Energy storage power supply residual life prediction method and system based on dynamic weight distribution

Abstract

The application discloses a kind of based on dynamic weight distribution energy storage power supply residual life prediction method and system, belong to energy storage power supply health management technical field.Pass through the voltage, current, temperature time series signal of the operation of energy storage power supply acquisition by multi-source sensor, degradation feature sequence is constructed using sliding window method;Degradation inflection point is identified by using curvature extreme detection and division stage, and the optimal model group is selected based on dynamic time warping algorithm matching historical degradation mode;Dynamic weight distribution is carried out to time step and feature dimension respectively using double-channel attention network, and the life prediction result can be generated by learning coefficient fusion spatio-temporal attention output;Combined with incremental learning mechanism, prediction residual and feature drift are monitored, and model parameters are updated using elastic weight consolidation algorithm and expanding degradation knowledge base.The application enhances the key information capture ability through spatio-temporal attention mechanism, and the model adaptability is optimized combined with incremental learning, which significantly improves the life prediction accuracy and working condition generalization performance of energy storage power supply.

Description

Technical Field

[0001] This invention relates to the field of energy storage power health management technology, specifically to a method and system for predicting the remaining lifespan of energy storage power based on dynamic weight allocation. Background Technology

[0002] In the field of lifetime prediction for electrochemical energy storage devices, traditional methods typically employ fixed models to model the entire lifecycle of the energy storage power source, making it difficult to adapt to the complex degradation characteristics in real-world applications. Existing technologies generally suffer from inaccurate identification of degradation stages. Because the degradation process of energy storage power sources often exhibits nonlinear characteristics (e.g., initial gradual decline, mid-term accelerated degradation, and final abrupt failure), most prediction models lack a dynamic correlation mechanism between degradation stages and model selection, leading to a significant increase in prediction bias after critical inflection points. Especially when energy storage power sources enter a rapid degradation phase, traditional time series prediction methods (such as autoregressive models and Kalman filters) often introduce lag errors due to the failure to adjust model parameters in a timely manner, causing the estimated remaining lifetime to deviate from the true value.

[0003] Existing feature fusion strategies typically employ fixed weight allocation, neglecting the dynamic changes in feature importance across different degradation stages. For example, temperature features may contribute little to lifetime prediction in the early stages of degradation, but become a critical indicator as failure approaches. However, traditional principal component analysis or static weighting methods cannot adaptively adjust feature weights. Furthermore, most systems rely on offline-trained static model libraries, lacking effective online learning mechanisms when encountering novel degradation patterns (such as sudden capacity drops or abrupt changes in cyclic stability). This results in insufficient model adaptability to new degradation features, leading to a continuous decline in prediction accuracy over long-term use. Summary of the Invention

[0004] The present invention aims to solve the above-mentioned technical problems by providing a method and system for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation. Its core lies in improving prediction accuracy through dynamic model selection and adaptive feature fusion.

[0005] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows:

[0006] A method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation includes the following steps:

[0007] S1: Collect multi-source sensor data during the operation of the energy storage power supply and extract features to generate a feature sequence containing historical degradation trajectories.

[0008] S2, the degradation stages of energy storage power sources are divided by inflection point detection algorithm, and prediction model groups are dynamically matched from the pre-built model library based on the similarity of degradation stages;

[0009] S3 utilizes a dual-channel attention network to dynamically allocate weights to the input feature sequence in terms of time and feature dimensions, generating remaining lifetime prediction results.

[0010] S4 continuously updates the prediction model parameters and degradation stage classification criteria through an incremental learning mechanism to achieve adaptive optimization of the model.

[0011] Furthermore, the multi-source sensing data mentioned in step S1 includes time-series signals such as voltage, current, temperature, and impedance spectrum. The specific steps of feature extraction and sequence generation are as follows:

[0012] S11. The voltage, current, and temperature time-series signals of the energy storage power supply during operation are acquired through a multi-source sensor array. The time-series signals are segmented using a sliding window method to ensure the effective preservation of the local continuity of degradation characteristics and the global evolution trend.

[0013] S12. A sliding window slides along the time axis with a fixed step size. The signal within the window is extracted in the time and frequency domains to generate a degenerate feature sequence X = [x1, x2, ..., x]. t ], where x t ∈R d Let represent the d-dimensional feature vector of the t-th window.

