A distributed energy storage system state evaluation method

Abstract

The application discloses a distributed energy storage system state evaluation method, mainly relates to the technical field of state evaluation, in order to solve the existing scheme exists key state characteristics are covered or lost, state sample is unevenly distributed in feature space, the problem of insufficient adaptability to sparse features and boundary samples. Including: obtaining fusion feature vector under different operating conditions, calculating enhanced feature vector; obtaining improved GELU activation function; calculating the class balance focusing loss function corresponding to each state label data class in the neural network; the local mode feature and the global correlation branch feature vector are mapped into the gate weight, and the gate fusion feature vector is obtained; the gate constraint loss function of neural network is constructed; using the improved GELU activation function for nonlinear transformation, outputting the predicted state class probability distribution; the total loss is weighted calculated through the class balance focusing loss function and the gate constraint loss function; taking the minimum total loss as the target.

Description

Technical Field

[0001] This application relates to the field of system state assessment technology, and in particular to a method for assessing the state of a distributed energy storage system. Background Technology

[0002] With the development of new energy technologies and intelligent power systems, distributed energy storage systems have been widely used as a key component for improving grid regulation capabilities and ensuring energy supply and demand balance. However, due to the data acquisition and processing involving multiple sensors, physical quantities, and measurement points, accurate assessment of the system's operating status faces the following challenges:

[0003] First, multi-source sensor data is highly heterogeneous. There are significant differences in dimensions, inconsistent noise levels, and nonlinear coupling relationships among sensors such as voltage, current, temperature, and impedance. Conventional data fusion methods, such as simple stitching or weighted averaging, cannot effectively integrate this information, which can easily lead to the masking or loss of key state features.

[0004] Secondly, the distribution of state samples in the feature space is uneven. Normal operating condition data is abundant, while fault state data, especially early fault symptoms, is relatively scarce, particularly in the state boundary regions where the distribution is sparse. This makes it difficult for the model to accurately define category boundaries, affecting the assessment accuracy.

[0005] Furthermore, traditional neural network models are not well-suited to sparse features and boundary samples. Activation functions such as ReLU have static saturation regions, which can easily lead to gradient vanishing and make it difficult to identify weak fault signals. At the same time, fixed-structure feature extraction networks ignore the distribution characteristics of samples in the feature space and lack the ability to adapt to local patterns and global dependencies. Summary of the Invention

[0006] This application provides a method, system, and medium for state assessment of distributed energy storage systems to address the problems of conventional data fusion methods, such as simple splicing or weighted averaging, which cannot effectively integrate this information and are prone to masking or losing key state features; uneven distribution of state samples in the feature space, with sparse distribution in the state boundary region, making it difficult for the model to accurately delineate category boundaries and affecting assessment accuracy; insufficient adaptability of traditional neural network models to sparse features and boundary samples; and fixed-structure feature extraction networks that ignore the distribution characteristics of samples in the feature space and lack the ability to adapt to local patterns and global dependencies.

[0007] Firstly, this application provides a method for assessing the state of a distributed energy storage system, the method comprising:

[0008] Several types of sensors are deployed at each preset key point in the distributed energy storage system to obtain sensor datasets under different operating conditions. The sensor dataset consists of several sensor data sequences, and the same sensor data sequence contains various types of sensor data at each preset key point at the same acquisition time. The state labeling data corresponding to the sensor data sequences is obtained. The sensor dataset is divided into training set and validation set.

[0009] The weight values ​​of each sensor data are quantified by calculating the information entropy of each sensor data in the training set; the mutual information between sensor data is quantified by calculating the mutual information between sensor data; adaptive weights are generated by combining the information entropy and the mutual information compensation factor; and the adaptive weights are used to perform weighted fusion of several types of sensor data in the same sensor data sequence in the training set to obtain a fused feature vector.

[0010] Multiple cluster structures are generated by density clustering of the fused feature vectors. Based on the distance from the fused feature vector to the cluster center and the distribution characteristics within the cluster, the local density value of each sensor data sequence is calculated. An enhancement coefficient is generated based on the local density value. The enhancement coefficient and the fused feature vector are then used to calculate the enhanced feature vector.

[0011] The improved GELU activation function of the neural network is obtained by multiplying the input value of the neural network neuron with the cumulative distribution function and the saturation characteristic; and the saturation characteristic is dynamically adjusted by combining the batch average gradient magnitude with the input value through the hyperbolic tangent function; the balance weights corresponding to each state labeled data category are constructed according to the amount of data of each state labeled data category in the fusion feature vector corresponding to the training set; and the class balance focusing loss function corresponding to each state labeled data category in the neural network is calculated using the balance weights and the prediction probability.

[0012] The neural network uses a lightweight convolution module to extract intra-cluster local pattern features corresponding to the enhanced feature vector. It employs a multi-head self-attention mechanism to output the global association branch feature vector corresponding to the enhanced feature vector. The intra-cluster local pattern features and the global association branch feature vector are mapped to gating weights through linear transformation and the sigmoid function, and then weighted and fused to obtain the gating fused feature vector.

[0013] By utilizing gate weights and local density values, a gated loss function for the neural network is constructed. The gated loss function is then input into the fully connected layer of the neural network, and a nonlinear transformation is performed using an improved GELU activation function to output the probability distribution of the predicted state category.

[0014] The total loss is calculated by weighting the loss function and the gated constraint loss function. The neural network is trained iteratively with the goal of minimizing the total loss. The neural network is then validated using a validation set until a well-trained neural network that meets the preset conditions is obtained.

[0015] In one implementation of this application, after obtaining a trained neural network that meets preset conditions, the method further includes:

[0016] Newly connected real-time sensor data streams are processed into sensor data sequences according to the acquisition time.

[0017] The weight values ​​of each sensor data are quantified by calculating the information entropy of each sensor data in the same sensor data sequence; the mutual information between sensor data is quantified by calculating the mutual information between sensor data in the same sensor data sequence; adaptive weights are generated by combining information entropy and mutual information compensation factors; and the adaptive weights are used to perform weighted fusion of several types of sensor data in the same sensor data sequence in the training set to obtain a fused feature vector.

[0018] Density clustering is performed on the fused feature vectors to generate multiple cluster structures. Based on the distance from the fused feature vectors to the cluster center and the distribution characteristics within the cluster, the local density value of each fused feature vector is calculated. An enhancement coefficient is generated based on the local density value. The enhancement coefficient and the fused feature vector are used to calculate the enhanced feature vector.

[0019] The enhanced feature vectors are input into the trained neural network to obtain the output predicted state category probability distribution.

[0020] In one implementation of this application, the weight values ​​of each sensor data point are quantified by calculating the information entropy of each sensor data point in the training set, specifically including:

[0021] The normalized proportion of each feature value is obtained by normalizing the absolute value of the feature value, and the information entropy is calculated based on the normalized proportion.

[0022] The weight values ​​of each sensor data are quantified based on the inverse relationship between information entropy and weight value.

[0023] In one implementation of this application, the mutual information between sensor data is calculated to quantify the mutual information compensation factor; an adaptive weight is generated by combining information entropy and the mutual information compensation factor; and the adaptive weight is used to perform weighted fusion of several classes of sensor data of the same sensor data sequence in the training set to obtain a fused feature vector, specifically including:

[0024] Through the formula:

[0025] Calculate the first Mutual information compensation factor for sensor data;

[0026] in, For the first Mutual information compensation factor for sensor data; For the first Data from individual sensors; For the first Data from individual sensors; To distinguish from Sensor data index; The number of sensor data points in a sensor data sequence; Features and Mutual information between them;

[0027] Through the formula:

[0028] Calculate the adaptive weights;

[0029] in, Indicates adaptive weights, Indicates the first Information entropy of sensor data;

[0030] Through the formula:

[0031] Calculate the fusion feature vector of each sensor data sequence;

[0032] in, This represents the fused feature vector.

