A method and system for dynamic management and control of energy storage batteries

Abstract

The application relates to the technical field of energy storage battery health state monitoring and dynamic management and control, and discloses an energy storage battery dynamic management and control method and system, wherein the energy storage battery dynamic management and control method comprises the following steps: collecting energy storage battery state data and environmental interference data through a multi-modal sensor network to obtain a multi-modal feature set; processing the preprocessed multi-modal feature set by using a multi-layer time sequence convolution network to extract time sequence features of a battery health state, and simultaneously realizing abnormal noise detection and suppression; establishing a mapping relationship between environmental interference and prediction reliability, representing a prediction result as a confidence interval and calculating a reliability index; dynamically adjusting prediction model parameters and structures to generate a health state prediction result and reliability evaluation; generating an energy storage battery dynamic management and control strategy with a safety margin; and effectively capturing battery state and environmental interference information through multi-modal signal acquisition and preprocessing technology to provide a high-quality data basis for health state prediction.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery health status monitoring and dynamic management technology, and more specifically, to a method and system for dynamic management of energy storage batteries. Background Technology

[0002] As key devices for energy storage and conversion, energy storage batteries are widely used in large-scale grid-side energy storage power stations, industrial microgrids, new energy vehicle battery management systems, and distributed energy storage systems. The health status of energy storage batteries directly affects the safety, reliability, and economic benefits of the system; therefore, accurate prediction and effective management of the health status of energy storage batteries are of great significance.

[0003] Currently, methods for predicting and managing the health status of energy storage batteries mainly include physical model-based methods, data-driven machine learning methods, and hybrid methods combining both. Physical model-based methods predict health status by establishing mathematical models of the battery's internal electrochemical reactions and physical processes, offering good interpretability, but they struggle to accurately describe the nonlinear degradation process under complex operating conditions. Data-driven machine learning methods train models using historical data, capturing complex nonlinear relationships, but they are highly dependent on data quality and have limited generalization ability when faced with unseen operating conditions. Hybrid methods combine the advantages of both, but still face many challenges under complex environmental disturbances.

[0004] As the application environment of energy storage batteries becomes increasingly complex, environmental interference factors such as electromagnetic interference, temperature fluctuations, and mechanical vibrations have a growing impact on health status prediction. Existing technologies mainly suffer from the following problems:

[0005] In multi-source noise interference environments, the prediction model struggles to effectively distinguish and suppress different types of noise interference, resulting in insufficient prediction accuracy. It has limited ability to identify subtle changes in characteristics such as micro-cycle discharge, making it unable to accurately capture minute changes in battery health status. It cannot effectively quantify the reliability of prediction results and lacks an assessment of prediction uncertainty. The control strategy is fixed and cannot be adaptively adjusted according to interference conditions and prediction reliability. It does not consider safety margin design under different reliability levels, posing potential safety risks.

[0006] Therefore, there is an urgent need for a technical solution that can cope with complex environmental interference, improve the accuracy and reliability of prediction, and achieve dynamic adaptive control. Summary of the Invention

[0007] This invention provides a dynamic management and control method and system for energy storage batteries, which solves the technical problems in related technologies such as insufficient accuracy in predicting the health status of energy storage batteries under complex environmental interference conditions, difficulty in assessing reliability, and inflexible management and control strategies.

[0008] This invention provides a method for dynamic management and control of energy storage batteries, comprising:

[0009] The system collects energy storage battery status data and environmental interference data through a multimodal sensor network, and performs noise reduction processing on the collected signals to obtain a multimodal feature set.

[0010] By using a multi-layer temporal convolutional network to process the preprocessed multimodal feature set, the temporal features of battery health status are extracted, while abnormal noise detection and suppression are achieved.

[0011] Based on the time-series characteristics of environmental disturbance data and battery health status, a mapping relationship between environmental disturbance and prediction reliability is established. An uncertainty quantification framework is constructed using evidence theory, and the prediction results are expressed as confidence intervals and a reliability index is calculated.

[0012] Based on the disturbance status and prediction reliability, the parameters and structure of the prediction model are dynamically adjusted to generate health status prediction results and their reliability assessment.

[0013] Based on health status prediction results and reliability assessment, a dynamic management and control strategy for energy storage batteries with safety margins is generated.

[0014] Furthermore, the multimodal sensor network includes a battery status sensor group and an environmental interference sensor group; the battery status sensor group is used to collect parameters related to battery health status, including voltage, current, temperature and impedance data; the environmental interference sensor group is used to collect interference signals that affect health status prediction, including electromagnetic field strength, environmental noise, vibration and humidity data.

[0015] Furthermore, the noise reduction process employs multi-scale wavelet decomposition technology, including:

[0016] Multi-scale wavelet decomposition was applied to the time-series data collected by each sensor;

[0017] An adaptive threshold function is applied to suppress noise in the wavelet coefficients obtained from the decomposition.

[0018] The denoised signal is reconstructed using inverse wavelet transform;

[0019] The signal is decomposed into three frequency domain feature sets: high frequency, mid frequency, and low frequency, each corresponding to a different physical meaning.

[0020] Furthermore, the multi-layer temporal convolutional network includes:

[0021] Causal convolutional layers ensure that the model only uses information from before the current time step;

[0022] Dilated convolutional layers expand the receptive field through dilated convolution, and the expansion rate increases exponentially with the number of layers.

[0023] Residual connectivity structures optimize gradient flow and avoid the gradient vanishing problem;

[0024] Skip connections connect features at different levels, enhancing the ability to fuse features at multiple scales.