[0014] Step S2 includes:

[0015] (1) Based on historical degradation data, the curvature extreme value detection automatically identifies degradation inflection points. The degradation stage is divided by calculating the second derivative curvature extreme point of the degradation curve. When the value exceeds the preset threshold and is a local maximum, it is determined to be a degradation stage inflection point.

[0016] (2) Construct a degradation pattern classification tree based on the number of inflection points and the duration of each stage;

[0017] (3) The similarity between the real-time degradation curve and the historical pattern is calculated using the dynamic time warping algorithm. Let the historical reference sequence be Y = {y1, y2, ..., y...}. n Construct an m×t cumulative distance matrix D, and select the prediction model group corresponding to the historical pattern with the smallest cumulative distance as the optimal model group.

[0018] Step S3 includes:

[0019] Construct a temporal attention layer and calculate the importance weights of each time step in the feature sequence;

[0020] Construct a feature attention layer to evaluate the contribution weights of multi-source sensing features;

[0021] By dynamically combining time and feature attention outputs through learnable fusion coefficients, a comprehensive representation vector is generated, which is then mapped to the remaining lifetime prediction result through a fully connected layer.

[0022] The construction of the time attention layer includes:

[0023] The input feature sequence captures forward and backward temporal dependencies through a bidirectional gated recurrent unit. Where x t It is the multidimensional feature vector of the t-th time window. It is a forward hidden state. It is a backward hidden state;

[0024] The hidden states output by the bidirectional gated recurrent units are concatenated into H, projected into a query matrix Q, a key matrix K, and a value matrix V. The association weights at each time step are calculated using a multi-head self-attention mechanism, with the weight distribution of each attention head determined by scaled dot product attention. Where A n It is the weight distribution of the attention head. It is the scaling factor;

[0025] The outputs of multiple attention heads are concatenated and then linearly transformed to generate a dynamic representation Z in the time dimension. time .

[0026] The construction of the feature attention layer includes:

[0027] Let the input feature sequence be X, and D be the feature dimension. The original features are linearly combined using a learnable parameter matrix and matrix multiplication.

[0028] A multi-head parallel computing strategy is adopted, with each head independently evaluating the global contribution weight of the feature dimension;

[0029] Dynamic feature representation Z is generated through feature selection and linear transformation. feature .

[0030] Step S4 includes:

[0031] Monitor the residual distribution and feature space similarity index of the prediction results;

[0032] When model performance degradation is detected, an online incremental learning algorithm is used to update the model parameters, retaining historical model knowledge while incorporating the new degradation pattern into the stage classification standard library.

[0033] The monitoring features include spatial similarity indicators such as:

[0034] The deviation between real-time data and historical distribution is measured using Mahalanobis distance, employing D... m Quantify the distribution difference, D m =(μ new -μ old )∑ -1 (μ new -μ old );

[0035] Where μ new μ is the mean vector of the feature vectors within the current time window. old Let D be the feature mean vector of the historical training data, and Σ be the covariance matrix of the historical data. m When the threshold is exceeded, a significant feature drift is determined to have occurred.

[0036] A system for predicting the remaining lifetime of an energy storage power source based on dynamic weight allocation, comprising:

[0037] Data acquisition module: includes a multi-source sensor array and a signal preprocessing unit, used to acquire multi-source sensor data during the operation of the energy storage power supply and generate degradation feature sequences;

[0038] Degradation analysis module: integrates inflection point detection algorithm and pattern matching engine, used to classify the degradation stages of energy storage power and match the optimal prediction model group based on similarity;

[0039] Prediction and inference module: Deploys a dynamic weight allocation network and a multi-model fusion unit to dynamically allocate weights to feature sequences and generate remaining lifetime prediction results through a dual-channel attention network;

[0040] Model update module: Used to update the prediction model parameters and degradation stage classification criteria through incremental learning mechanism, realizing incremental learning and degradation knowledge base maintenance functions.

[0041] The dynamic weight allocation network includes:

[0042] The temporal attention subnetwork is implemented using gated recurrent units combined with a self-attention mechanism.

[0043] The feature attention subnetwork captures cross-modal feature associations through a multi-head attention layer;

[0044] The dynamic fusion layer uses differentiable weight parameters to balance the contribution ratio of spatiotemporal attention output.

