Intelligent online capacity checking method and system for energy storage battery based on deep learning

Abstract

The invention discloses an intelligent online capacity checking method and system for an energy storage battery based on deep learning. The intelligent online capacity checking method comprises the following steps: firstly, acquiring a historical operation data set of a target storage battery and a target energy storage system; encoding the high-dimensional time sequence features through an auto-encoder, and extracting target battery state features; associated storage battery data are screened based on degradation track similarity measurement; constructing a target storage battery data set, inputting the target storage battery data set into the time sequence diagram convolutional hybrid network, cooperatively analyzing and matching preset features through a multi-head diagram attention mechanism, and outputting system-level degradation characterization and degradation attribution reports; and generating capacity check decision suggestions in combination with system-level representation and attribution results. According to the method, the collaborative degradation law of the battery system is dynamically captured through deep learning, online capacity checking without shutdown is realized, and the accuracy of health assessment and the scientificity of decision making are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and system for intelligent online capacity verification of energy storage batteries based on deep learning. Background Art

[0002] As a core component of new energy storage systems, the health of energy storage batteries directly impacts system safety and economic viability. Traditional battery capacity verification methods often rely on offline full-discharge testing, which requires system interruption and significant investment in manpower and resources. This leads to high costs and impacts on system continuity. Furthermore, manual analysis of historical operating data makes it difficult to accurately capture the temporal correlations of battery degradation and the synergistic effects of multiple batteries, resulting in inadequate and time-sensitive capacity decisions. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for intelligent online capacity verification of energy storage batteries based on deep learning.

[0004] In a first aspect, an embodiment of the present invention provides a method for intelligent online capacity verification of energy storage batteries based on deep learning, comprising:

[0005] Obtain a set of historical operating data of the target battery and target energy storage system;

[0006] Extracting target battery state features of the battery from an operating segment in a historical operating data set of the target battery and the target energy storage system; the target battery state features are obtained by encoding high-dimensional time series features of the battery through an autoencoder;

[0007] determining, based on a degradation trajectory similarity metric between the target battery and the target battery state characteristics of the respective operation segments, a battery data set associated with the target battery from the historical operation data set of the target energy storage system;

[0008] Based on the associated battery data set and the target battery, a target battery data set is obtained, and the target battery data set is input into a time series-graph convolution hybrid network; the time series-graph convolution hybrid network includes a preset battery state feature set, and the time series-graph convolution hybrid network is trained based on a historical operation data set of a training energy storage system;

[0009] Through the time series-graph convolution hybrid network, the preset battery state features corresponding to each battery in the target battery data set are matched in the preset battery state feature set, and a multi-source feature collaborative analysis based on the multi-head graph attention mechanism is performed on each matched preset battery state feature to output a system-level degradation representation and degradation attribution report of the target battery data set; the system-level degradation representation is used to determine the health state compatibility between the target battery and the target energy storage system, and combined with the degradation attribution report to generate capacity decision recommendations.

[0010] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0011] Compared with the existing technology, the beneficial effects provided by the present invention include: using a deep learning-based intelligent online capacity-based capacity-based method and system for energy storage batteries disclosed by the present invention, by obtaining a set of historical operating data of the target battery and the target energy storage system; encoding high-dimensional time series features through an autoencoder to extract target battery status features; screening related battery data based on degradation trajectory similarity metrics; constructing a target battery data set and merging the input time series graph convolution hybrid network, using a multi-head graph attention mechanism to collaboratively analyze and match preset features, outputting system-level degradation representation and degradation attribution reports; combining system-level representation and attribution results to generate capacity-based decision recommendations. This method dynamically captures the collaborative degradation laws of the battery system through deep learning, achieves online capacity-based ... BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0013] Figure 1 A schematic flow chart of the steps of a method for intelligent online capacity verification of energy storage batteries based on deep learning provided by an embodiment of the present invention;

[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0016] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0017] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of an intelligent online capacity control method for energy storage batteries based on deep learning provided in an embodiment of the present disclosure. The intelligent online capacity control method for energy storage batteries based on deep learning is introduced in detail below.