[0025] Furthermore, the abnormal noise detection employs the Maharanobis distance method, including:

[0026] Calculate the mean vector and covariance matrix of the features;

[0027] Calculate the Mahalanobis distance for each sample point;

[0028] Set an adaptive threshold; when the distance to Mahalanobis exceeds the threshold, it is marked as an outlier.

[0029] For detected anomalies, noise suppression is achieved using interpolation or predictive replacement methods.

[0030] Furthermore, establishing the mapping relationship between environmental disturbances and prediction reliability includes:

[0031] The interference data collected by the environmental interference sensor group is subjected to feature extraction to obtain the interference feature vector;

[0032] By training with historical data, a mapping relationship between interference features and prediction errors is established;

[0033] The impact of different types of interference on the prediction results is quantified into an influence coefficient matrix;

[0034] The predicted compensation value is calculated based on the influence coefficient matrix.

[0035] Furthermore, the uncertainty quantification framework is constructed based on evidence theory and includes:

[0036] Define a decision framework that includes possible states or categories;

[0037] Calculate the basic probability assignment for each source of evidence;

[0038] Calculate the belief function based on the basic probability assignment;

[0039] Calculate the likelihood function;

[0040] The prediction results are represented as intervals, and the uncertainty is calculated as the width of the prediction interval.

[0041] Furthermore, the step of dynamically adjusting the prediction model parameters and structure includes:

[0042] Calculate the sensitivity matrix of the model parameters to the interference features;

[0043] Calculate the parameter adjustment amount based on the sensitivity matrix and the current interference situation;

[0044] Choose model structures of different complexities based on the reliability level;

[0045] The model's ability to capture long-term dependencies and gradient flow efficiency are enhanced by employing residual connection and recursive connection structures.

[0046] Furthermore, the steps for generating the dynamic control strategy include:

[0047] Based on the reliability index of the prediction results, the safety margin is adaptively calculated;

[0048] Based on the health status prediction results and safety margin, optimize the battery charge and discharge control strategy;

[0049] Optimize load distribution and thermal management strategies based on health status predictions and reliability assessments;

[0050] Implement preventative maintenance and adaptive correction measures for abnormal conditions.

[0051] This invention provides a dynamic management and control system for energy storage batteries, used to execute the aforementioned dynamic management and control method for energy storage batteries, comprising:

[0052] The data acquisition and preprocessing module is used to acquire battery status and environmental interference information through a multimodal sensor network and perform signal noise reduction processing;

[0053] The feature extraction and noise reduction module is used to analyze battery health status features and filter abnormal noise using a multi-layer temporal convolutional network.

[0054] The reliability assessment module is used to construct a correlation model between environmental disturbances and prediction accuracy and to quantify prediction uncertainty based on evidence theory.

[0055] The adaptive inference module is used to dynamically optimize the prediction algorithm structure and parameters based on real-time interference and reliability indicators.

[0056] The management strategy generation module is used to calculate the safety margin and generate differentiated management plans based on the health status prediction results and their confidence levels.

[0057] The beneficial effects of this invention are as follows: by using multimodal signal acquisition and preprocessing technology, it effectively captures battery status and environmental interference information, providing a high-quality data foundation for health status prediction, and the signal-to-noise ratio is improved compared with traditional methods;

[0058] By employing a multi-layer temporal convolutional network and the Mahalanobis distance anomaly detection method, efficient feature extraction and noise suppression are achieved, improving the resistance to noise and interference signals and enhancing the accuracy of microcirculation discharge pattern recognition.

[0059] By constructing a mapping model between interference and reliability and a framework for quantifying uncertainty in evidence theory, a quantitative assessment of the reliability of prediction results is achieved, extending the prediction results from point estimation to interval representation, providing a basis for reliability assessment for decision-making, and providing early warning of sudden interference.

[0060] The dynamic model adjustment mechanism based on interference conditions and prediction reliability enables the system to adapt to different operating conditions, improving adaptability while reducing computational resource consumption and supporting millisecond-level response.

[0061] By adopting adaptive safety margin design and differentiated control strategies, the system effectively avoids risks such as overcharging and discharging while ensuring system performance, reducing the incidence of safety risk events and extending the overall lifespan of the battery pack. Attached Figure Description

[0062] Figure 1 This is a flowchart of a dynamic management and control method for energy storage batteries according to the present invention;

[0063] Figure 2 This is a flowchart of the multimodal signal acquisition and preprocessing steps of the present invention;

[0064] Figure 3 This is a detailed flowchart of the temporal convolution feature extraction and noise suppression steps of the present invention;

[0065] Figure 4 This is a detailed flowchart of the interference reliability mapping and uncertainty quantification steps of the present invention;

[0066] Figure 5 This is a detailed flowchart of the adaptive reasoning and decision optimization steps of the present invention;

[0067] Figure 6 This is a detailed flowchart of the dynamic control strategy generation steps of the present invention. Detailed Implementation

[0068] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0069] At least one embodiment of the present invention discloses a method for dynamic management and control of energy storage batteries, such as... Figures 1 to 5 As shown, it includes:

[0070] Step 1: Collect energy storage battery status data and environmental interference data through a multimodal sensor network, and perform noise reduction processing on the collected signals to obtain a multimodal feature set;

[0071] This step involves collecting data related to the energy storage battery and environmental interference data through a multimodal sensor network, and then performing noise reduction processing on the collected signals to extract feature information in different frequency domains.