[0045] The degradation analysis module includes:

[0046] The online inflection point detection unit is used to identify the degradation stage changes in real time based on sliding window curvature analysis, and to locate the inflection point of the historical degradation trajectory through the curvature extreme value detection algorithm.

[0047] The pattern similarity calculation unit is used to perform degradation trajectory matching using dynamic time warping and dynamic programming algorithms, and to calculate the similarity between the real-time degradation curve and historical patterns.

[0048] The model update module includes:

[0049] The feature drift detection unit is used to measure the deviation between real-time data and historical distribution using Mahalanobis distance;

[0050] Incremental learning units are used to update model parameters using an elastic weight consolidation algorithm, and regularization constraints are used to prevent the model from forgetting historical knowledge.

[0051] The degradation knowledge base maintenance unit is used to automatically update the classification criteria for degradation stages and the corresponding model parameters, incorporate new degradation patterns into the knowledge base, and train dedicated prediction models.

[0052] The predictive reasoning module includes:

[0053] The dual-channel attention network consists of a temporal attention subnetwork, a feature attention subnetwork, and a dynamic fusion layer. The temporal attention subnetwork is implemented using gated recurrent units combined with a self-attention mechanism. The feature attention subnetwork captures cross-modal feature associations through a multi-head attention layer. The dynamic fusion layer uses differentiable weight parameters to balance the contribution ratio of spatiotemporal attention output.

[0054] The multi-model fusion unit is used to call upon the best model group in the recall and dynamically allocate weights based on the historical accuracy of the models to generate the final lifetime prediction value.

[0055] By adopting the above system and method, the present invention has the following advantages:

[0056] 1. Improve prediction accuracy:

[0057] By using a dual-channel attention network to achieve dynamic weight allocation of the time and feature dimensions, it can accurately focus on key time steps and important features (such as temperature features at the end of degradation), solving the problem that traditional fixed weight allocation cannot adapt to the dynamic changes in feature importance.

[0058] Based on the similarity matching of degradation stages, the optimal model group is selected for different degradation stages (slow in the early stage, accelerated in the middle stage, and rapid failure in the late stage) to reduce the prediction bias caused by nonlinear degradation characteristics.

[0059] 2. Enhanced generalization performance under various operating conditions:

[0060] By employing a dynamic time warping algorithm to match historical degradation patterns, the model can effectively address the diverse degradation trajectories under different individual energy storage power sources and complex operating conditions, thereby improving its adaptability to different scenarios.

[0061] The incremental learning mechanism enables the model to continuously learn new degradation patterns (such as sudden capacity drops) through feature drift detection and elastic weight consolidation algorithms, thus avoiding a decline in prediction accuracy after long-term use.

[0062] 3. Achieve adaptive optimization:

[0063] By utilizing curvature extremum detection to automatically identify degradation inflection points and divide them into stages, and combining an incremental learning mechanism to dynamically update the degradation stage classification criteria, the system can respond in real time to the stage changes of the degradation process.

[0064] The model update process retains historical knowledge while expanding the degradation knowledge base to ensure rapid adaptation to new degradation patterns, forming a closed-loop mechanism of "detection-update-optimization".

[0065] 4. Comprehensively capture key information:

[0066] Multi-source sensors collect multi-dimensional data such as voltage, current, temperature, and impedance spectrum. Combined with the sliding window method, time-domain, frequency-domain, and statistical features are extracted to fully preserve the local continuity and global evolution trend of the degradation process.

[0067] In a dual-channel attention network, the temporal attention subnetwork captures temporal dependencies, while the feature attention subnetwork captures cross-modal feature associations. A dynamic fusion layer balances the contributions of both, enhancing the ability to capture key information.

[0068] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a flowchart of the energy storage power supply remaining lifetime prediction method based on dynamic weight allocation of degradation stage provided in the embodiments of the present invention;

[0071] Figure 2 This is a structural diagram of the dual-channel attention network according to an embodiment of the present invention;

[0072] Figure 3 This invention relates to the prediction effect of the energy storage power supply remaining lifetime prediction method based on dynamic weight allocation of degradation stage in different power sources.