[0033] In one implementation of this application, density clustering is performed on the fused feature vectors to generate multiple cluster structures. Based on the distance from the fused feature vectors to the cluster centers and the intra-cluster distribution characteristics, the local density value of each sensor data sequence is calculated, specifically including:

[0034] Through the formula:

[0035] Calculate the local density value of each fused feature vector;

[0036] in, Let be the local density value of the i-th sensor data sequence;

[0037] Let be the fused feature vector of the i-th sensor data sequence;

[0038] For clusters The standard deviation is calculated as follows: ;

[0039] The first generation of density clustering A cluster; Let be the fused feature vector of the j-th sensor data sequence; Cluster index for density clustering; For clusters The number included; For clusters The center vector; It is an L2 norm; It is a natural exponential function.

[0040] In one implementation of this application, an enhancement coefficient is generated based on local density values, and an enhanced feature vector is calculated using the enhancement coefficient and the fused feature vector, specifically including:

[0041] Through the formula:

[0042] Calculate the enhanced feature vector;

[0043] in, Let be the enhanced feature vector of the i-th sensor data sequence; The enhancement coefficient for the i-th sensor data sequence is expressed as follows:

[0044] ;

[0045] To enhance the intensity factor; For clusters The center vector.

[0046] In one implementation of this application, the input value of a neural network neuron is multiplied by the cumulative distribution function and the saturation characteristic to obtain an improved GELU activation function for the neural network; and the saturation characteristic is dynamically adjusted by combining the batch average gradient magnitude with the input value using the hyperbolic tangent function, specifically including:

[0047] Through the formula:

[0048] To obtain the improved GELU activation function for neural networks;

[0049] in, For the improved GELU activation function, Output values ​​for the improved GELU activation function;

[0050] Input values ​​for neural network neurons;

[0051] The cumulative distribution function of the standard normal distribution is calculated as follows:

[0052] ,in, is the normalization constant for the normal distribution; This represents the value from negative infinity to the current input value of a neuron in the neural network. Definite integral; dt is a natural constant; t is the integration variable; dt is an integral infinitesimal element.

[0053] The initial value for the saturation characteristic is set to 1;

[0054] The saturation characteristic function for dynamically adjusting saturation characteristics is expressed as:

[0055] ,in, The average gradient magnitude for the batch is calculated as follows: ;

[0056] This is the scaling factor; It is the hyperbolic tangent function; Batch size; The loss function of the i-th sensor data sequence applied to the input values ​​of the neural network neurons The gradient.

[0057] In one implementation of this application, a balanced weight is constructed for each state-labeled data category based on the amount of data in each state-labeled data category of the fusion feature vector corresponding to the training set; using the balanced weight and prediction probability, a class-balanced focusing loss function is calculated for each state-labeled data category in the neural network, specifically including:

[0058] Through the formula:

[0059] Construct balanced weights corresponding to the data categories labeled with each state;

[0060] In the formula, Let c be the balanced weight of the c-th category; c is the category index. c represents the total number of categories in the status label data. The number of sensor data sequences in the c-th category; As a balancing regulator;

[0061] Through the formula:

[0062] Calculate the equilibrium focusing loss function;

[0063] in, Indicates class-balanced focusing loss; Label the data category for the true state of the i-th sensor data sequence;

[0064] for Balanced weights for labeling the data categories of the true state of the i-th sensor data sequence;

[0065] Label the data categories for the true state of the i-th sensor data sequence to the model. The predicted probability;

[0066] This represents the prediction error term; As a focusing factor; It is a logarithmic function.

[0067] Secondly, this application provides a distributed energy storage system state assessment system, the system comprising:

[0068] The data collection module is used to deploy several types of sensors at various preset key points in the distributed energy storage system to obtain sensor datasets under different operating conditions. The sensor dataset consists of several sensor data sequences, and the same sensor data sequence contains various types of sensor data at each preset key point at the same acquisition time. The module also acquires the state labeling data corresponding to the sensor data sequences and divides the sensor dataset into training and validation sets.

[0069] The data processing module is used to quantify the weight values ​​of each sensor data by calculating the information entropy of each sensor data in the training set; to quantify the mutual information compensation factor between sensor data by calculating the mutual information between sensor data; to generate adaptive weights by combining the information entropy and the mutual information compensation factor; to use the adaptive weights to perform weighted fusion of several classes of sensor data of the same sensor data sequence in the training set to obtain a fused feature vector; to perform density clustering on the fused feature vector to generate multiple cluster structures; to calculate the local density value of each sensor data sequence based on the distance from the fused feature vector to the cluster center and the distribution characteristics within the cluster; to generate an enhancement coefficient based on the local density value; and to calculate the enhanced feature vector using the enhancement coefficient and the fused feature vector.

[0070] The neural network module is used to multiply the input values ​​of neural network neurons by the cumulative distribution function and saturation characteristics to obtain the improved GELU activation function of the neural network; and to dynamically adjust the saturation characteristics by combining the batch average gradient magnitude with the input values ​​through the hyperbolic tangent function; to construct the balance weights corresponding to each state-labeled data category based on the amount of data in each state-labeled data category corresponding to the fusion feature vector in the training set; and to calculate the class-balanced focusing loss function corresponding to each state-labeled data category in the neural network using the balance weights and prediction probabilities; the neural network uses a lightweight convolution module to extract the intra-cluster local pattern features corresponding to the enhanced feature vector, and adopts a multi-head self-attention mechanism to output the global key corresponding to the enhanced feature vector. The system generates a clustered branch feature vector; it maps the local pattern features within the cluster and the global associated branch feature vectors to gating weights through linear transformation and the Sigmoid function, and then weights and fuses them to obtain a gating fusion feature vector; it constructs a gating constraint loss function for the neural network using the gating weights and local density values; it inputs the gating constraint loss function into the fully connected layer of the neural network, performs a nonlinear transformation using an improved GELU activation function, and outputs the probability distribution of the predicted state class; it calculates the total loss by weighting the class-balanced focusing loss function and the gating constraint loss function; it iteratively trains the neural network with the goal of minimizing the total loss, and validates the neural network using a validation set until a well-trained neural network that meets the preset conditions is obtained.

[0071] Thirdly, this application provides a state assessment device based on a distributed energy storage system, the device comprising:

[0072] processor;

[0073] And a memory that stores executable code, which, when executed, causes the processor to execute a state assessment method for a distributed energy storage system as described above.

[0074] As can be seen from the above technical solutions, this application has the following advantages:

[0075] I. Improved accuracy of multi-source data fusion:

[0076] An adaptive weighting mechanism constructed using information entropy and mutual information compensation factors enables dynamic weighted fusion of data from multiple sensor types. This scheme directly addresses the feature masking problem caused by traditional simple splicing or weighted averaging. By quantifying data weights and mutual compensation relationships, it ensures that feature information under key operating conditions is fully preserved, providing a high-fidelity fused feature vector for subsequent state assessment.