[0072] Step 1.1, Multimodal data acquisition;

[0073] Construct a multimodal sensor network, which includes the following two types of sensor groups;

[0074] A battery status sensor array is used to collect parameters related to the battery's health status, including but not limited to:

[0075] Voltage sensors monitor the voltage of individual battery cells and the overall battery voltage;

[0076] Current sensor to monitor charging and discharging current;

[0077] Temperature sensor to monitor battery surface and ambient temperature;

[0078] Impedance measurement unit collects electrochemical impedance spectroscopy data of the battery.

[0079] An environmental interference sensor array is used to collect interference signals that may affect battery health status prediction, including but not limited to:

[0080] Electromagnetic field strength sensor to detect changes in the intensity of the environmental electromagnetic field;

[0081] Environmental noise sensors monitor acoustic and electrical noise;

[0082] Vibration sensors monitor mechanical vibration interference;

[0083] A humidity sensor to monitor changes in ambient humidity.

[0084] All sensor data are collected synchronously through the data acquisition unit. The sampling frequency is configured according to the variation characteristics of different parameters. For example, the sampling frequency of voltage and current is usually 1-10kHz, the sampling frequency of temperature is 1-10Hz, and the impedance measurement frequency is 0.1-0.01Hz.

[0085] Step 1.2, signal preprocessing and feature separation;

[0086] The acquired raw signal is preprocessed, and multi-scale wavelet decomposition is used for signal denoising and feature separation. The specific steps are as follows:

[0087] Signal noise reduction processing is performed on the time-series data collected by each sensor. Multi-scale wavelet decomposition is applied to decompose the signal into wavelet coefficients of different scales:

[0088] ;

[0089] in For a moment wavelet basis functions, Indicates the scale parameter. Indicates the translation parameter. For the corresponding wavelet coefficients, For a moment Time series data, This indicates integration.

[0090] Thresholding: Applying a threshold function to wavelet coefficients to suppress noise.

[0091] ;

[0092] in The threshold value is determined adaptively based on the signal-to-noise ratio. These are the processed wavelet coefficients; This is a threshold function used to suppress noise.

[0093] Signal reconstruction: The denoised signal is reconstructed using inverse wavelet transform.

[0094] ;

[0095] in For the reconstructed signal, This represents the summation operation. and These are the indices of the scale parameter and the translation parameter, respectively. Represents the processed wavelet coefficients. Indicates time Scale parameter is Translation parameters The wavelet basis functions.

[0096] Frequency domain feature separation decomposes a signal into three frequency domain feature sets: high frequency, mid frequency, and low frequency, each corresponding to different physical meanings.

[0097] Low-frequency feature set This corresponds to changes in the basic state of the battery (such as SOC, SOH, etc.), with a frequency range of 0-0.1Hz.

[0098] Intermediate frequency feature set Corresponding to the dynamic characteristics of battery charging and discharging, the frequency range is 0.1-10Hz;

[0099] High-frequency feature set This corresponds to noise and interference information, with a frequency range above 10Hz.

[0100] Through the above multimodal signal acquisition and preprocessing, a multimodal feature set after noise reduction is output. This provides a high-quality data foundation for subsequent feature extraction and analysis.

[0101] Step 2: Use a multi-layer temporal convolutional network to process the preprocessed multimodal feature set, extract the temporal features of battery health status, and simultaneously achieve abnormal noise detection and suppression.

[0102] This step takes the multimodal feature set output from step 1 as input, uses a multi-layer temporal convolutional network to process these preprocessed multimodal feature data, extracts multi-scale temporal features of battery health status, and simultaneously achieves abnormal noise detection and suppression, and identifies micro-cycle discharge mode features.

[0103] Step 2.1, Construction of a multi-layer temporal convolutional network;

[0104] A multi-layer temporal convolutional network integrating causal convolution and dilated convolution is constructed. This network contains the following key components:

[0105] Causal convolutional layers ensure that the model only uses information from before the current time step, avoiding information leakage. Their mathematical expression is:

[0106] ;

[0107] in For output features, For convolution kernel parameters, The kernel size is [size]. for Input at any time This is the time offset. This indicates a summation operation.

[0108] Dilated convolutional layers, by inserting holes between convolutional kernel elements, expand the receptive field and capture long-term dependencies; they are defined as follows:

[0109] ;

[0110] in The expansion rate increases exponentially with the number of network layers. , (for layer indexes) To expand the output features of the convolution, For convolution kernel parameters, for Input at any moment.

[0111] Residual connectivity structures are used to optimize gradient flow and avoid the gradient vanishing problem.

[0112] ;

[0113] in For the first The output of the layer, For the first The nonlinear transformation function of the layer, For the first The output of the layer.

[0114] Skip connections connect features at different levels, enhancing multi-scale feature fusion capabilities:

[0115] ;

[0116] in For the first Layer weight coefficients, For the first The output of the layer, Features following skip connections This indicates a summation operation.

[0117] Step 2.2, Detection of abnormal noise at the distance to Maharanobis;

[0118] The extracted features are then analyzed using the Mahalanobis distance method to detect anomalous noise.

[0119] Calculate the mean vector and covariance matrix of the features:

[0120] ;

[0121] ;

[0122] in For feature vectors, For the sample size, It is the mean vector. Let covariance matrix be the variance matrix. This represents the summation operation. For sample index, express transpose, Subtract the mean vector from the feature vector.

[0123] Calculate the Mahalanobis distance for each sample point:

[0124] ;

[0125] in For sample points The distance to Maharanobis It is the inverse of the covariance matrix. This represents the square root operation.