[0073] Figure 4 This is a structural diagram of the energy storage power supply remaining lifetime prediction system based on dynamic weight allocation of degradation stage provided in an embodiment of the present invention. Detailed Implementation

[0074] Specific embodiments of the invention will now be described in detail. Although the invention is described in conjunction with these specific embodiments, it should be understood that the invention is not intended to be limited to these specific embodiments. Rather, these embodiments are intended to cover alternative, modified, or equivalent embodiments that may be included within the spirit and scope of the invention as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details. In other instances, well-known processes have not been described in detail in order not to unnecessarily obscure the invention.

[0075] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0076] The present invention will now be described in further detail with reference to the full text.

[0077] Combined with appendix Figures 1-4 A method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation of degradation stages includes the following steps:

[0078] S1: Collect multi-source sensor data during the operation of the energy storage power supply and extract features to generate a feature sequence containing historical degradation trajectories.

[0079] S2, the degradation stages of energy storage power sources are divided by inflection point detection algorithm, and prediction model groups are dynamically matched from the pre-built model library based on the similarity of degradation stages;

[0080] S3 utilizes a dual-channel attention network to dynamically allocate weights to the input feature sequence in terms of time and feature dimensions, generating remaining lifetime prediction results.

[0081] S4 continuously updates the prediction model parameters and degradation stage classification criteria through an incremental learning mechanism to achieve adaptive optimization of the model.

[0082] In a preferred embodiment of the energy storage power supply remaining lifetime prediction method based on dynamic weight allocation of degradation stage of the present invention, step S1 specifically includes:

[0083] Operating parameters, including voltage, current, temperature, and impedance spectrum time-series signals, are collected in real time through a multi-source sensor array deployed on the energy storage power equipment. The sensor data undergoes time-series alignment and denoising by a signal preprocessing unit to eliminate transmission delays and high-frequency noise interference. The preprocessing employs a wavelet threshold denoising algorithm to ensure signal fidelity.

[0084] The time-series signal is segmented using a sliding window method, with the window length and sliding step size set. Within each window, time-domain features, frequency-domain features, and statistical features are extracted to generate a multidimensional degradation feature sequence.

[0085] Specifically: S11: Collect multi-dimensional time-series signals such as voltage, current, and temperature during the operation of the energy storage power supply through a multi-source sensor array, and use the sliding window method to perform time-series segmentation on the original signals to ensure the effective preservation of the local continuity of degradation characteristics and the global evolution trend.

[0086] The S12 sliding window slides along the time axis with a fixed step size. The signal within the window is processed by time-domain and frequency-domain feature extraction to generate a degradation feature sequence X = [x1, x2, ..., x...]. t ], where x t ∈R d Let represent the d-dimensional feature vector of the t-th window.

[0087] Furthermore, the specific steps of the dynamic matching prediction model group based on the similarity of the degradation stage described in step S2 are as follows:

[0088] S21, in the degradation stage division, the inflection point of the historical degradation trajectory is located based on the curvature extreme value detection algorithm. The degradation curve y is calculated using the following formula. t The extreme points of the curvature of the second derivative of f(t) are used to identify the locations of abrupt changes in the degradation rate, thereby dividing the degradation process.

[0089]

[0090] Where κ(t) is the extreme point of the second derivative curvature. When κ(t) exceeds a preset threshold and is a local maximum, it is determined to be an inflection point of the degradation stage.

[0091] S22: The degradation patterns corresponding to each stage are matched with the real-time degradation trajectory using the Dynamic Time Warping (DTW) algorithm to calculate the cumulative distance under the optimal path to quantify the similarity. Let the historical reference sequence be Y = {y1, y2, ..., y...} n Construct the m×t cumulative distance matrix D:

[0092] D(i,j)=d(x i ,y j )+min{D(i-1,j),D(i,j-1),D(i-1,j-1)};

[0093] Where d(x) i ,y j )=|x i -y j| represents the Euclidean distance, and D represents the similarity between the two sequences. A smaller value indicates that the degradation patterns are more similar. The prediction model group corresponding to the historical pattern with the smallest cumulative distance is selected as the optimal model group.

[0094] For the time-series problem of remaining lifetime of energy storage power sources, traditional recurrent neural networks only consider the impact of historical data sequences on the current moment, which may overlook some implicit information. Therefore, the dual-channel attention network proposed in this invention uses the degradation feature sequence extracted from multiple sources as input, and realizes the prediction of remaining lifetime through a spatiotemporal dual attention mechanism and dynamic fusion.