[0077] II. Optimization capability of feature space distribution:

[0078] This approach, based on density clustering, calculates local density values ​​and generates enhancement coefficients, directly addressing the problem of uneven distribution in the feature space. By quantifying the distance from samples to cluster centers and the distribution characteristics within clusters, it significantly improves the visibility of samples in boundary regions, enabling neural networks to more clearly identify class boundaries and thus improving their adaptability to sparse samples.

[0079] III. Adaptive Optimization of Neural Network Structures:

[0080] The improved GELU activation function, combined with the hyperbolic tangent function to dynamically adjust saturation characteristics, directly enhances the model's ability to express sparse features. Meanwhile, the class-balanced focusing loss function, constructed by balancing the weights of state-labeled data categories, directly mitigates the impact of uneven sample distribution on training. The fusion of branch features from the lightweight convolutional module and the multi-head self-attention mechanism further enables adaptive extraction of local patterns and global dependencies.

[0081] IV. Synergistic Effects of Loss Function Design:

[0082] The weighted combination of the gating constraint loss function and the class-balanced focusing loss function directly serves the dual objectives of model training: strengthening the rationality of feature selection through the correlation between gating weights and local density values, and ensuring equal learning opportunities for each state class through class-balanced weights. This collaborative design directly improves the model's evaluation robustness under complex conditions. Attached Figure Description

[0083] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0084] Figure 1 This is a flowchart of a distributed energy storage system status assessment method provided in an embodiment of this application.

[0085] Figure 2 This is a schematic diagram comparing the effects of multi-source sensor data fusion provided in an embodiment of this application.

[0086] Figure 3 This is a schematic diagram comparing the enhancement effect of density-sensitive features provided in an embodiment of this application.

[0087] Figure 4 This is a schematic diagram comparing improved activation function gradient stability analysis activation functions provided in an embodiment of this application.

[0088] Figure 5This is a schematic diagram of the internal structure of a distributed energy storage system state assessment system provided in an embodiment of this application.

[0089] Figure 6 This is a schematic diagram of the internal structure of a distributed energy storage system status assessment device provided in an embodiment of this application. Detailed Implementation

[0090] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0092] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0093] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0094] The embodiment provides a method for assessing the state of a distributed energy storage system, such as... Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps:

[0095] Step 110: Deploy several types of sensors at each preset key point in the distributed energy storage system to obtain sensor datasets under different operating conditions; obtain state labeling data corresponding to the sensor data sequences; divide the sensor dataset into training set and validation set.

[0096] The sensor dataset consists of several sensor data sequences, and the same sensor data sequence contains various types of sensor data at each preset key point at the same acquisition time.

[0097] It should be noted that deploying several types of sensors at each preset key point in the distributed energy storage system can specifically be done by deploying a network of four types of sensors—voltage, current, temperature, and impedance—at each key measurement point in the distributed energy storage system, with a total of 10 spatial locations, to achieve synchronous monitoring throughout the entire lifecycle.

[0098] The data acquisition process uses high-precision sampling equipment to acquire raw monitoring signals in real time at a fixed frequency. The frequency can be set to 1kHz, covering various operating conditions such as charging and discharging, static placement, and fault simulation.

[0099] The voltage sensor records the instantaneous voltage fluctuations and ripple characteristics at each measurement point, the current sensor captures the amplitude and harmonic components of the charging and discharging current, the temperature sensor monitors the temperature gradient between the battery surface and the internal core region, and the impedance sensor obtains the characteristic values ​​of the electrochemical impedance spectrum through the AC injection method.

[0100] Data labeling is performed by domain experts who mark different state categories based on the operation logs. In one embodiment, the categories of state labeling data include: normal charging and discharging, overvoltage state, undervoltage state, overtemperature state, and local aging.

[0101] Step 120: Quantify the weight value of each sensor data by calculating the information entropy of each sensor data in the training set; quantify the mutual information compensation factor between sensor data by calculating the mutual information between sensor data; generate adaptive weights by combining information entropy and mutual information compensation factor; and use the adaptive weights to perform weighted fusion of several types of sensor data in the same sensor data sequence in the training set to obtain a fused feature vector.

[0102] It should be noted that the sensor monitoring data of distributed energy storage systems, including four types of sensor monitoring data such as voltage, current, temperature and impedance, have the characteristics of multi-source heterogeneity. This is manifested in the significant differences in the dimensions of different sensors, uneven noise distribution, and nonlinear correlation between features. Conventional weighted average fusion methods ignore the nonlinear dependence and noise characteristics between features, which can easily lead to the loss of key state information and cannot effectively handle the problem of multi-source heterogeneity.

[0103] This step processes the raw sensor data using an entropy-weighted correlation dual-driven fusion strategy. It integrates multi-source features in an adaptive weighting manner, reducing the weight of high-noise features and strengthening the contribution of relevant features, thereby addressing the multi-source heterogeneity problem and preserving key state information. The specific steps are as follows:

[0104] 1) Feature information entropy weight allocation:

[0105] The information entropy of each sensor data point is calculated. Specifically, the normalized proportion of each feature value is obtained by normalizing the absolute value of the normalized feature value (sensor data), and the information entropy is calculated based on this proportion. Information entropy reflects the amount of information or uncertainty of a feature. Features with high information entropy have high noise or scattered information, and their weights are reduced during fusion. Features with low information entropy have concentrated information and are given higher weights, thereby suppressing the influence of noise and highlighting information-rich features. This is expressed as:

[0106] ,

[0107] In the formula, For the first The information entropy of sensor data characterizes the amount of information or uncertainty of that feature (sensor data);

[0108] For the first The nth sample (the nth in the training set) The first sensor data sequence) dimensional eigenvalues ​​(the first) The first sensor data sequence The normalized proportion of data from each sensor is calculated as the ratio of the absolute value of the feature value, ensuring that it is non-negative and sums to 1. The calculation method is expressed as follows: ;

[0109] The number of samples (the total number of sensor data sequences in the training set);

[0110] For sample index, ;

[0111] For sensor data indexing, , to These are the values ​​taken from the 1st to the 10th voltage measuring point. to These are the values ​​taken from the 1st to the 10th current measuring point. to These are the values ​​taken from the 1st to the 10th temperature measuring point. to These are the values ​​taken from the 1st to the 10th impedance measurement point;

[0112] For the j-th sensor data in the i-th sensor data sequence in the training set;

[0113] This represents the absolute value of the j-th sensor data in the i-th sensor data sequence in the training set;

[0114] The absolute value operator is represented by the symbol.

[0115] The natural logarithm is... Logarithm with base 0.

[0116] 2) Generation of cross-feature mutual information compensation factors:

[0117] A mutual information compensation factor is defined. This factor quantifies the correlation between features by calculating their mutual information. Feature pairs with high mutual information have strong dependencies, so their compensation factor is increased to strengthen their weights. Feature pairs with low mutual information have weak dependencies, so their weights are adjusted less. This addresses the problem of neglecting nonlinear correlations between features and enhances the contribution of related features. It is expressed as:

[0118] ,

[0119] In the formula, For the first A mutual information compensation factor for sensor data is used to adjust weights to reflect the correlation between sensor data.

[0120] For the first A vector of sensor data;

[0121] For the first A vector of sensor data;

[0122] To distinguish from Sensor data index;

[0123] The number of sensor data points in a sensor data sequence, taking 10 measurement points as an example, includes the values ​​from four types of sensors: voltage, current, temperature, and impedance, at the 10 measurement points. That is, the feature dimension is... ;

[0124] Features and The mutual information between them is calculated as follows: ;

[0125] Features and The joint probability distribution of ;

[0126] Features The marginal probability distribution;

[0127] Features The marginal probability distribution.