[0018] Step S201, obtaining a set of historical operating data of a target battery and a target energy storage system;

[0019] Step S202: extracting target battery state features of the battery from the historical operation data set of the target battery and the target energy storage system; the target battery state features are obtained by encoding the high-dimensional time series features of the battery through an autoencoder;

[0020] Step S203: determining an associated battery data set of the target battery from the historical operation data set of the target energy storage system based on a degradation trajectory similarity metric between the target battery and the target battery state characteristics of the operation segments;

[0021] Step S204: obtaining a target battery data set based on the associated battery data set and the target battery, and inputting the target battery data set into a time series-graph convolution hybrid network; the time series-graph convolution hybrid network includes a preset battery state feature set, and the time series-graph convolution hybrid network is trained based on a historical operation data set of a training energy storage system;

[0022] In step S205, the preset battery status features corresponding to each battery in the target battery data set are matched in the preset battery status feature set through the time series-graph convolution hybrid network, and a multi-source feature collaborative analysis based on a multi-head graph attention mechanism is performed on each matched preset battery status feature to output a system-level degradation representation and a degradation attribution report of the target battery data set; the system-level degradation representation is used to determine the health status compatibility between the target battery and the target energy storage system, and a capacity decision recommendation is generated in combination with the degradation attribution report.

[0023] In an embodiment of the present invention, for example, the energy storage system of a new energy power station is used as an application scenario. The system includes an energy storage unit (i.e., the target energy storage system) composed of 200 lithium iron phosphate batteries, which is mainly used to smooth out wind power fluctuations. Recently, operation and maintenance personnel discovered that the battery numbered B32 (i.e., the target battery) showed abnormal temperature rise during the charging and discharging process. It is necessary to use the deep learning-based intelligent online capacity evaluation method described in this embodiment to evaluate its health status and generate capacity decision recommendations. The following describes the specific implementation process in detail with the server as the execution subject:

[0024] First, the server collects historical operating data from the target battery and target energy storage system from the battery management system (BMS) of each battery through the IoT gateway deployed in the energy storage system. The target battery B32's historical operating data includes the last 12 months of minute-by-minute time series data, including voltage (range: 2.8V-3.6V), current (charge / discharge rate: 0.5C-2C), temperature (25°C-40°C), state of charge (SOC, 20%-90%), state of health (SOH, initially 98%), and internal resistance (initial 8mΩ). The target energy storage system's historical operating data covers the past three years of global operating logs for the same batch of 200 batteries (including B32). This data includes the number of charge / discharge cycles per battery (500-2000), deep discharge frequency (3-8 times per month), ambient temperature fluctuations (10°C-45°C), and system-level charge / discharge strategies (such as high-rate discharge to respond to peak / valley price fluctuations). The server aligns the above data by timestamp and stores it as a structured historical operation data set, where each "operation segment" is defined as 7 consecutive days of operation data (for example, January 1-7, 2023 is one segment, with a total of 52 segments).

[0025] Next, the server needs to extract the target battery status characteristics from the target battery and each operating segment. Specifically, for each operating segment of the target battery B32 and the other 199 batteries in the system (such as the January 2023 segment of B32 and the May 2022 segment of B15), the server first converts its multi-dimensional operating data into a "multi-dimensional battery fingerprint": the time dimension extracts the "voltage platform length" (for example, the duration of the 3.2V platform from SOC50% to 70%) and "temperature rise rate" (the rate at which the temperature rises from 25°C to 35°C within 1 hour of charging) of the daily charge and discharge curve; the statistical dimension calculates the weekly average internal resistance growth rate (such as 0.05mΩ / week) and the hysteresis voltage of the SOC-voltage curve (the difference between the discharge end voltage and the charging start voltage); the event dimension marks whether it has experienced abnormal operating conditions such as overcharging (voltage>3.65V), over-discharge (voltage<2.7V), and high-temperature storage (temperature>40°C for 24 hours). These fingerprints are concatenated in chronological order to form a "feature chain of the battery instance to be evaluated" (for example, the feature chain of B32 is [F1_202301, F1_202302, ..., F1_202312], where each F1 contains a fingerprint of 15 dimensions). The server then inputs this feature chain into a trained target feature compression encoder (an autoencoder optimized based on historical training data), which contains a preset high-dimensional time series feature set (a 200-dimensional degradation feature template extracted from 100,000 battery cell data). The encoder first matches the template most similar to each fingerprint in the current feature chain in the high-dimensional time series feature set (for example, the "shortened voltage platform length" feature of B32 is matched to the "initial capacity decay" template), and then performs cross-domain feature fusion (compressing 200 dimensions to 50 dimensions) on the matched high-dimensional features (such as capacity decay, internal resistance growth, SEI film thickening, etc.) through the fully connected layer. Finally, the fused features are mapped to the health factor space (SOH prediction value of 0-1) and the target battery status features are output (for example, the January 2023 segment features of B32 are [0.95, 0.08, 0.8], corresponding to the SOH prediction value, internal resistance growth rate, and thermal runaway risk level, respectively).