[0126] Set adaptive threshold ,when When this happens, it is marked as an anomaly.

[0127] For detected outliers, noise suppression is performed using interpolation or predictive replacement methods:

[0128] ;

[0129] in For interpolation functions, The size of the time window. express arrive Data at any given time express arrive Data at any given time For the corrected data, for The raw data at that moment for The distance to Maharanobis The threshold value is used.

[0130] Step 2.3, Multi-head self-attention mechanism and microcirculation discharge pattern recognition;

[0131] A multi-head self-attention mechanism was introduced to identify microcirculation discharge pattern characteristics;

[0132] Construct the query, key, and value matrix:

[0133] ;

[0134] ;

[0135] ;

[0136] in , , These represent the weight matrices for the query, key, and value, respectively. These are input features. It is a query matrix. It is a key matrix. It is a value matrix.

[0137] Calculate attention weights:

[0138] ;

[0139] in It is the dimension of the key vector. yes transpose, It is a normalized exponential function. It's a loss of attention.

[0140] Implementing a multi-head attention structure:

[0141] ;

[0142] ;

[0143] in It's about the number of heads to focus on. It is the output weight matrix. It is multi-headed attention output. This indicates a splicing operation. It is the first The output of each attention head, , , They represent the first A weight matrix for each attention head's query, key, and value. Indicates the index of the attention head.

[0144] Adaptive weighting of features is achieved through a channel attention mechanism:

[0145] ;

[0146] ;

[0147] in For channel The weighting coefficients, For channel Features For activation function, It is a multilayer perceptron. This is a global average pooling operation. These are the weighted features.

[0148] Through the above-described temporal convolutional feature extraction and noise suppression steps, a multi-scale temporal feature representation of the battery health status is output. and the characteristics of the identified microcirculation discharge modes This provides high-quality feature input for subsequent steps.

[0149] Step 3: Based on the time-series characteristics of environmental disturbance data and battery health status, establish the mapping relationship between environmental disturbance and prediction reliability, apply evidence theory to construct an uncertainty quantification framework, express the prediction results as confidence intervals and calculate the reliability index.

[0150] This step takes the environmental disturbance data collected in step 1 and the health status features extracted in step 2 as inputs, establishes a mapping relationship between environmental disturbance and prediction reliability, applies evidence theory to construct an uncertainty quantification framework, and extends the prediction results from point estimation to confidence interval representation, providing a basis for reliability assessment for decision-making.

[0151] Step 3.1, Construction of the interference and reliability mapping model;

[0152] Construct a mapping relationship between interference patterns and prediction reliability, and quantify the impact of different interferences on prediction results:

[0153] Interference feature extraction: Features are extracted from the interference data collected by the environmental interference sensor group to obtain the interference feature vector;

[0154] ;

[0155] in Represents the set of interference feature vectors. , , They represent the 1st, 2nd, and 3rd respectively. One interference feature, This represents the total number of interference features.

[0156] Construct an interference impact model, train it using historical data, and establish a mapping relationship between interference characteristics and prediction errors:

[0157] ;

[0158] in Indicates reliability assessment metrics, The mapping function can be implemented using a multilayer perceptron:

[0159] ;

[0160] in and The first Layer weights and biases For activation function, For the number of network layers, As the input of the interference feature vector, The reliability evaluation metric is the output of the mapping function. Indicates the index of the network layer. This is the output layer weight matrix. For output layer bias, This is the first layer weight matrix. For the first layer bias, This is the penultimate layer weight matrix. This is the second-to-last layer of bias.

[0161] Impact quantification quantifies the degree of influence of different types of disturbances on the prediction results into an impact coefficient matrix. :

[0162] ;

[0163] in Indicates the first Class-A interference on the first The degree of influence of each predictive indicator is calculated using the Jacobian matrix. express right The partial derivatives of .

[0164] Interference compensation calculation, based on the influence coefficient matrix, calculates the predicted compensation value:

[0165] ;

[0166] in The predicted compensation value is used to adjust the original prediction results. For the first The influence coefficient vector of interference-like phenomena For the first Eigenvalues ​​of interference-like features This represents the summation operation. This represents the total number of interference types.

[0167] Step 3.2, an uncertainty quantification framework based on evidence theory;

[0168] An uncertainty quantification framework is constructed using evidence theory to achieve an interval representation of prediction results:

[0169] Construct a recognition framework and define a decision framework:

[0170] ;

[0171] in Represents the set of decision frames. , , They represent the 1st, 2nd, and 3rd respectively. One possible state or category The total number of states or categories.

[0172] Basic probability allocation calculation, for each source of evidence (e.g., different sensors, different feature sets), calculate their basic probability allocation :

[0173] ;

[0174] in for a subset of As a source of evidence The credibility weights are provided by the interference and reliability mapping model. As a source of evidence Pair subset Support This represents the summation operation. express All subsets of the decision frame (i.e. subsets of the decision frame, representing possible combinations of states or categories). As a source of evidence Pair subset The basic probability allocation.

[0175] Belief function calculation: The belief function is calculated based on the basic probability assignment.

[0176] ;

[0177] in The basic probability assignments after merging are obtained using Dempster's merging rule. For a subset belief level This represents the summation operation. express All subsets of.

[0178] Likelihood function calculation:

[0179] ;

[0180] in for The supplement, For a subset likelihood To The degree of belief in the supplement. This represents the summation operation. express All subsets of.

[0181] Interval representation generation, which represents the prediction results as intervals. Rather than a single-point estimate, the uncertainty is calculated as follows:

[0182] ;

[0183] in For a subset Uncertainty, For likelihood, For belief level.

