscillation in the later stage, which significantly improves the model convergence efficiency and generalization ability.

[0178] Step 150: Output the health status probability of each health status category through the residual attention deep neural network; calculate the dynamic focus loss based on the health status probability of each health status category and the preset loss weight corresponding to each category, obtain the dynamic focus loss, temporal continuity loss, and attention sparse regularization loss, and calculate the total loss function.

[0179] Specifically, to avoid the majority class-dominated loss caused by sample imbalance, this application calculates the number of samples for each category and constructs corresponding weights based on this value to ensure that rare categories occupy a higher proportion in the loss function. At the same time, the focus factor of each category is dynamically adjusted according to the number of samples, so that prediction errors of rare categories are more severely penalized. The dynamic focus loss function can effectively improve the model's ability to recognize difficult samples by focusing on the probability output of each category, which is expressed as:

[0180] ;

[0181] ;

[0182] ;

[0183] Where, is the dynamic focus loss function;

[0184] is the loss weight of category c;

[0185] is the classification probability of the cth class;

[0186] Dynamically focus on the parameters and adjust the loss weight of difficult samples;

[0187] is the total sample number;

[0188] is the number of samples in the cth category;

[0189] For preset constants, for example, ;

[0190] Represents the median of the number of samples in each category.

[0191] Calculate the total loss function:

[0192] In order to collaboratively optimize the accuracy of health status classification and the consistency of time series prediction, this application adopts a multi-task coupling loss mechanism. The total loss function consists of three parts: dynamic focus classification loss function Solve the problem of sample imbalance; time series continuity loss function Constrain the prediction results of adjacent time windows to conform to the Markov transition law; attention sparse regularization loss function Suppress excessive dispersion of attention weights and improve feature selection focus.

[0193] Specifically, the total loss function is defined as a weighted summation:

[0194]

[0195] Where, is the total loss function;

[0196] is the first loss weight, for example, set to 0.6;

[0197] is the second loss weight, for example, set to 0.2;

[0198] is the third loss weight, for example, set to 0.2.

[0199] The temporal continuity loss function is obtained by calculating the KL divergence of the state probabilities of adjacent time windows. The Markov transfer matrix is ​​back-propagated to the training process to force the network to learn the continuity of state evolution. The calculation method is expressed as:

[0200] ;

[0201] Where, is the KL divergence calculation function;

[0202] Indicates the State probability vector of the time window;

[0203] is the state transition matrix.

[0204] The attention sparsity regular loss function calculates the negative entropy of the attention weight to enhance sparsity, and restricts the attention mechanism to avoid redundant activation, and the calculation method is represented as:

[0205] ;

[0206] In the formula, is the number of network layers of the residual attention deep neural network;

[0207] is the number of attention heads;

[0208] is the weight parameter of the e-th attention head of the i-th layer of the residual attention deep neural network.

[0209] Step 160, based on the total loss function and the adaptive learning rate, update the trainable parameters of the iteration; when the preset stopping iteration condition is met, the trained residual attention deep neural network is obtained.

[0210] Based on the total loss function and the adaptive learning rate, the trainable parameters of the iteration are updated, which can be specifically:

[0211] Based on the total loss function and the adaptive learning rate, the network parameters are updated by using the gradient clipping method. First, the gradient of the trainable parameters of the current round is calculated, and the L2 norm of the gradient is clipped to constrain its maximum amplitude not to exceed the preset threshold. Then, the updated momentum term is used to update the momentum term, and the weight decay term is added to suppress overfitting. Then, based on the updated momentum term, the trainable parameters of the next round of iteration are generated, and the specific steps are as follows:

[0212] First, the gradient of the trainable parameter is calculated, which is represented as:

[0213] ;

[0214] In the formula, is the partial derivative symbol;

[0215] is the trainable parameter of the residual attention deep neural network of the e-th iteration.

[0216] Further, the gradient norm is constrained by using the gradient clipping method to prevent the gradient explosion phenomenon in the training process, which is represented as:

[0217] ;

[0218] , is the clipping threshold, such as 5.0;

[0219] ​​ is the gradient of the trainable parameters after clipping;

[0220] is the L2 norm.

[0221] Furthermore, the momentum term is updated based on the gradient of the pruned trainable parameters, which is expressed as:

[0222] ;

[0223] Where, For the The momentum term of the round iteration is used to smooth the parameter update of the current round;

[0224] For the The momentum term of the round iteration;

[0225] is the momentum coefficient, for example, set to 0.9;

[0226] is the weight decay strength, e.g., set to .

[0227] Furthermore, the trainable parameters of the residual attention deep neural network are updated based on the momentum term, which is expressed as:

[0228] ;

[0229] Where, For the Trainable parameters of the iterative residual attention deep neural network.