[0128] It should be noted that mutual information To measure the nonlinear dependency between two features, specifically quantify the spatial correlation of the same sensor type at different measurement points, such as the mutual information between voltage measurement point 1 and measurement point 2, or the coupling relationship between different sensor types at the same measurement point, such as the mutual information between voltage and temperature at a certain measurement point. The larger the value, the stronger the dependency.

[0129] 3) Adaptive weighted fusion feature generation:

[0130] An adaptive weight is generated by combining information entropy and mutual information compensation factor. This adaptive weight is then used to perform weighted fusion of features from multiple sensor sources. Entropy weights reduce the impact of high-noise features, while mutual information compensation factor enhances the contribution of relevant features. The resulting fused feature vector is expressed as:

[0131] ,

[0132] In the formula, To fuse feature vectors, calculations are performed for each sensor data sequence;

[0133] For the first The adaptive weights for each sensor data point have values ​​between 0 and 1, and the sum of all weights is 1. The calculation method is expressed as follows: .

[0134] It should be noted that the information entropy weight and the mutual information compensation factor are nonlinearly coupled, through... Item and The product structure achieves a dual driving force, on the one hand, On the one hand, the weights of high-noise features are directly suppressed, while the features with concentrated information are preserved. As a mutual information compensation factor, it amplifies the weights of features such as strong spatial correlation between adjacent voltage measurement points or cross-physical quantity coupling such as voltage and temperature at a certain measurement point. This achieves dual-factor synergistic noise suppression while strengthening the implicit physical correlation between features. For example, when a current sensor experiences high-frequency noise... When the temperature rises, conventional methods that directly reduce the weight of the measuring point may lose the coupling information between the measuring point and other temperature measuring points. However, this application achieves this by... The compensation mechanism for related feature pairs retains the associated features across sensor types, enabling the fused features to reflect the synergistic mechanism of multiple physical fields in distributed energy storage systems, which is something that traditional weighted average or single entropy weight methods cannot achieve.

[0135] In one embodiment, a comparative analysis of the fusion effect of multi-source sensor data is performed. The distribution of features from the original sensor data and the fused data is compared using three-dimensional spatial visualization. To intuitively demonstrate the clustering effect of adaptive fusion of multi-source sensor data on various categories, three categories—normal charging / discharging, over-temperature state, and local aging—are selected for visualization, showcasing the effectiveness of the entropy weight correlation dual-driven fusion strategy. In the original feature map on the left, the voltage, current, and temperature sensor data exhibit a high degree of overlap in three-dimensional space. The sample points from the normal state, over-temperature state, and local aging state are mixed together. In particular, the over-temperature state (red dots) and the local aging state (blue dots) show significant overlap in the region of approximately 3.6 volts, 50 amperes, and 20 degrees Celsius, reflecting the problem of large dimensional differences and weak feature correlation in the original multi-source heterogeneous data. Figure 2 As shown, in the fusion feature map of this application, the feature space changes significantly after entropy weight denoising and mutual information compensation. The three types of states form clearly separated clustering regions in the three fusion feature dimensions, each forming clusters in relatively independent intervals. This proves that the dual-drive fusion strategy of this application effectively captures cross-sensor physical relationships such as voltage-temperature coupling and current-impedance correlation, providing a high-quality feature foundation for subsequent state classification.

[0136] Step 130: Perform density clustering on the fused feature vector to generate multiple cluster structures. Based on the distance from the fused feature vector to the cluster center and the distribution characteristics within the cluster, calculate the local density value of each sensor data sequence. Generate an enhancement coefficient based on the local density value. Calculate the enhanced feature vector using the enhancement coefficient and the fused feature vector.

[0137] It should be noted that the state data of distributed energy storage systems is sparsely distributed at the decision boundary, resulting in blurred state boundaries. Conventional standardization methods cannot effectively improve the distribution characteristics of the boundary region, making it difficult to distinguish the transition region between adjacent states and affecting the clarity of the decision boundary of the classification model.

[0138] This application processes fused features using density-sensitive feature scaling technology, calculates the local density of samples using clustering results, and adaptively enhances boundary samples to compress the distribution of boundary regions and improve the clarity of decision boundaries. The specific steps are as follows:

[0139] 1) Cluster-based local density calculation:

[0140] Density clustering is performed on the fused features to generate multiple cluster structures. Based on the distance from the sample to the cluster center and the distribution characteristics within the cluster, the local density value of each sample is calculated. Low density values ​​correspond to samples in the cluster boundary or sparse region, while high density values ​​correspond to samples in the cluster core region. This quantifies the distribution density of samples in the feature space, as shown below:

[0141] ,

[0142] In the formula, Let be the local density value of the i-th sample (the i-th sensor data sequence in the training set). The larger the value, the closer the sample is to the cluster center.

[0143] Let be the fusion feature vector of the i-th sample;

[0144] For clusters The standard deviation is calculated as follows: ;

[0145] The first generation of density clustering Each cluster is a set of samples (training set);

[0146] This is the fused feature vector of the j-th sample (the j-th sensor data sequence in the training set);

[0147] Cluster index for density clustering;

[0148] For clusters The number of samples included;

[0149] For clusters The center vector;

[0150] It is the L2 norm, calculated in the same way as the Euclidean distance;

[0151] It is a natural exponential function.

[0152] 2) Enhanced centripetal effect in sparse boundary regions:

[0153] Enhancement coefficients are generated based on local density values. Sample positions are adjusted using density-sensitive linear interpolation, resulting in the contraction of low-density boundary samples towards the cluster center, while high-density core samples retain their original values. This compresses the sparsity of the boundary region, sharpens the decision boundary, and improves the clarity of state classification. This can be represented as:

[0154] ,

[0155] In the formula, This is the enhanced feature vector of the i-th sample (the i-th sensor data sequence in the training set);

[0156] Let be the enhancement coefficient for the i-th sample. The lower the density, the smaller the value. The calculation method is expressed as:

[0157] To enhance the strength factor and control the boundary compressive strength, for example, ;

[0158] For clusters The center vector.

[0159] It should be noted that enhancing feature vectors In the adaptive acquisition process, local density is combined with dynamic enhancement coefficients, causing samples in sparse regions to adaptively move towards high-density regions, based on local density values. Dynamic adjustment enhancement coefficient Using the exponential function Map density values ​​to The interval achieves a smooth contraction of boundary samples towards the cluster center. Time represents the boundary sample, Enhanced features Approaching the cluster center The distribution sparsity of the compressed boundary region, when Time represents the core sample. Enhanced features retain original values To avoid distortion in the core area, density-sensitive linear interpolation is used to sharpen the decision boundary while maintaining the distribution of core data, thus solving the boundary ambiguity problem that conventional standardization methods cannot handle.

[0160] In one embodiment, such as Figure 3 As shown, the effect of density-sensitive feature enhancement is analyzed through the original feature distribution map and the enhanced distribution map. Taking the normal state and the overheated state as examples, the distribution changes in the boundary region are revealed by the kernel density estimation curve, verifying the effect of density-sensitive feature enhancement technology on improving the clarity of the decision boundary. In the original feature distribution map, the normal state (green curve) and the overheated state (red curve) have a significant overlap in the feature value range of 1.0-2.5, reflecting the serious problem of blurred boundary between the two states in the original feature space. In the enhanced distribution map, key changes occur. The peak value of the normal state distribution shifts to the left from the feature value of 1.0 to around 1.1, and the peak value of the overheated state distribution shifts to the right from 2.0 to around 1.7. The overlap of the two distributions is significantly narrowed. The distribution change reflects the effect of the density-sensitive scaling strategy. After detecting low-density sample points, the adaptive enhancement coefficient shrinks them towards their respective cluster centers, effectively compressing the sparse region of the decision boundary.