[0184] Step 3.3, Real-time interference impact assessment and reliability index calculation;

[0185] Based on the currently detected interference and historical model performance, the reliability of the prediction results is evaluated in real time:

[0186] Interference status detection, based on environmental interference sensor data, detects the current interference status:

[0187] ;

[0188] in The disturbance state vector, This is the interference data currently being collected. This is the interference detection function.

[0189] Reliability index calculation: Taking into account disturbance conditions, historical model performance, and prediction uncertainty, the reliability index is calculated as follows:

[0190] ;

[0191] in This is a reliability index, with a value range of [value missing]. , , , These represent the weights for the impact of disturbance states, historical accuracy, and uncertainty, respectively. For interference sensitivity parameters, Let the norm of the disturbance state vector be denoted as . This represents the model's historical accuracy under similar conditions. To predict uncertainty.

[0192] Reliability level classification divides the reliability index into different levels for subsequent decision-making adjustments:

[0193] ;

[0194] in For reliability level, Reliability index;

[0195] Through interference and reliability mapping and uncertainty quantification steps, the confidence interval representation of the prediction results is output. and reliability index This provides a basis for reliability assessment for subsequent adaptive decision-making.

[0196] Step 4: Based on the interference situation and prediction reliability, dynamically adjust the prediction model parameters and structure to generate health status prediction results and their reliability assessment.

[0197] This step integrates the outputs of the first three steps. Based on the interference information obtained in step 3 (the original interference data from step 1) and the prediction reliability assessment, the prediction model parameters and structure are dynamically adjusted. Combined with the multimodal feature information extracted in step 2, optimized health status prediction results and reliability assessments are generated.

[0198] Step 4.1, Adaptive model parameter adjustment;

[0199] Based on the interference conditions and reliability index, the parameters and structure of the prediction model are dynamically adjusted:

[0200] Model parameter sensitivity analysis, calculation of model parameters:

[0201] ;

[0202] in Represents the set of model parameters. , , They represent the 1st, 2nd, and 3rd respectively. Each model parameter This represents the total number of model parameters.

[0203] Relative to interference characteristics Sensitivity matrix :

[0204] ;

[0205] in, Indicates the index of the model parameter. Index representing interference features, Indicates the first The model parameter affects the first... Sensitivity to interference features; The model loss function, For the first Each model parameter As a feature of interference, For elements of the sensitivity matrix, This indicates that the loss function is applied to the parameters. and interference characteristics The second-order partial derivatives of .

[0206] Parameter adaptive adjustment is performed by calculating the parameter adjustment amount based on the sensitivity matrix and the current interference situation.

[0207] ;

[0208] in To adjust the step size, This represents the current interference characteristic value. For reliability index, For parameter adjustment amount, This represents the summation operation. For the interference feature dimension, Let be an element of the sensitivity matrix, representing the th The model parameter affects the first... Sensitivity of each interference feature.

[0209] Adaptive model structure switching: Selecting model structures of different complexities based on reliability levels.

[0210] ;

[0211] in , and These represent model structures of high, medium, and low complexity, respectively. For reliability level, The selected model.

[0212] Step 4.2, Long-term dependency enhancement and gradient flow optimization;

[0213] The model's ability to capture long-term dependencies and gradient flow efficiency are enhanced by employing residual and recursive connection structures:

[0214] Recursive join structure: Introducing recursive joins to capture long-term dependencies:

[0215] ;

[0216] in This represents the hidden state at the current moment. For the current input, This is the hidden state from the previous moment. For nonlinear transformation functions, gated recursive units (GRUs) can be used for implementation:

[0217] ;

[0218] ;

[0219] ;

[0220] ;

[0221] in To update the gate, control the degree to which the hidden state information from the previous moment is retained; To reset the gate, control the degree of influence of the previous hidden state on the current candidate hidden state; This represents a candidate hidden state, indicating new information at the current moment. The current hidden state is a weighted combination of the previous hidden state and the candidate hidden states. This represents the hidden state from the previous moment, containing historical information; The input features at the current time; Hadamard product (element-wise multiplication) represents the multiplication of elements at corresponding positions in a vector or matrix; The activation function maps the input to a range of 0 to 1. It is the hyperbolic tangent activation function; The parameters used to update the gate weight matrix are learned to update the control information. The weight matrix for resetting the gate is used to learn the parameters that control the resetting of historical information; The weight matrix represents the candidate hidden states and is used to learn new information. This indicates that the hidden state of the previous moment will be removed. and current input Concatenate along the feature dimension; This represents the forgetting factor, which controls the degree to which hidden state information from the previous moment is discarded.

[0222] Residual connectivity optimization: Applying residual connectivity in multilayer structures to optimize gradient flow.

[0223] ;

[0224] in Normalization operations on the representation layer improve training stability; For the first The nonlinear transformation function of the layer, For the first The output of the layer, For the first The output of the layer.

[0225] Gradient clipping to prevent gradient explosion:

[0226] ;

[0227] in For the gradient of the parameters, The cropping threshold, For the clipping function, restrict the gradient value to... Within the range.

[0228] Step 4.3, Multimodal Feature Fusion and Health Status Prediction;

[0229] A nonlinear feature fusion algorithm is used to integrate multimodal information to form the final health status prediction result:

[0230] Weighted feature fusion: Weighted fusion of features from different modalities.