[0095] In a specific implementation case, such as Figure 4 As shown, the dual-channel attention network structure includes an input layer, a BIGRU layer, a feature projection layer, a temporal attention sub-network, a feature attention sub-network, a dynamic fusion layer, and a fully connected layer.

[0096] Further, step S3 involves constructing a dual-channel attention network, specifically as follows:

[0097] S31, the input feature sequence first enters the temporal attention sub-network, where forward and backward temporal dependencies are captured by a bidirectional gated recurrent unit (BIGRU):

[0098]

[0099] x t It is the multidimensional feature vector of the t-th time window. It is a forward hidden state. It is a backward hidden state.

[0100] S32, the hidden states output by BIGRU are concatenated into H, which is then projected into a query matrix Q, a key matrix K, and a value matrix V. The association weights at each time step are calculated using a multi-head self-attention mechanism. The weight distribution of each attention head is determined by scaled dot product attention.

[0101]

[0102] A n It is the weight distribution of the attention head. It is a scaling factor used to control the order of magnitude of the dot product result.

[0103] S33, the outputs of multiple attention heads are concatenated and then linearly transformed to generate a dynamic representation Z in the time dimension. time This highlights the impact of key time segments on the degradation process.

[0104] Z time =[A1V1,A2V2,…,A n V n ];

[0105] S34, Let the input feature sequence be X, and D be the feature dimension. To capture the interaction relationships between features, a learnable parameter matrix W is designed. E ∈R D×D The original features are linearly combined using matrix multiplication:

[0106] X′=XW E ;

[0107] A multi-head parallel computing strategy is adopted, with each head independently evaluating the global contribution weight of the feature dimension:

[0108]

[0109] x′ is the t-th component of X′, w mj and b mj σ represents the weights and biases corresponding to feature j in the m-th attention head, and σ is the Sigmoid function.

[0110] Z is generated through feature selection and linear transformation. feature This enhances the ability to focus on key characteristics such as current and voltage.

[0111]

[0112] β m It is the feature weight vector of the m-th head, and ⊙ represents the element-wise multiplication broadcast along the feature dimension, i.e., the feature vector x at each time step. t by β m Weighted, Let be the projection matrix of the m-th head, which maps the weighted features to the subspace.

[0113] S35, the dynamic fusion layer achieves adaptive integration of spatiotemporal attention through learnable weight parameters α. The fusion coefficients are generated by concatenating spatiotemporal representations and then passing them through a gating mechanism.

[0114] α=σ(w α [Z time Z feature ]+b α );

[0115] Where w α and b α These are the corresponding weights and biases.

[0116] The ratio of the dynamic equilibrium time of this coefficient to the contribution of feature attention forms a comprehensive characterization:

[0117] H final =α·Z time +(1-α)·Z feature ;

[0118] Ultimately, the fully connected layer maps the synthesized representation to the remaining lifetime prediction:

[0119]

[0120] Where w p and b p These are the corresponding weights and biases.

[0121] Furthermore, the incremental learning mechanism described in step S4 continuously updates the prediction model parameters, including:

[0122] S41, First, to achieve continuous adaptation of the prediction model to new degradation patterns, a closed-loop optimization process is implemented, triggering parameter updates and knowledge base expansion through feature drift detection. Firstly, the system monitors the deviation of the input feature distribution from the historical baseline in real time, employing D... m Quantitative distribution differences:

[0123] D m =(μ new -μ old )∑ -1 (μ new -μ old );

[0124] Where, μ new μ is the mean vector of the feature vectors within the current time window. old Let D be the feature mean vector of the historical training data, and Σ be the covariance matrix of the historical data. m When the threshold θ is exceeded, a significant feature drift is determined, triggering the model update process.

[0125] S42, based on this, the above system uses the elastic weight consolidation algorithm to update the model parameters. Its core lies in protecting historical knowledge through regularization constraints. The loss function is designed as follows:

[0126]

[0127] Where Loss(θ) is the standard loss for the new data, θ i,old For the model parameters before the update, F i These are the diagonal elements of the Fisher information matrix, used to quantize the parameter θ. i The importance of historical tasks is represented by λ, a trade-off coefficient that controls the balance between old and new knowledge. Parameter updates are achieved through gradient descent.