[0026] The server then calculates the degradation trajectory similarity based on the target battery and the target battery state characteristics of each operating segment to determine the associated battery data set. Specifically, the dynamic time warping (DTW) algorithm is used to measure the shape similarity of two time series characteristic curves (the smaller the DTW distance, the more similar the trajectories). For example, the characteristic trajectory of B32 is [0.95, 0.93, 0.91, ..., 0.85] (SOH decreases year by year), and the characteristic trajectory of B15 in the system from January 2020 to December 2021 is [0.96, 0.94, 0.92, ..., 0.86], with a DTW distance of 0.03 (high similarity); the characteristic trajectory of B47 from June 2019 to May 2020 is [0.98, 0.95, 0.90, ..., 0.80], with a DTW distance of 0.1 (moderate similarity). The server sorts the data from smallest to largest by DTW distance, selects the top 20 most similar operating segments (corresponding to 15 different batteries), and aligns these segments according to the "charge and discharge cycle phase" (that is, unifies them to the same cycle number node. For example, if the current cycle number of B32 is 1200, the segments of the associated battery are selected in the cycle number range of 1100-1300) to ensure consistency in the degradation stage. Finally, the associated battery data set (including operating data around 1200 cycles for 15 batteries) is obtained.

[0027] Subsequently, the server integrates the current operating data of the target battery B32 (1200 cycles in the last month) with the associated data set to form a target battery data set (aligned data of 16 batteries in total). This set contains the time series features of each battery (voltage, current, and temperature curves for the past 7 days), state features (target battery state features, such as SOH prediction value, internal resistance rate) and associated information (similarity score with B32, such as the similarity score of B15 is 95 points). The server inputs this data set into a trained time series-graph convolution hybrid network, which is trained based on 2000 sets of historical data of training energy storage systems. It includes a time series convolution layer (processing the time series data of each battery, such as the time dependence of the voltage curve), a graph convolution layer (constructing an association graph between batteries, with nodes as batteries and edge weights as similarity scores) and a preset battery state feature set (containing 1000 feature templates of typical degradation states, such as "normal aging", "overcharge damage", "thermal runaway precursor", etc.).

[0028] Finally, a hybrid time-series-graph convolutional network analyzes the target battery data set and outputs the results. The network first matches each battery to the closest degradation template within a preset feature set (for example, B32 is matched to the "SEI film thickening caused by high-rate discharge" template, and B15 is matched to the same template). It then uses a multi-head graph attention mechanism (eight attention heads) to analyze the interactions between the batteries. For example, because B32 and B15 are in the same battery cluster, the coupled temperature fields cause their internal resistance growth rate to be 10% higher than when they operate independently. Meanwhile, because B47 is located at a vent, temperature fluctuations are minimal, resulting in a degradation rate 5% lower than the average. The network ultimately outputs a system-level degradation characterization (e.g., “the system-level health is 82%, of which B32 is 80%, and the associated battery averages 83%, which is 3% lower than the standard system with the same number of cycles”) and a degradation attribution report (the main cause is located through attention weight visualization: the high-rate discharge of B32 (average 1.5C, standard 0.8C) causes abnormal thickening of the SEI film (internal resistance growth rate 0.1mΩ / week, standard 0.05mΩ / week), while thermal coupling with B15 (the temperature difference between the two is less than 5°C, resulting in heat accumulation) exacerbates the degradation). The server combines system-level degradation characterization and attribution reports to generate core capacity decision recommendations: immediately perform online core capacity testing on B32 (measure the actual capacity by discharging it to 2.7V with a low current, and the actual capacity is expected to be 78% of the initial capacity); adjust the charge and discharge rate limit of the battery cluster where B32 is located to 1.0C and increase the ventilation volume (target temperature difference >10°C); increase the temperature sampling frequency of similar batteries such as B15 (from 1 time per minute to 5 times per minute) to continuously monitor changes in internal resistance.

[0029] Through the above implementation process, the server realizes intelligent online capacity verification of energy storage batteries based on deep learning technology. It can accurately evaluate the battery health status and provide decision support without system shutdown, effectively solving the problem of high cost and impact on system operation of traditional offline capacity verification.

[0030] In the embodiment of the present invention, the extraction of target battery state features of the battery from the operation segments in the historical operation data set of the target battery and the target energy storage system can be performed through the following examples.