[0231] ;

[0232] in The characteristics after fusion Indicates the first One feature mode, The corresponding weight coefficients are automatically determined by the attention mechanism. This represents the summation operation. Number of characteristic modes:

[0233] ;

[0234] ;

[0235] in and Let represent the query vector and feature transformation matrix in the attention mechanism, respectively; both are learnable parameters. Represents an exponential function. The hyperbolic tangent activation function is used. express transpose, Features Attention scores are calculated. Through this attention mechanism, the model can automatically learn the importance weights of different feature modalities, making the fusion process more intelligent and adaptive.

[0236] Cross-modal feature complementarity: Establishing a cross-modal feature correlation matrix :

[0237] ;

[0238] in For each element in the correlation matrix, the element represents the first element. The first characteristic mode and the second The degree of correlation between feature modes Representing the eigenvector transpose, For the first The first characteristic mode and the second The mapping matrix between the feature modes.

[0239] Final health status prediction. By integrating the above features, a health status prediction result is generated:

[0240] ;

[0241] in For predictive values ​​of health status, For the prediction function, consider fused features, cross-modal correlations, and reliability indices. The characteristics after fusion This is a cross-modal feature correlation matrix. This is a reliability index.

[0242] Reliability-perceived interval prediction. The prediction interval width is adjusted based on the reliability index.

[0243] ;

[0244] in For the prediction interval, To predict the lower bound, To predict the upper bound, For predictive values ​​of health status, The basic interval width parameter, This is a reliability index; the lower the reliability, the wider the range. and These represent subtraction and addition operations, respectively.

[0245] Through adaptive reasoning and decision optimization steps, the system outputs health status predictions that take into account the impact of disturbances and reliability assessments, including predicted values. and prediction interval This provides a decision-making basis for the final dynamic control strategy.

[0246] Step 5: Based on the health status prediction results and reliability assessment, generate a dynamic management and control strategy for energy storage batteries with a safety margin.

[0247] This step uses the health status prediction results generated in step 4 and their reliability assessment (step 3). Using the battery's real-time status data and environmental conditions data obtained in step 1 as the core input, a dynamic management strategy for energy storage batteries with a safety margin is generated to ensure system performance while avoiding safety risks such as overcharging and discharging.

[0248] Step 5.1, calculate the safety margin based on reliability;

[0249] Based on the reliability index of the prediction results, the safety margin is adaptively calculated:

[0250] Mapping relationship between safety margin and reliability:

[0251] ;

[0252] in For safety margin, Based on the safety margin parameter, For reliability index, and These represent the amplification factor and nonlinear adjustment parameter, respectively, indicating the degree of influence of control reliability on the safety margin. , , express of Power of 1.

[0253] Integration of forecast range and safety margin:

[0254] ;

[0255] in The prediction interval after considering the safety margin, To safely predict the lower bound, To safely predict the upper bound, This is the lower bound of the original prediction. This is the upper bound of the original prediction. For safety margin.

[0256] Step 5.2, Optimize the charging and discharging strategy;

[0257] Based on health status prediction results and safety margins, optimize the battery charge and discharge control strategy:

[0258] Charging cutoff conditions adjusted:

[0259] ;

[0260] in This is the adjusted state of charge cutoff. For the maximum permissible state of charge, This is the nominal state of charge cutoff. To address the security margins of the SOC, This represents the function that takes the minimum value.

[0261] Discharge cutoff condition adjustment:

[0262] ;

[0263] in This is the adjusted discharge cutoff state of charge. For the minimum permissible state of charge, This is the nominal discharge cutoff state of charge. To address the security margins of the SOC, This represents the function that takes the maximum value.

[0264] Charge and discharge power limits:

[0265] ;

[0266] ;

[0267] in This is the adjusted maximum charging power. This is the adjusted maximum discharge power. This is the nominal maximum power. This is the power adjustment coefficient. For reliability index, This represents the power adjustment factor.

[0268] Step 5.3, Load Distribution and Thermal Management Strategy;

[0269] Optimize load distribution and thermal management strategies based on health status predictions and reliability assessments:

[0270] For multi-battery systems, the load distribution strategy allocates the load based on health status and reliability:

[0271] ;

[0272] in To be assigned to the The load ratio of each battery For the first The health status of each battery For the first The predicted reliability index of a battery. It is the sum of the products of all battery health status and reliability indices. For total load demand, For the number of batteries, The summation variable, representing the battery number, is used to calculate the sum of the products of the health status and reliability index of all batteries. This formula ensures that batteries with better health and more reliable predictions can handle more load, thereby optimizing overall system performance and extending battery pack lifespan.

[0273] Thermal management strategy adjustment:

[0274] ;

[0275] in For cooling intensity, Based on cooling intensity, To adjust the amount, This is the reliability index for temperature prediction. This indicates the degree of unreliability in temperature prediction.

[0276] Step 5.4, Fault prevention measures and adaptive correction;

[0277] Implement preventative maintenance and adaptive correction measures for abnormal conditions:

[0278] Anomaly detection is based on the deviation between the predicted interval and the actual measured value to detect abnormal states.

[0279] ;

[0280] in These are abnormal scores. These are actual measured values. For predicted values, To be the upper bound of the safe prediction interval, To provide a lower bound for the safe prediction interval, This represents the absolute value of the difference between the actual value and the predicted value. This represents the half-width of the safe prediction interval.

[0281] Preventive maintenance triggered:

[0282] ;

[0283] in This is a maintenance flag; a value of 1 indicates that maintenance is required, and a value of 0 indicates that maintenance is not required. The threshold for abnormal scores. For reliability index, The minimum acceptable reliability index, Represents a logical OR operation.