[0128]

[0129] Where η is the learning rate (e.g., 0.001), and F is constrained. i The magnitude of the change effectively prevents the model from forgetting historical degradation patterns.

[0130] Furthermore, the specific steps for updating the classification criteria for the degradation stage described in step S4 are as follows:

[0131] The system continuously analyzes the degradation trajectory for novel inflection point patterns. When an inflection point combination not recorded in the historical classification tree is detected, a degradation knowledge base expansion process is initiated: first, a density clustering algorithm is used to group the new inflection point sequences, generating new degradation pattern categories; then, sufficient samples are collected to train a dedicated prediction model, which is added to the model library, and the classification tree node structure is reconstructed. Finally, the updated model group needs to be tested on the retained validation set to ensure that the update process does not introduce performance degradation.

[0132] In specific implementation of this invention, such as Figure 3 As shown, the remaining lifetime prediction method for energy storage power sources based on the dynamic weight allocation of degradation stages according to embodiments of the present invention is demonstrated for the prediction performance of the remaining lifetime of different energy storage power sources (power sources 1, 2, and 3). The three groups of power sources exhibit significantly different degradation trajectories, but the method can accurately identify the inflection points of the degradation stages of each power source and maintain high-precision prediction performance under their respective different degradation modes.

[0133] This invention also applies for a power storage remaining lifetime prediction system based on dynamic weight allocation of degradation stage, which is described below in conjunction with... Figure 2 The system described in this invention,

[0134] This system consists of four core modules.

[0135] Data acquisition module: includes a multi-source sensor array and a signal preprocessing unit, used to acquire multi-source sensor data during the operation of the energy storage power supply and generate degradation feature sequences;

[0136] Degradation analysis module: integrates inflection point detection algorithm and pattern matching engine, used to classify the degradation stages of energy storage power and match the optimal prediction model group based on similarity;

[0137] Prediction and inference module: Deploys a dynamic weight allocation network and a multi-model fusion unit to dynamically allocate weights to feature sequences and generate remaining lifetime prediction results through a dual-channel attention network;

[0138] Model update module: Used to update the prediction model parameters and degradation stage classification criteria through incremental learning mechanism, realizing incremental learning and degradation knowledge base maintenance functions.

[0139] The system employs a distributed architecture to dynamically identify the degradation stages of energy storage power sources and accurately predict their remaining lifespan. It supports real-time data stream processing and offline knowledge base updates, with modules working collaboratively to form a closed-loop optimization process.

[0140] The data acquisition module mainly consists of a module measurement unit and a data preprocessing unit. The module hardware includes voltage sensors, current sensors, temperature sensors, and an impedance spectrum measurement unit, covering multi-dimensional status monitoring of the energy storage power supply. The data preprocessing unit performs time-series alignment on the raw data, eliminates deviations caused by transmission delays in multi-source signals, and filters out high-frequency noise using a wavelet threshold denoising algorithm. The preprocessed data is segmented using a sliding window method, and time-domain features, frequency-domain features, and statistical features are extracted within the window, ultimately generating a multi-dimensional degradation feature sequence.

[0141] The degradation analysis module mainly consists of an inflection point detection unit and a pattern similarity calculation unit. First, the module uses the online inflection point detection unit to perform curvature extremum analysis on historical degradation trajectories, dividing the degradation process into stages. Then, the pattern similarity calculation unit employs a dynamic time warping algorithm to align the real-time degradation curve with reference trajectories in the historical pattern library, quantifying their similarity.

[0142] The prediction inference module achieves high-precision lifetime prediction through a dual-channel attention network and a multi-model fusion strategy. The dual-channel attention network consists of a temporal attention sub-network and a feature attention sub-network, generating a temporal dynamic representation. The feature attention sub-network evaluates the contribution of multi-source features using a learnable projection matrix and multi-head parallel weights, generating a feature dynamic representation. The dynamic fusion layer generates learnable weight coefficients through a gating mechanism, fusing the temporal and feature dynamic representations into a comprehensive representation. The multi-model fusion unit calls upon the recalled optimal model group and generates the final lifetime prediction value based on a weighted strategy (weights are dynamically allocated by the model's historical accuracy).