[0161] Step 140: Multiply the input value of the neural network neuron with the cumulative distribution function and saturation characteristic to obtain the improved GELU activation function of the neural network; and dynamically adjust the saturation characteristic by combining the batch average gradient magnitude with the input value through the hyperbolic tangent function; construct the balance weights corresponding to each state labeled data category according to the amount of data of each state labeled data category corresponding to the fusion feature vector of the training set; and calculate the class balance focusing loss function corresponding to each state labeled data category in the neural network using the balance weights and prediction probability.

[0162] In some embodiments, this step may be further specified as follows:

[0163] S141. Define the improved GELU activation function:

[0164] In the state classification of distributed energy storage systems, the conventional ReLU activation function is prone to losing subtle state features in sparse activation scenarios, causing key state information such as early fault signs to be ignored. Existing activation functions lack adaptability to dynamic gradient changes and cannot solve the gradient vanishing and gradient exploding problems at the same time.

[0165] This application integrates Gaussian gating and adaptive saturation mechanisms, employs an improved GELU activation function to dynamically adjust the nonlinear characteristics of the activation function, and adaptively adjusts the saturation region according to the gradient magnitude. This enhances the nonlinear response when the gradient vanishes and suppresses the saturation effect when the gradient explodes, thereby preserving subtle state characteristics. The specific steps are as follows:

[0166] 1) Gaussian probability gating transformation:

[0167] A Gaussian gating mechanism based on the cumulative distribution function of the standard normal distribution is constructed. The input value is multiplied by the probability gate, which preserves the continuity of the input signal, avoids hard saturation, and enhances the response capability to subtle state features. This is expressed as:

[0168] ,

[0169] In the formula, For the improved GELU activation function, Output values ​​for the improved GELU activation function;

[0170] Input values ​​for neurons;

[0171] The cumulative distribution function of the standard normal distribution represents the cumulative probability of an input value within the normal distribution. It is calculated as follows: ,in is the normalization constant for the normal distribution;

[0172] Represents the range from negative infinity to the current input value. The definite integral;

[0173] It is a natural constant;

[0174] t is the integration variable, representing the independent variable in the probability density function of the standard normal distribution. Its physical meaning is the intermediate variable of the input value z of the input neuron during the integration process.

[0175] dt is an integral infinitesimal element, representing the infinitesimal increment of the integral variable t;

[0176] This is a dynamic saturation function that controls the nonlinear intensity of the activation function, with an initial value set to 1.

[0177] 2) Gradient-driven dynamic adjustment of saturation characteristics:

[0178] By combining the batch average gradient magnitude with the input value using the hyperbolic tangent function, a dynamic saturation function is calculated. This adjusts the saturation characteristics of the activation function, maintaining standard Gaussian gating characteristics when the gradient is small and suppressing the output magnitude when the gradient is large. This achieves an adaptive balance between gradient vanishing and exploding, improving the detection sensitivity of fault symptoms. This can be expressed as:

[0179] ,

[0180] In the formula, The average gradient magnitude of the batch is calculated as follows: ;

[0181] The scaling factor adjusts the intensity of the input value's influence on the saturation coefficient, such as... ;

[0182] It is the hyperbolic tangent function;

[0183] Batch size;

[0184] The loss function for the i-th sample (the i-th sensor data sequence in the training set) applied to the neuron input value The gradient.

[0185] In one embodiment, such as Figure 4As shown, gradient stability analysis of the improved activation function is performed. A dual comparison of the activation function output curve and gradient change curve reveals the innovative advantages of the improved Gaussian error linear unit in dynamic gradient adjustment. In the upper activation function comparison graph, the conventional linear rectifier unit outputs zero when the input is negative, exhibiting hard saturation. While the Gaussian error linear unit maintains a negative response, the change is gradual. The improved Gaussian error linear unit (red curve) of this application exhibits a steeper nonlinear response in the input value range of -1 to 1, especially in the micro-input range of -0.5 to 0.5, where its output slope is significantly higher than that of the conventional method, enhancing the model's sensitivity to weak fault signals. The gradient stability diagram below is more significant. The gradient of a conventional linear rectifier unit is always 1 when the input is greater than zero (blue horizontal line), but drops to zero in the negative value region, posing a risk of gradient truncation. The method of this application (red curve) maintains a smooth gradient transition in a wide range of input values ​​from -3 to 3. It avoids the zero gradient dead zone of conventional linear rectifier units and the gradient decay of Gaussian error linear units near the input value of ±2. The smooth gradient characteristics across the entire range are beneficial for capturing the weak characteristics of early faults in distributed energy storage systems.

[0186] It should be noted that the saturation characteristics of conventional activation functions such as ReLU or GELU are static. However, fault symptom signals are usually weak and sparse, and static activation functions easily miss these subtle features. The improved GELU activation function uses a dynamic saturation function. Dynamically adjust the saturation zone. The batch average gradient magnitude By combining it with the hyperbolic tangent function, an adaptive balance between gradient stability and nonlinear response is achieved. The term enables dynamic adjustment of input dependencies, especially when the gradient is small. To maintain the continuous probability gating property of the standard GELU activation function and avoid the hard saturation problem of the ReLU activation function, when the gradient is large, By using the saturation characteristics of the hyperbolic tangent function, the output amplitude is suppressed to prevent gradient explosion. The dynamic saturation mechanism can automatically adjust the sensitivity according to the gradient state, thereby improving the detection rate of early faults.

[0187] S142. Define class-balanced focus loss:

[0188] Distributed energy storage system state classification suffers from a significant long-tail distribution problem, such as the scarcity of fault samples. Conventional cross-entropy loss functions treat all samples equally, neglecting the mining of hard examples and class imbalance, resulting in insufficient ability of the model to identify rare state categories.

[0189] This application employs a class-balanced focusing loss function, and through dual-mechanism collaborative optimization, it combines a class frequency weighting factor to suppress majority class dominance and integrates a focusing factor to strengthen the training weights of difficult examples. The specific steps are as follows:

[0190] 1) Calculation of class-frequency inverse proportional balancing weights:

[0191] The number of samples in each class of the training set is counted, and a balance weight term inversely proportional to the class frequency is constructed. This balance weight term is then adjusted exponentially by a balance adjustment factor to reduce the loss weight of the majority class samples, thereby alleviating class imbalance and suppressing the dominant influence of the majority class on model training. This is expressed as:

[0192] ,

[0193] In the formula, The balancing weights for the c-th category (state labeling data category);

[0194] c is the category index. ;

[0195] c represents the total number of categories;

[0196] Let c be the number of samples in the c-th category;

[0197] To balance the adjustment factors and control the intensity of the penalty for class imbalance, for example, .

[0198] 2) Difficult case prediction probability focusing correction:

[0199] Incorporating a prediction probability modulation factor into the cross-entropy loss function amplifies the loss for samples with low true class prediction probabilities. This modulation factor, through a focusing factor, controls the reinforcement intensity of difficult examples, forcing the model to focus on classification boundary samples and rare class samples, thus improving the ability to identify difficult examples. This can be expressed as:

[0200] ,

[0201] In the formula, To balance the focus loss;

[0202] Label the data category of the true state of the i-th sample (the i-th sensor data sequence in the training set);

[0203] for Balanced weights for labeling the data categories of the true state of the i-th sample;

[0204] Label the data category of the model for the true state of the i-th sample. The predicted probability;

[0205] This represents the prediction error term; the larger the value, the higher the difficulty of classifying the sample (sensor data sequence in the training set).