[0284] Adaptive policy adjustment: Based on historical data and the current state, the control policy is adaptively adjusted.

[0285] ;

[0286] in For the control strategy in the next moment, As the current strategy, This is the best historical strategy. The learning rate controls the policy update speed. This represents the difference between the current strategy and the optimal strategy.

[0287] Through a dynamic management strategy generation step, a comprehensive management strategy is output, including charge / discharge strategies, load allocation, thermal management, and fault prevention measures. This strategy is adaptive and safe, dynamically adjusting the safety margin based on the reliability of the prediction results to effectively avoid safety risks such as overcharging and discharging, thereby improving the overall safety and lifespan of the battery system.

[0288] A dynamic management and control system for energy storage batteries, used to execute the aforementioned dynamic management and control method for energy storage batteries, includes:

[0289] The data acquisition and preprocessing module is used to acquire battery status and environmental interference information through a multimodal sensor network and perform signal noise reduction processing;

[0290] The feature extraction and noise reduction module is used to analyze battery health status features and filter abnormal noise using a multi-layer temporal convolutional network.

[0291] The reliability assessment module is used to construct a correlation model between environmental disturbances and prediction accuracy and to quantify prediction uncertainty based on evidence theory.

[0292] The adaptive inference module is used to dynamically optimize the prediction algorithm structure and parameters based on real-time interference and reliability indicators.

[0293] The management strategy generation module is used to calculate the safety margin and generate differentiated management plans based on the health status prediction results and their confidence levels.

[0294] Here, the present invention provides an implementation example:

[0295] A practical application test was conducted in a large industrial microgrid energy storage power station. This power station contains 100 lithium-ion battery packs with a total capacity of 10MWh, primarily used for peak shaving and valley filling of loads and grid frequency regulation. The power station is located within an industrial park, surrounded by large electrical equipment and substations, resulting in complex and variable environmental interference.

[0296] The testing environment has the following characteristics:

[0297] Electromagnetic interference sources: multiple high-power frequency converters, transformers, and motor equipment;

[0298] Temperature range: -10℃ to 45℃, with a daily temperature difference of up to 20℃;

[0299] Load characteristics: Intermittent high-power discharge and micro-cycle charge-discharge alternate;

[0300] Sudden interference: Unpredictable interference such as voltage fluctuations on the power grid side and sudden electromagnetic pulses.

[0301] Test system configuration:

[0302] Battery status sensors: Each battery group is equipped with 8 voltage, current and temperature sensors, and 1 impedance measurement unit;

[0303] Environmental interference sensors: 10 electromagnetic field strength sensors, 8 environmental noise sensors, 12 vibration sensors, and 6 humidity sensors;

[0304] Computing platform: Edge computing server, configured with Intel i7 processor and 16GB memory, used for model inference and management policy generation.

[0305] The noise reduction of the acquired raw signal by multi-scale wavelet decomposition technology effectively improves the signal quality.

[0306] The system continuously collects data over a 6-month operating cycle, and the data sampling frequency configuration is shown in Table 1:

[0307] Table 1: Data Acquisition Parameters and Preprocessing Results

[0308]

[0309] The denoised signal is decomposed into three feature sets: low-frequency, mid-frequency, and high-frequency, providing a high-quality data foundation for subsequent analysis. The system can automatically identify and mark outliers in the data, effectively avoiding prediction bias caused by abnormal data.

[0310] The Mahalanobis distance method was applied for abnormal noise detection, and the characteristics of microcirculation discharge modes were identified through a multi-head self-attention mechanism. Performance under different operating conditions is shown in Table 2.

[0311] Table 2: Results of Abnormal Noise Detection and Microcirculation Discharge Pattern Recognition

[0312]

[0313] Compared with traditional methods, this implementation method can maintain a high anomaly detection accuracy in complex interference environments, while the micro-circulation discharge pattern recognition accuracy is increased from 60% of the traditional method to an average of 89.9%.

[0314] By constructing a mapping model between interference and reliability, the impact of different types of interference on the prediction results is quantified, and a reliability index is generated. The impact coefficients and reliability indices for different interference types are shown in Table 3.

[0315] Table 3: Relationship between Interference Type and Reliability Index

[0316]

[0317] The system can assess environmental interference in real time and dynamically adjust the prediction interval width according to the degree of interference to ensure that the prediction interval can include the actual state value, thus providing a reliable assessment basis for decision-making.

[0318] Under different disturbance conditions, the system dynamically adjusts the parameters and structure of the prediction model through adaptive inference and decision optimization techniques to generate optimized health status prediction results and control strategies. Table 4 shows a comparison of the root mean square error (RMSE) of SOH predictions under different disturbance intensities, as well as the system's safety and lifespan extension effects.

[0319] Table 4: Analysis of Adaptive Inference Performance and System Security Benefits

[0320]

[0321] Under extremely strong interference, the prediction accuracy of this scheme is improved by 53.6% compared with traditional methods, the incidence of safety events is reduced by 65.8%, the expected battery life is extended by 20.2%, and the consumption of computing resources is reduced by 46.8%, which verifies the effectiveness of the adaptive inference mechanism.

[0322] This implementation method operated continuously for 6 months in an energy storage power station, collecting a large amount of operational data for long-term performance verification.

[0323] The health status prediction performance of this implementation method and the traditional method under different operating conditions was compared. The data came from statistical results under different disturbance environments during 6 months of continuous operation. As shown in Table 5:

[0324] Table 5: Comparison of Prediction Accuracy and Reliability During Long-Term Operation

[0325]

[0326] This implementation method outperforms traditional methods in all prediction metrics, especially in environments with strong interference, where the improvement in prediction performance is more significant, verifying the robustness and adaptability of the system.