[0143] The model update module mainly consists of a feature drift detection unit, an incremental learning unit, and a degradation stage maintenance unit. The feature drift detection unit monitors the deviation of the input feature distribution from the historical baseline in real time based on Mahalanobis distance. When the deviation exceeds a threshold for three consecutive windows, the incremental learning process is triggered. The incremental learning unit uses an elastic weight consolidation algorithm to update the model parameters. Its loss function adds a regularization term to the standard loss to constrain the variation of key historical parameters and prevent the model from forgetting existing knowledge. The degradation stage maintenance unit is responsible for expanding new degradation patterns: when an unrecorded inflection point combination is detected, a similar algorithm is used to group the new inflection point sequences. After collecting sufficient samples, a dedicated model is trained and stored in the knowledge base. The updated model needs to be tested on a reserved validation set to ensure the system's adaptability and stability to new degradation patterns.

[0144] This invention collects voltage, current, and temperature time-series signals of energy storage power supply operation through multi-source sensors and constructs a degradation feature sequence using the sliding window method. It identifies degradation inflection points and divides the system into stages using curvature extremum detection, matches historical degradation patterns based on a dynamic time warping algorithm, and selects the optimal model group. A dual-channel attention network dynamically assigns weights to the time step and feature dimensions, and generates lifetime prediction results by fusing spatiotemporal attention outputs with learnable coefficients. An incremental learning mechanism monitors prediction residuals and feature drift, and an elastic weight consolidation algorithm updates model parameters and expands the degradation knowledge base. The system includes modules for data acquisition, degradation analysis, prediction inference, and model updating, achieving dynamic identification and adaptive modeling of degradation stages. This invention enhances the ability to capture key information through a spatiotemporal attention mechanism and optimizes model adaptability through incremental learning, significantly improving the accuracy of energy storage power supply lifetime prediction and its generalization performance under operating conditions.

[0145] The present invention and its embodiments have been described above. This description is not restrictive, and the embodiments shown throughout are only one of the embodiments of the present invention. The actual structure is not limited to this. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation, characterized in that, Includes the following steps: S1: Collect multi-source sensor data during the operation of the energy storage power supply and extract features to generate a feature sequence containing historical degradation trajectories. S2, the degradation stages of energy storage power sources are divided by inflection point detection algorithm, and prediction model groups are dynamically matched from the pre-built model library based on the similarity of degradation stages; S3 utilizes a dual-channel attention network to dynamically assign weights to the input feature sequence in terms of time and feature dimensions, generating remaining lifetime prediction results. S4 continuously updates the prediction model parameters and degradation stage classification criteria through an incremental learning mechanism to achieve adaptive optimization of the model; Step S2 includes: (1) Based on the curvature extreme value detection of historical degradation data, the degradation inflection point is automatically identified. The degradation stage is divided by calculating the curvature extreme point of the second derivative of the degradation curve. When it exceeds the preset threshold and is a local maximum, it is determined to be the inflection point of the degradation stage. (2) Construct a degradation pattern classification tree based on the number of inflection points and the duration of each stage; (3) The similarity between the real-time degradation curve and the historical pattern is calculated using the dynamic time warping algorithm. Let the historical reference sequence be... Construct an m×t cumulative distance matrix D, and select the prediction model group corresponding to the historical pattern with the smallest cumulative distance as the optimal model group; Step S3 includes: Construct a temporal attention layer and calculate the importance weights of each time step in the feature sequence; Construct a feature attention layer to evaluate the contribution weights of multi-source sensing features; By dynamically combining time and feature attention output through learnable fusion coefficients, a comprehensive representation vector is generated, which is then mapped to the remaining lifetime prediction result through a fully connected layer. The construction of the time attention layer includes: The input feature sequence captures forward and backward temporal dependencies through a bidirectional gated recurrent unit. , where x t It is the multidimensional feature vector of the t-th time window. It is a forward hidden state. It is a backward hidden state; The hidden states output by the bidirectional gated recurrent units are concatenated into H, projected into a query matrix Q, a key matrix K, and a value matrix V. The association weights at each time step are calculated using a multi-head self-attention mechanism, with the weight distribution of each attention head determined by scaled dot product attention. ,in It is the weight distribution of the attention head. It is the scaling factor; The outputs of multiple attention heads are concatenated and then linearly transformed to generate a dynamic representation Z in the time dimension. time .