[0206] As a focusing factor, the loss amplification intensity of difficult examples is adjusted, such as... ;

[0207] It is a logarithmic function, with the default base being the natural constant.

[0208] It should be noted that the balanced focusing loss addresses the long-tail distribution problem through a dual-mechanism synergy. This item reduces the majority class weight by combining a balancing adjustment factor with power-law adjustment, thereby alleviating class imbalance. This method amplifies the loss for samples with low prediction probabilities, forcing the model to focus on classification boundaries and rare samples. This achieves a unified framework that integrates class balance and hard-case focus, rather than simply combining them. The product form of the terms provides double reinforcement to the difficult samples of the minority class, further enhancing the model's sensitivity to rare faults. This results in the gradient update magnitude of the difficult samples of the minority class being significantly higher than that of the easily distinguishable samples of the majority class, thus naturally forming an adaptation to the long-tailed distribution during model training.

[0209] Step 150: The neural network uses a lightweight convolution module to extract the intra-cluster local pattern features corresponding to the enhanced feature vector. It adopts a multi-head self-attention mechanism to output the global association branch feature vector corresponding to the enhanced feature vector. The intra-cluster local pattern features and the global association branch feature vector are mapped to gating weights through linear transformation and Sigmoid function, and then weighted and fused to obtain the gating fused feature vector.

[0210] It should be noted that the state assessment of distributed energy storage systems needs to simultaneously handle multi-scale spatial features, correlations between different measurement points, and ambiguity of state boundaries. Conventional single network architectures are difficult to adaptively integrate local and global features and ignore the prior knowledge of sample distribution revealed by density clustering, resulting in insufficient generalization ability of the state decision surface.

[0211] This application employs a density-aware gated hybrid neural network, and achieves multi-scale feature collaborative optimization through a dual-branch heterogeneous architecture and a dynamic feature routing mechanism. The specific steps are as follows:

[0212] S151, Dual-branch feature extraction:

[0213] 1) Local density branch feature extraction:

[0214] The local density branch uses a lightweight convolutional module to process the difference between the sample and the cluster center, extracting local pattern features within the cluster and enhancing the network's sensitivity to boundary region features. This is represented as:

[0215] ,

[0216] In the formula, For the first The local density branch feature vector of a sample (the i-th sensor data sequence in the training set);

[0217] This is a one-dimensional convolution operation;

[0218] For clusters The corresponding convolution kernel weight matrix is ​​a trainable parameter;

[0219] For clusters The center vector.

[0220] 2) Global correlation branch feature extraction:

[0221] The global correlation branch employs a multi-head self-attention mechanism to enhance feature vectors. It captures cross-measurement point feature dependencies through query matrices, key matrices, and value matrices. The output, after layer normalization, retains global spatial coupling information, thus resolving the nonlinear correlation problem of sensor data from different measurement points. This is represented as:

[0222] ,

[0223] In the formula, For the first The global correlation branch feature vector of each sample (the i-th sensor data sequence in the training set);

[0224] For layer normalization operation;

[0225] This is a multi-head self-attention mechanism;

[0226] For query matrix Key matrix Value matrix Multi-head self-attention output;

[0227] For querying the matrix, ;

[0228] The key matrix, ;

[0229] For value matrices, ;

[0230] The projection weight matrix is ​​a trainable parameter for querying.

[0231] Let be the key projection weight matrix, which are trainable parameters;

[0232] The projected weight matrix is ​​the value, and the parameters are trainable.

[0233] It should be noted that the multi-head self-attention mechanism allows each element in the input sequence (sensor data sequence) to dynamically pay attention to all other elements in the sequence. Elements correspond to the feature values ​​of the enhanced feature vectors corresponding to different measurement points and different types of sensors in this application. Information is aggregated based on their correlations. Specifically, the input enhanced feature vectors are converted into three different representations: a query matrix, a key matrix, and a value matrix. The query matrix represents what information the current element wants to "ask," the key matrix represents what "identifier" information other elements can provide for matching the query, and the value matrix contains the actual information content that other elements want to convey. Then, the relationship between the query vector and all key vectors is calculated. Similarity scores are normalized into attention weights, and then these attention weights are used to perform a weighted summation of the corresponding value matrices to obtain a new representation that reflects global correlation. Here, "multi-head" means that this process is repeated in parallel multiple times, with each head using different linear transformation parameters to learn different types of dependencies in the input sequence. Finally, the new representations calculated by all heads are concatenated and integrated into the final output vector through another linear transformation. This enables the global correlation branch to effectively capture the complex, nonlinear spatial coupling and dependencies between different measurement points and different physical quantities in the distributed energy storage system, providing global feature information for state assessment.

[0234] The bi-branch feature vectors are mapped to gating weights using linear transformation and the sigmoid function. Then, the gating weights are used to weight and fuse the local density branch features and the global correlation branch features, achieving adaptive integration of the bi-branch features, as shown below:

[0235] ,

[0236] ,

[0237] In the formula, For the first The gating weights of each sample (the i-th sensor data sequence in the training set);

[0238] The weight matrix is ​​the gated linear transformation weight matrix, and the parameters are trainable.

[0239] The gating weight vector is a trainable parameter;

[0240] for The transpose of is a trainable parameter;

[0241] This is the gated bias term, which is a trainable parameter;

[0242] Use the Sigmoid activation function;

[0243] For the first Gated fusion feature vectors of each sample.

[0244] Step 160: Construct the gated loss function of the neural network using gated weights and local density values; input the gated loss function into the fully connected layer of the neural network, perform nonlinear transformation using the improved GELU activation function, and output the predicted state category probability distribution.

[0245] In some embodiments, this step may be further specified as follows:

[0246] Gated density-dependent constraints:

[0247] A gated loss term is constructed, which forces the gate weights to be correlated with local density values. The gate weights for low-density samples tend to strengthen local branch features, while high-density samples retain global features. The feature routing strategy is optimized using density prior knowledge, as follows:

[0248] ,

[0249] In the formula, This is the gate constraint loss.

[0250] Fully connected classifier decision:

[0251] The fused features are input into a fully connected layer, and a non-linear transformation is performed using an improved GELU activation function. The output is then processed by a softmax function to generate a predicted class probability distribution, thus completing the final decision for state classification. This can be represented as:

[0252] ,

[0253] For the first The probability distribution of the predicted state class of each sample (the i-th sensor data sequence in the training set);

[0254] For the Softmax function;

[0255] is the classifier weight matrix, which consists of trainable parameters;

[0256] is the weight matrix of the fully connected layer, which are trainable parameters;

[0257] These are bias terms for the fully connected layer and are trainable parameters;

[0258] The classifier bias term is a trainable parameter.

[0259] Step 170: Calculate the total loss by weighting the class-balanced focusing loss function and the gated constraint loss function; train the neural network iteratively with the goal of minimizing the total loss, and verify the neural network using the validation set until a well-trained neural network that meets the preset conditions is obtained.