[0327] After applying this implementation method, the incidence of system security incidents decreased, with an average reduction of 58.9%. The advantages of this method became more pronounced with increasing interference intensity, achieving a security incident reduction rate of 65.8% under extremely strong interference conditions. Simultaneously, the battery capacity degradation rate slowed, the expected lifespan was extended by an average of 18.5%, and maintenance costs were reduced by approximately 32.7%, fully validating the effectiveness of the reliability-based dynamic management strategy.

[0328] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for dynamic management and control of energy storage batteries, characterized in that, Comprise: Collecting energy storage battery state data and environmental interference data through a multi-modal sensor network, and performing noise reduction processing on the collected signals to obtain a multi-modal feature set, the noise reduction processing using multi-scale wavelet decomposition technology, comprising: Applying multi-scale wavelet decomposition to the time series data collected by each sensor; Applying an adaptive threshold function to the wavelet coefficients obtained by decomposition to suppress noise; Reconstructing the noise-reduced signal through inverse wavelet transform; Decomposing the signal into three frequency domain feature sets: high frequency, medium frequency and low frequency, corresponding to different physical meanings respectively; Using a multi-layer time series convolution network to process the pre-processed multi-modal feature set to extract the time series features of the battery health state, while achieving abnormal noise detection and suppression, the abnormal noise detection using the Mahalanobis distance method, comprising: Calculating the mean vector and covariance matrix of the features; Calculating the Mahalanobis distance of each sample point; Setting an adaptive threshold, when the Mahalanobis distance is greater than the threshold, marking it as an abnormal point; For the detected abnormal points, use interpolation or prediction replacement method for noise suppression; Based on the environmental interference data and the time series features of the battery health state, a mapping relationship between the environmental interference and the prediction reliability is established, an uncertainty quantification framework is constructed using evidence theory to represent the prediction result as a confidence interval and calculate the reliability index, the uncertainty quantification framework being based on evidence theory, comprising: Defining a decision framework containing possible states or categories; Calculating the basic probability assignment of each evidence source; Calculating the belief function based on the basic probability assignment; Calculating the likelihood function; Representing the prediction result as an interval and calculating the uncertainty as the width of the prediction interval; According to the interference condition and the prediction reliability, dynamically adjusting the prediction model parameters and structure to generate the health state prediction result and its reliability evaluation, the steps of dynamically adjusting the prediction model parameters and structure comprising: Calculating the sensitivity matrix of the model parameters with respect to the interference features; Based on the sensitivity matrix and the current interference condition, calculating the parameter adjustment amount; Selecting a model structure of different complexity according to the reliability level; Using residual connection and recursive connection structure to enhance the long-term dependence capture ability and gradient flow efficiency of the model; Based on the health state prediction result and the reliability evaluation, generating a dynamic management and control strategy for the energy storage battery with a safety margin, the steps of generating the dynamic management and control strategy comprising: Adaptively calculating the safety margin according to the reliability index of the prediction result; Based on the health state prediction result and the safety margin, optimizing the battery charge and discharge control strategy; According to the health state prediction and reliability evaluation, optimizing the load distribution and thermal management strategy; Implementing preventive maintenance and adaptive correction measures for abnormal states; Mapping relationship between safety margin and reliability: ； wherein is a safety margin, is a base safety margin parameter, is a reliability index, and respectively represent amplification factors and non-linear adjustment parameters that control the degree of influence of reliability on safety margin, , , represents power.​ 2. The method of claim 1, wherein, The multi-modal sensor network comprises a battery state sensor group and an environmental interference sensor group; the battery state sensor group is used to collect battery health state related parameters, including voltage, current, temperature and impedance data; the environmental interference sensor group is used to collect interference signals affecting health state prediction, including electromagnetic field intensity, environmental noise, vibration and humidity data.

3. The method of claim 1, wherein, The multi-layer time series convolution network comprises: Causal convolution layer, ensuring that the model only uses information before the current time; Dilated convolution layer, expanding the receptive field through dilated convolution, with the expansion rate increasing exponentially with the number of layers; Residual connection structure, optimizing gradient flow and avoiding gradient vanishing problem; Jump connection, connecting features of different levels to enhance multi-scale feature fusion capability.

4. The method of claim 1, wherein, The mapping relationship between environmental interference and prediction reliability includes: Feature extraction is performed on the interference data collected by the environmental interference sensor group to obtain an interference feature vector; Through historical data training, a mapping relationship between interference features and prediction errors is established; The influence of different types of interference on the prediction result is quantified as an influence coefficient matrix; Based on the influence coefficient matrix, a prediction compensation value is calculated.

5. A dynamic management system for energy storage batteries, comprising: A dynamic management method for energy storage batteries according to any one of claims 1-4, comprising: A data acquisition and preprocessing module for acquiring battery state and environmental interference information through a multi-modal sensor network and performing signal denoising processing; A feature extraction and denoising module for analyzing battery health state features and filtering abnormal noise using a multi-layer time series convolution network; A reliability evaluation module for constructing an environmental interference and prediction accuracy correlation model and quantifying prediction uncertainty based on evidence theory; An adaptive reasoning module for dynamically optimizing the prediction algorithm structure and parameters according to real-time interference conditions and reliability indicators; A management strategy generation module for calculating safety margins based on health state prediction results and their confidence and generating differentiated management schemes.

Images