[0230] When the preset stop iteration condition is met, the trained residual attention deep neural network is obtained, which can be specifically:

[0231] During the training process, the loss change on the validation set is continuously monitored. If the current training round reaches the preset maximum training round limit, the iteration is stopped immediately.

[0232] At the same time, an early stopping mechanism is used for dynamic judgment. That is, when the value of the total loss function does not show a significant decrease within the consecutive set rounds, which is manifested as not falling below the historical optimal verification loss by more than the preset threshold, or the adaptive learning rate has decayed to the preset minimum learning rate lower limit, the training process is also terminated.

[0233] Step 170: Obtain current monitoring data, calculate the enhanced features of the current monitoring data, and then use the trained residual attention deep neural network to evaluate the health status of the energy storage system.

[0234] Specifically, after the residual attention deep neural network is trained, real-time health state classification is performed according to the following process:

[0235] 1) For the input sensor data stream (current monitoring data), time-frequency dual-domain normalization is performed, and is outputted;

[0236] 2) The fluctuation entropy vector is calculated, and is spliced to obtain ;

[0237] 3) The is inputted into the residual attention deep neural network, sequentially passes through the residual attention module and the mixed regularization layer, and the feature after the mixed regularization processing is outputted ;

[0238] 4) The health state probability vector is calculated through the health state probabilistic output layer, and the index with the maximum probability is taken as the predicted state.

[0239] In an embodiment, as shown in Figure 5 , to verify the performance advantage of the residual attention deep neural network compared with a conventional model, the comparison objects include a support vector machine, a random forest, a long short-term memory network and a convolutional neural network, and the column chart clearly shows that the conventional model has a significantly low accuracy rate in the “mild degradation” and “moderate degradation” states, indicating that it is difficult to capture early weak degradation features, the conventional deep learning model is improved in the middle and late stages, but still has missed judgment for early degradation, and the method of the present application maintains a high accuracy rate in the five health states, especially in the early degradation stage, the experimental results show that the residual attention module strengthens the degradation sensitive features through dynamic adjustment of feature weights, the Markov output layer constrains the state transition continuity, and the dynamic focal loss function balances the sample imbalance problem, which solves the defects of the conventional model in modeling the gradual degradation.

[0240] In addition, the present application Figure 6 provides a kind of energy storage system health state evaluation system for the embodiment of the present application.As shown in Figure 6 , the system provided by the embodiment of the present application mainly includes:

[0241] The acquisition module 210 is used to collect the original monitoring data of the energy storage system in real time, and store the original monitoring data according to the preset time window;

[0242] The normalization calculation module 220 is configured to perform time scale uniform processing on the original monitoring data in a preset time window to obtain time normalization results, perform frequency domain processing on the original monitoring data to obtain frequency domain compression results, and calculate normalized data after time domain and frequency domain processing based on the time normalization results and the frequency domain compression results.

[0243] The enhancement calculation module 230 is configured to model fluctuation characteristics of a time sequence based on the normalized data after time domain and frequency domain processing, in combination with a fluctuation entropy mechanism, and calculate features of enhanced monitoring data.

[0244] The learning rate adjustment module 240 is configured to construct a residual attention deep neural network composed of an input layer, a plurality of residual attention modules, a hybrid regularization layer, and a state output layer stacked in sequence, and autonomously adjust a current adaptive learning rate by monitoring variance of a parameter gradient of each round in a training process of the residual attention deep neural network.

[0245] The loss calculation module 250 is configured to output health state probabilities of each health state category through the residual attention deep neural network, calculate a dynamic focus loss based on the health state probabilities of each health state category and preset loss weights corresponding to the categories, and obtain a total loss function based on the dynamic focus loss, a time sequence continuity loss, and an attention sparsity regularization loss.

[0246] The iterative training module 260 is configured to update trainable parameters of iterations based on the total loss function and the adaptive learning rate, and obtain a trained residual attention deep neural network when a preset stopping iteration condition is met.

[0247] The real-time evaluation module 270 is configured to obtain current monitoring data, calculate features of enhanced current monitoring data, and then perform health state evaluation of the energy storage system by using the trained residual attention deep neural network.

[0248] So far, the technical solutions of the present disclosure have been described in combination with the foregoing embodiments, but those skilled in the art can easily understand that the protection scope of the present disclosure is not limited to these specific embodiments. Those skilled in the art can split and combine the technical solutions in the above-described embodiments, or make equivalent changes or replacements to related technical features, without deviating from the technical principles of the present disclosure. Any changes, equivalent replacements, improvements, etc. made within the technical concept and / or technical principles of the present disclosure will fall within the protection scope of the present disclosure.