[0260] In some embodiments, this step may be further specified as follows:

[0261] Total loss calculation:

[0262] The total loss is a weighted average of the class-balanced focusing loss and the gating constraint loss. The gating constraint strength factor controls the contribution of density prior knowledge to model training, achieving collaborative optimization of classification accuracy and feature routing, expressed as:

[0263] ,

[0264] In the formula, This represents the total loss during model training.

[0265] For example, the gate constraint strength factor. .

[0266] It should be noted that, The term represents the local density value obtained from unsupervised clustering. Gating weights of supervised networks To associate, At that time, force the gating weights of low-density samples That is, to strengthen the global interconnected branches, and In this scenario, high-density samples tend to branch into local density branches, transforming the spatial distribution knowledge of samples into a feature routing strategy. This makes boundary samples more reliant on global features, capturing cross-measurement point correlations, while core samples focus on local patterns, improving the physical interpretability of the decision surface. Furthermore... The negative correlation design of the terms forms an information bottleneck, preventing the gating mechanism from overfitting noise, enabling the network to adaptively distinguish between sparse fault samples that require global correlation analysis and typical state samples that require local pattern recognition, thus solving the defect of multi-branch networks ignoring sample distribution characteristics.

[0267] Loop iteration and parameter update:

[0268] The model training employs a mini-batch stochastic gradient descent strategy, with the total loss... To optimize the objective, each iteration consists of two phases: forward propagation and backward propagation.

[0269] In the forward stage, the batch-enhanced feature vectors are input into the density-aware hybrid network, which then sequentially performs local density branch convolution, global correlation branch self-attention calculation, density-gated fusion, and classification decision to generate the predicted state class probability distribution. ;

[0270] Backpropagation phase based on total loss Calculate the gradient and update the trainable parameters of the model through the adaptive moment estimate optimizer;

[0271] After each training cycle, the state classification accuracy and rare fault recall are evaluated on the validation set. The iteration is terminated when the loss function converges or the early stopping condition is triggered, and the optimal parameters are saved. The early stopping condition can be set to the upper limit of 1000 iterations.

[0272] After obtaining a trained neural network that meets the preset conditions, the method also includes:

[0273] Newly connected real-time sensor data streams are processed into sensor data sequences according to the acquisition time.

[0274] The weight values ​​of each sensor data are quantified by calculating the information entropy of each sensor data in the same sensor data sequence; the mutual information between sensor data is quantified by calculating the mutual information between sensor data in the same sensor data sequence; adaptive weights are generated by combining information entropy and mutual information compensation factors; and the adaptive weights are used to perform weighted fusion of several types of sensor data in the same sensor data sequence in the training set to obtain a fused feature vector.

[0275] Density clustering is performed on the fused feature vectors to generate multiple cluster structures. Based on the distance from the fused feature vectors to the cluster center and the distribution characteristics within the cluster, the local density value of each fused feature vector is calculated. An enhancement coefficient is generated based on the local density value. The enhancement coefficient and the fused feature vector are used to calculate the enhanced feature vector.

[0276] The enhanced feature vectors are input into the trained neural network to obtain the output predicted state category probability distribution.

[0277] As described above, this embodiment employs an adaptive weighted fusion strategy driven by both information entropy weights and mutual information compensation factors to address the issues of dimensional differences, nonlinear dependencies, and uneven noise in multi-source heterogeneous sensor data, effectively preserving key state information. Based on clustering-based local density estimation and enhanced coefficient interpolation strategies, it achieves smooth compression of boundary samples towards the center, significantly improving the clarity of decision boundaries for state classification, particularly suitable for ambiguous state transition regions. A gradient-driven dynamic saturation adjustment mechanism is adopted to enable the activation function to have gradient adaptive capabilities, solving the problems of gradient vanishing and exploding, and enhancing the model's response to early, weak fault features. By employing local density branches and global association branches, combined with density-driven gated fusion and constraint loss, and introducing prior knowledge of sample distribution, it achieves collaborative optimization of feature extraction and feature routing, improving the model's interpretability and generalization ability.

[0278] In addition, this application Figure 5 This application provides a distributed energy storage system state assessment system as an embodiment. For example... Figure 5 As shown in the embodiments of this application, the system mainly includes:

[0279] The data collection module 210 is used to deploy several types of sensors at each preset key point of the distributed energy storage system, thereby obtaining sensor datasets under different operating conditions; wherein, the sensor dataset consists of several sensor data sequences, and the same sensor data sequence contains various types of sensor data at each preset key point at the same acquisition time; obtain the state labeling data corresponding to the sensor data sequence; and divide the sensor data set into training set and validation set;

[0280] The data processing module 220 is used to quantify the weight values ​​of each sensor data by calculating the information entropy of each sensor data in the training set; to quantify the mutual information compensation factor between sensor data by calculating the mutual information between sensor data; to generate adaptive weights by combining the information entropy and the mutual information compensation factor; to use the adaptive weights to perform weighted fusion of several classes of sensor data of the same sensor data sequence in the training set to obtain a fused feature vector; to perform density clustering on the fused feature vector to generate multiple cluster structures; to calculate the local density value of each sensor data sequence based on the distance from the fused feature vector to the cluster center and the distribution characteristics within the cluster; to generate an enhancement coefficient based on the local density value; and to calculate the enhanced feature vector using the enhancement coefficient and the fused feature vector.

[0281] Neural network module 230 is used to multiply the input values ​​of neural network neurons by the cumulative distribution function and saturation characteristics to obtain the improved GELU activation function of the neural network; and to dynamically adjust the saturation characteristics by combining the batch average gradient magnitude with the input values ​​through the hyperbolic tangent function; to construct the balance weights corresponding to each state labeled data category based on the amount of data in each state labeled data category of the fusion feature vector corresponding to the training set; to calculate the class balance focusing loss function corresponding to each state labeled data category in the neural network using the balance weights and prediction probabilities; and to use a lightweight convolution module to extract the intra-cluster local pattern features corresponding to the enhanced feature vector, and to output the full-valued feature vector corresponding to the enhanced feature vector using a multi-head self-attention mechanism. Local correlation branch feature vectors are generated; local pattern features within clusters and global correlation branch feature vectors are mapped to gating weights through linear transformation and the Sigmoid function, and then weighted and fused to obtain gating fused feature vectors; gating constraint loss function of neural network is constructed using gating weights and local density values; gating constraint loss function is input into fully connected layer of neural network, and nonlinear transformation is performed using improved GELU activation function to output predicted state class probability distribution; total loss is calculated by weighting class balance focusing loss function and gating constraint loss function; neural network is iteratively trained with the goal of minimizing total loss, and the neural network is validated using validation set until a well-trained neural network that meets preset conditions is obtained.

[0282] The above are method embodiments of this application. Based on the same inventive concept, embodiments of this application also provide a distributed energy storage system state assessment device. Figure 6 As shown, the device includes: a processor; and a memory storing executable code thereon, which, when executed, causes the processor to perform a distributed energy storage system state assessment method as described in the above embodiments.