2. The method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation according to claim 1, characterized in that: Step S1 includes: S11. Obtain voltage, current, and temperature time-series signals during the operation of the energy storage power supply through a multi-source sensor array, and segment the time-series signals using a sliding window method. S12. The sliding window slides along the time axis at fixed steps. The signal within the window is processed by time-domain and frequency-domain feature extraction to generate a degradation feature sequence.

3. The method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation according to claim 1, characterized in that: The construction of the feature attention layer includes: Let the input feature sequence be X, and D be the feature dimension. The original features are linearly combined using a learnable parameter matrix and matrix multiplication. A multi-head parallel computing strategy is adopted, with each head independently evaluating the global contribution weight of the feature dimension; Dynamic feature representation Z is generated through feature selection and linear transformation. feature .

4. The method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation according to claim 1, characterized in that: Step S4 includes: Monitor the residual distribution and feature space similarity index of the prediction results; When model performance degradation is detected, an online incremental learning algorithm is used to update the model parameters, retaining historical model knowledge while incorporating the new degradation pattern into the stage classification standard library.

5. The method for predicting the remaining lifetime of energy storage power sources based on dynamic weight allocation according to claim 4, characterized in that: The monitoring features include spatial similarity indicators such as: The deviation between real-time data and historical distribution is measured using Mahalanobis distance, and D is employed. m Quantify distribution differences, ; Where μ new μ is the mean vector of the feature vectors within the current time window. old Let D be the feature mean vector of the historical training data, and Σ be the covariance matrix of the historical data. m When the threshold is exceeded, a significant feature drift is determined to have occurred.

6. A system for predicting the remaining lifetime of an energy storage power source based on dynamic weight allocation, characterized in that, include: Data acquisition module: includes a multi-source sensor array and a signal preprocessing unit, used to acquire multi-source sensor data during the operation of the energy storage power supply and generate degradation feature sequences; Degradation analysis module: integrates inflection point detection algorithm and pattern matching engine, used to classify the degradation stages of energy storage power and match the optimal prediction model group based on similarity; Prediction and inference module: Deploys a dynamic weight allocation network and a multi-model fusion unit to dynamically allocate weights to feature sequences and generate remaining lifetime prediction results through a dual-channel attention network; Model update module: Used to update the prediction model parameters and degradation stage classification criteria through incremental learning mechanism, realizing incremental learning and degradation knowledge base maintenance functions.

7. The energy storage power supply remaining lifetime prediction system based on dynamic weight allocation according to claim 6, characterized in that: The dynamic weight allocation network includes: The temporal attention subnetwork is implemented using gated recurrent units combined with a self-attention mechanism. The feature attention subnetwork captures cross-modal feature associations through a multi-head attention layer; The dynamic fusion layer uses differentiable weight parameters to balance the contribution ratio of spatiotemporal attention output. The degradation analysis module includes: an online inflection point detection unit, used to identify degradation stage transitions in real time based on sliding window curvature analysis, and to locate inflection points in historical degradation trajectories using a curvature extreme value detection algorithm; and a pattern similarity calculation unit, used to match degradation trajectories using dynamic time warping and dynamic programming algorithms, and to calculate the similarity between real-time degradation curves and historical patterns. The model update module includes: The feature drift detection unit is used to measure the deviation between real-time data and historical distribution using Mahalanobis distance; Incremental learning units are used to update model parameters using an elastic weight consolidation algorithm, and regularization constraints are used to prevent the model from forgetting historical knowledge. The degradation knowledge base maintenance unit is used to automatically update the classification criteria for degradation stages and the corresponding model parameters, incorporate new degradation patterns into the knowledge base, and train a dedicated prediction model. The predictive reasoning module includes: The dual-channel attention network consists of a temporal attention subnetwork, a feature attention subnetwork, and a dynamic fusion layer. The temporal attention subnetwork is implemented using gated recurrent units combined with a self-attention mechanism. The feature attention subnetwork captures cross-modal feature associations through a multi-head attention layer. The dynamic fusion layer uses differentiable weight parameters to balance the contribution ratio of spatiotemporal attention output. The multi-model fusion unit is used to call upon the best model group retrieved and dynamically allocate weights based on the historical accuracy of the models to generate the final lifetime prediction value.

Images