Claims

1. A method for evaluating the health status of an energy storage system, characterized in that: The method comprises: Collect the original monitoring data of the energy storage system in real time and store the original monitoring data according to the preset time window; The original monitoring data within the preset time window is processed in a unified time scale to obtain a time normalization result, and the original monitoring data is processed in the frequency domain to obtain a frequency domain compression result; the normalized data after time domain and frequency domain processing is calculated based on the time normalization result and the frequency domain compression result; Based on the normalized data after time domain and frequency domain processing, the fluctuation characteristics of the time series are modeled in combination with the fluctuation entropy mechanism to calculate the characteristics of the enhanced monitoring data; Construct a residual attention deep neural network consisting of an input layer, several residual attention modules, a hybrid regularization layer, and a state output layer stacked in sequence; by monitoring the variance changes of the parameter gradients in each round of residual attention deep neural network training, the current adaptive learning rate is automatically adjusted; Through the residual attention deep neural network, the health state probability of each health state category is output; according to the health state probability of each health state category and the preset loss weight corresponding to each category, the dynamic focus loss is calculated; the dynamic focus loss, temporal continuity loss, and attention sparse regularization loss are obtained to calculate the total loss function; wherein, the dynamic focus loss is calculated, specifically including: By formula: , calculate the dynamic focusing parameters of category c ;in, represents the number of samples in the cth category, c∈[1,C], C represents the total number of health status categories, represents the median number of samples in each category; By formula: , calculate the loss weight of category c ;in, represents the total number of samples, Indicates a preset constant; By formula: , calculate the dynamic focus loss ;in, represents the health status probability of category c; The attention sparse regularization loss function calculates the negative entropy of the attention weight to enhance sparsity and constrain the attention mechanism to avoid redundant activation. The calculation method is expressed as: Where, is the number of network layers of the residual attention deep neural network; is the number of attention heads; The residual attention deep neural network Tier The weight parameters of the attention heads; Based on the total loss function and adaptive learning rate, the iterative trainable parameters are updated; when the preset stopping conditions are met, the trained residual attention deep neural network is obtained; The current monitoring data is obtained, the features of the enhanced current monitoring data are calculated, and then the trained residual attention deep neural network is used to evaluate the health status of the energy storage system.

2. The method for evaluating the health status of an energy storage system according to claim 1, wherein: Real-time collection of raw monitoring data from the energy storage system, including: Real-time collection of raw monitoring data through a sensor network deployed on key components of the energy storage system; The data sources of the original monitoring data include the voltage sensor, current sensor, temperature sensor of the battery cell / module, and the charge and discharge status information recorded by the battery management system; The original monitoring data is collected using synchronous high-frequency sampling above 1 kHz.

3. The method for evaluating the health status of an energy storage system according to claim 1, wherein: The original monitoring data within the preset time window is processed in a unified time scale to obtain a time normalization result, and the original monitoring data is processed in the frequency domain to obtain a frequency domain compression result; The normalized data after time domain and frequency domain processing are calculated based on the time normalization results and frequency domain compression results, including: In the time dimension, the mean and standard deviation of the original monitoring data within the preset time window are calculated, and based on this, the data at the current moment is centered and scaled to unify the time scale and obtain the time normalization result; The time normalization result is mapped to the frequency domain space through the frequency domain transformation method to obtain the spectrum amplitude; the spectrum amplitude is compressed to the preset amplitude dynamic range using the logarithmic function; The time normalization result is multiplied element by element by the frequency domain compression result to obtain the normalized data after time domain and frequency domain processing.

4. The method for evaluating the health status of an energy storage system according to claim 1, wherein: Based on the normalized data after time domain and frequency domain processing, the fluctuation characteristics of the time series are modeled in combination with the fluctuation entropy mechanism, and the enhanced monitoring data features are calculated, including: Extracting normalized data at different time points through a sliding time window and calculating the distance between the extracted data and several pre-generated center vectors; According to the distance, the exponential mapping is used to construct the probability that the normalized data belongs to each center, and the fluctuation entropy vector corresponding to the normalized data is calculated based on the probability of each center; The normalized data is concatenated with the corresponding fluctuation entropy vector to obtain the enhanced monitoring data features.