[0283] Specifically, the server deploys several types of sensors at various preset key points in the distributed energy storage system to obtain sensor datasets under different operating conditions. Each sensor dataset consists of several sensor data sequences, with each sequence containing various sensor data from different preset key points at the same acquisition time. The system acquires state labeling data corresponding to the sensor data sequences. The sensor dataset is divided into training and validation sets. The information entropy of each sensor data in the training set is calculated to quantify the weight values. The mutual information between sensor data is calculated to quantify the mutual information compensation factor. Adaptive weights are generated by combining information entropy and mutual information compensation factors. These adaptive weights are then used to weight and fuse several types of sensor data from the same sensor data sequence in the training set to obtain a fused feature vector. Density clustering is performed on the fused feature vector to generate multiple cluster structures. Based on the distance from the fused feature vector to the cluster center and the intra-cluster distribution characteristics, the local density value of each sensor data sequence is calculated. An enhancement coefficient is generated based on the local density value. Using the enhancement coefficient and the fused feature vector, an enhanced feature vector is calculated. Finally, the input value of the neural network neuron is multiplied by the cumulative distribution function and saturation characteristics to obtain the improved GELU activation function of the neural network. The algorithm iterates through the following steps: First, it calculates the batch average gradient magnitude and dynamically adjusts the saturation characteristics by combining the input value with the hyperbolic tangent function. Second, it constructs balanced weights for each state-labeled data category based on the amount of data in each category of the fusion feature vector corresponding to the training set. Third, it calculates the class-balanced focusing loss function for each state-labeled data category in the neural network using the balanced weights and prediction probabilities. Fourth, it uses a lightweight convolution module to extract intra-cluster local pattern features corresponding to the enhanced feature vector, and employs a multi-head self-attention mechanism to output the global association branch feature vector corresponding to the enhanced feature vector. Fifth, it maps the intra-cluster local pattern features and the global association branch feature vector to gate weights through linear transformation and the Sigmoid function, and then weights and fuses them to obtain the gated fusion feature vector. Sixth, it constructs the gated constraint loss function of the neural network using the gate weights and local density values. Seventh, it inputs the gated constraint loss function into the fully connected layer of the neural network, performs a nonlinear transformation using the improved GELU activation function, and outputs the predicted state category probability distribution. Eighth, it calculates the total loss by weighting the class-balanced focusing loss function and the gated constraint loss function. Finally, it iteratively trains the neural network with the goal of minimizing the total loss, and validates the neural network using a validation set until a well-trained neural network that meets the preset conditions is obtained.

[0284] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for assessing the state of a distributed energy storage system, characterized in that, The method includes: Several types of sensors are deployed at each preset key point in the distributed energy storage system to obtain several sensor data sequences under different operating conditions, and a sensor dataset is constructed; the state labeling data corresponding to the sensor data sequences is obtained; the sensor dataset is divided into training set and validation set. By calculating the data from each sensor in the training set, the weight values ​​of each sensor data and the mutual information compensation factor between the sensor data are quantified; adaptive weights are generated by combining information entropy and mutual information compensation factor, and then the fused feature vector is obtained. Using the fused feature vector, the local density value of each sensor data sequence is calculated; using the local density value and the fused feature vector, an enhanced feature vector is calculated; enhancement coefficients are generated based on the local density value, and the sample positions are adjusted through density-sensitive linear interpolation to shrink low-density boundary samples towards the cluster center, while retaining the original values ​​of high-density core samples. This compresses the sparsity of the boundary region distribution, sharpens the decision boundary, and improves the clarity of state classification, as expressed below: In the formula, Let be the enhanced feature vector of the i-th sample; Let be the enhancement coefficient for the i-th sample. The lower the density, the smaller the value. The calculation method is expressed as: ; To enhance the strength factor and control the boundary compressive strength; For clusters The center vector; Let be the local density value of the i-th sample; the larger the value, the closer the sample is to the cluster center. Let be the fusion feature vector of the i-th sample; An improved GELU activation function for the neural network is obtained by utilizing the input values ​​of neural network neurons, the cumulative distribution function, and the saturation characteristic. Balance weights for each state labeling data category are constructed based on the fusion feature vector corresponding to the training set. The class-balanced focusing loss function corresponding to each state labeling data category in the neural network is calculated using the balance weights and prediction probabilities. The improved GELU activation function is obtained by multiplying the input values ​​of the neural network neurons by the cumulative distribution function and the saturation characteristic. The batch average gradient magnitude is combined with the input value using the hyperbolic tangent function to calculate the dynamic saturation function, thus adjusting the saturation characteristics of the activation function. This is expressed as: In the formula, The average gradient magnitude of the batch is calculated as follows: ; This is a scaling factor that adjusts the intensity of the input value's influence on the saturation coefficient. It is the hyperbolic tangent function; Batch size; The loss function for the i-th sample is applied to the neuron input value. The gradient; Extract the intra-cluster local pattern features corresponding to the enhanced feature vector and output the global association branch feature vector corresponding to the enhanced feature vector; map the intra-cluster local pattern features and the global association branch feature vector to gating weights, and perform weighted fusion to obtain the gating fusion feature vector, specifically including: The neural network uses a lightweight convolution module to extract intra-cluster local pattern features corresponding to the enhanced feature vector. It employs a multi-head self-attention mechanism to output the global association branch feature vector corresponding to the enhanced feature vector. The intra-cluster local pattern features and the global association branch feature vector are mapped to gating weights through linear transformation and the sigmoid function, and then weighted and fused to obtain the gating fused feature vector. A gated constraint loss function for the neural network is constructed using gate weights and local density values; the gated constraint loss is... The gated fused feature vector is input into the fully connected layer of the neural network, and a nonlinear transformation is performed using an improved GELU activation function to output the predicted state class probability distribution; where, The term represents the local density value obtained from unsupervised clustering. Gating weights of supervised networks To establish a connection; The total loss is calculated by weighting the balanced focusing loss function and the gated constraint loss function; the neural network is iteratively trained with the goal of minimizing the total loss, and the neural network is validated using a validation set until a well-trained neural network that meets the preset conditions is obtained; the balanced focusing loss function is expressed as: In the formula, To balance the focus loss; Label the data category of the true state of the i-th sample; Balanced weights for labeling the data categories of the true state of the i-th sample; Label the data category of the model for the true state of the i-th sample. The predicted probability; This represents the prediction error term; As a focusing factor, it adjusts the loss amplification intensity of difficult sample cases; It is a logarithmic function, with the default base being the natural constant.

2. The distributed energy storage system state assessment method according to claim 1, characterized in that, The same sensor data sequence contains various types of sensor data at preset key points at the same acquisition time.

3. The distributed energy storage system state assessment method according to claim 1, characterized in that, By calculating the data from each sensor in the training set, the weight values ​​of each sensor data and the mutual information compensation factor between the sensor data are quantified; adaptive weights are generated by combining information entropy and mutual information compensation factor, and then a fused feature vector is obtained, specifically including: The weight values ​​of each sensor data are quantified by calculating the information entropy of each sensor data in the training set; the mutual information between sensor data is quantified by calculating the mutual information compensation factor between sensor data. An adaptive weight is generated by combining information entropy and mutual information compensation factor. The adaptive weight is then used to perform weighted fusion of several types of sensor data in the same sensor data sequence to obtain a fused feature vector.

4. The distributed energy storage system state assessment method according to claim 1, characterized in that, Using the fused feature vector, the local density value of each sensor data sequence is calculated. Specifically, this includes: performing density clustering on the fused feature vector to generate multiple cluster structures, and calculating the local density value of each sensor data sequence based on the distance from the fused feature vector to the cluster center and the distribution characteristics within the cluster.

5. The distributed energy storage system state assessment method according to claim 1, characterized in that, The weight values ​​of each sensor data point are quantified by calculating the information entropy of each sensor data point in the training set, specifically including: The normalized percentage of each sensor data is obtained by normalizing the absolute value of the sensor data, and the information entropy is calculated based on the normalized percentage. The weight values ​​of each sensor data are quantified based on the inverse relationship between information entropy and weight value.

Images