5. The method for evaluating the health status of an energy storage system according to claim 1, wherein: Construct a residual attention deep neural network consisting of an input layer, several residual attention modules, a hybrid regularization layer, and a state output layer stacked in sequence, specifically including: The input layer receives the features of the enhanced monitoring data and maps the features of the enhanced monitoring data to the initial high-dimensional feature space of the network; The residual path in the residual attention module extracts preset common features from the features of the enhanced monitoring data in the high-dimensional feature space, and the attention path dynamically adjusts the feature weights; through the preset common features, the features of the enhanced monitoring data, the self-attention weights, and the Sigmoid activation function, the features processed by the residual attention are obtained; The hybrid regularization layer randomly discards a preset part of the residual attention module output with a preset probability to obtain the first feature, and adopts the spectral normalization mechanism to perform spectral normalization on the residual attention module output to obtain the second feature; the feature after the hybrid regularization of the first feature and the second feature is obtained; The state output layer obtains the basic probability of each health state category through the conventional classification function; Based on the number differences between each health state category, a state transition matrix is ​​constructed; The health state probability is calculated based on the basic probability, the features after hybrid regularization, and the state transition matrix.

6. The method for evaluating the health status of an energy storage system according to claim 1, wherein: By monitoring the variance changes of parameter gradients in each round of residual attention deep neural network training, the current adaptive learning rate is adjusted autonomously, including: Calculate the ratio of the gradient variance of the current round to the gradient variance of the previous round; where the current round is greater than 1; If the current gradient variance is large, the adaptive learning rate is relaxed; otherwise, the adaptive learning rate is gradually reduced; the relaxation / reduction of the learning rate is achieved through an exponential decay mechanism.

7. The method for evaluating the health status of an energy storage system according to claim 1, wherein: Obtain dynamic focus loss, temporal continuity loss, attention sparse regularization loss, and calculate the total loss function, including: Obtain the corresponding weight values ​​of dynamic focus loss, temporal continuity loss, and attention sparse regularization loss, and then weight them to obtain the total loss function.

8. The method for evaluating the health status of an energy storage system according to claim 1, wherein: Based on the total loss function and adaptive learning rate, the trainable parameters of the update iteration are as follows: Calculate the gradient of the trainable parameters using the total loss function and the trainable parameters of the current round; Adopt gradient clipping to clip the gradient of trainable parameters; Use the clipped gradient and adaptive learning rate to update the momentum term, and add a preset weight decay term to suppress overfitting; Generate the trainable parameters for the next iteration based on the current momentum term, the updated momentum term, the adaptive learning rate, the current round of trainable parameters, and the gradient of the pruned trainable parameters.

9. An energy storage system health status assessment system, characterized in that: The system comprises: The acquisition module is used to collect the original monitoring data of the energy storage system in real time and store the original monitoring data according to the preset time window; The normalization calculation module is used to perform time-scale uniform processing on the original monitoring data within the preset time window to obtain a time normalization result, and perform frequency domain processing on the original monitoring data to obtain a frequency domain compression result; the normalized data after time domain and frequency domain processing is calculated based on the time normalization result and the frequency domain compression result; The enhanced computing module is used to model the fluctuation characteristics of the time series based on the normalized data processed in the time and frequency domains, combined with the fluctuation entropy mechanism, and calculate the characteristics of the enhanced monitoring data; The learning rate adjustment module is used to construct a residual attention deep neural network composed of an input layer, several residual attention modules, a hybrid regularization layer, and a state output layer. By monitoring the variance changes of the parameter gradients in each round of residual attention deep neural network training, the current adaptive learning rate is automatically adjusted. The loss calculation module is used to output the health status probability of each health status category through the residual attention deep neural network; calculate the dynamic focus loss based on the health status probability of each health status category and the preset loss weight corresponding to each category; obtain the dynamic focus loss, temporal continuity loss, and attention sparse regularization loss to calculate the total loss function; The dynamic focus loss is calculated, specifically including: By formula: , calculate the dynamic focusing parameters of category c ;in, represents the number of samples in the cth category, c∈[1,C], C represents the total number of health status categories, represents the median number of samples in each category; By formula: , calculate the loss weight of category c ;in, represents the total number of samples, Indicates a preset constant; By formula: , calculate the dynamic focus loss ;in, represents the health status probability of category c; The attention sparse regularization loss function calculates the negative entropy of the attention weight to enhance sparsity and constrain the attention mechanism to avoid redundant activation. The calculation method is expressed as: Where, is the number of network layers of the residual attention deep neural network; is the number of attention heads; The residual attention deep neural network Tier The weight parameters of the attention heads; The iterative training module is used to update the iterative trainable parameters based on the total loss function and adaptive learning rate; when the preset stopping condition is met, the trained residual attention deep neural network is obtained; The real-time evaluation module is used to obtain current monitoring data, calculate the enhanced features of the current monitoring data, and then use the trained residual attention deep neural network to evaluate the health status of the energy storage system.

Images