[0031] Generate a feature chain of the battery instance to be evaluated based on the multi-dimensional battery fingerprint of the battery instance to be evaluated, and input the feature chain of the battery instance to be evaluated into a target feature compression encoder; the battery instance to be evaluated is an operation segment in a historical operation data set of the target battery or the target energy storage system, and the target feature compression encoder includes a high-dimensional time series feature set;

[0032] Through the target feature compression encoder, in the high-dimensional time series feature set, the high-dimensional time series features corresponding to the multi-dimensional battery fingerprint of the battery instance to be evaluated are matched, cross-domain feature fusion and health factor mapping are performed on the matched high-dimensional time series features, and the target battery state features of the battery instance to be evaluated are output.

[0033] In an exemplary embodiment of the present invention, in an energy storage system of a new energy power station, the server performs target battery state feature extraction for target battery B32 and each operating segment in the target energy storage system's historical operating data as follows: First, the server generates a "battery instance feature chain to be evaluated" for each battery instance to be evaluated (including the operating segments of target battery B32 and the other 199 batteries in the system). Taking target battery B32 as an example, its operating segment from January 1 to 7, 2023 (i.e., one operating segment) contains minute-by-minute data such as voltage, current, temperature, and SOC. The server first extracts a "multi-dimensional battery fingerprint" from the fragment: Time dimension: Analyze the daily charge and discharge curves and calculate the duration of the 3.2V voltage platform (for example, when charging on January 3, the voltage maintained at 3.2V for 45 minutes during the process of SOC rising from 50% to 70%); Calculate the temperature rise rate within 1 hour of charging (for example, when charging on January 5, the temperature rose from 25°C to 35°C at a rate of 10°C / hour); Statistical dimension: Calculate the average internal resistance growth rate for the week (for example, The internal resistance of the B32 increased from 8.1mΩ on January 1 to 8.4mΩ on January 7, with an average weekly growth rate of 0.043mΩ / day); the hysteresis voltage of the SOC-voltage curve was extracted (for example, the voltage was 3.1V at the end of discharge and 3.0V at the beginning of charging the next day, with a hysteresis voltage of 0.1V); and the event dimension: marking whether abnormal operating conditions occurred that week (for example, on January 2, due to the peak-shaving demand of the power grid, the B32 experienced an over-discharge (the voltage dropped to 2.68V), triggering an abnormal flag). The fingerprints of the above 15 dimensions are concatenated in chronological order (for example, the fingerprint from January 1 to 7, 2023 is F1_20230101-07, and from January 8 to 14, 2023 is F1_20230108-14), thus forming the "feature chain of the battery instance to be evaluated" of B32 (such as [F1_20230101-07, F1_20230108-14, ..., F1_20231225-31]). Subsequently, the server inputs the feature chain into the trained "target feature compression encoder". The encoder is constructed based on historical training data (such as the degradation data of 100,000 batteries) and internally stores a "high-dimensional time series feature set" (containing 200 degradation feature templates, such as the "initial capacity attenuation template" corresponding to a shortened voltage platform and a slow increase in internal resistance; the "overcharge damage template" corresponds to an abnormal increase in voltage and a surge in internal resistance, etc.). The encoder first performs "high-dimensional time series feature matching": for each fingerprint in the feature chain (such as B32's "over-discharge mark + internal resistance weekly increase of 0.043mΩ"), it searches for the most similar template in the high-dimensional time series feature set (for example, it matches the "initial template of SEI film damage caused by over-discharge", the typical characteristics of which are: internal resistance weekly increase of 0.04-0.05mΩ after over-discharge, and voltage hysteresis of 0.08-0.12V).After the matching is completed, the encoder performs "cross-domain feature fusion" on the matched high-dimensional features (such as "SEI film damage", "internal resistance growth", "voltage hysteresis" and other 200-dimensional features): through the fully connected layer and ReLU activation function, redundant features are eliminated (for example, the two strongly correlated features of "internal resistance growth rate" and "voltage hysteresis" are merged into "interface impedance damage degree"), and finally the 200-dimensional features are compressed to 50 dimensions to form the fused intermediate features. Finally, the encoder converts the intermediate features into interpretable "target battery state features" through "health factor mapping": the 50-dimensional intermediate features are mapped to the health factor space using a linear mapping layer (for example, a 0-1 SOH prediction value, a 0-0.2 internal resistance growth rate level, and a 0-1 thermal runaway risk level). For example, after processing the January 2023 operating segment of B32, the output target battery state characteristics are [0.94, 0.045, 0.2], which respectively indicate that the corresponding SOH prediction value of this segment is 94%, the internal resistance growth rate is 0.045mΩ / day, and the thermal runaway risk level is 20% (low risk). For other operating segments in the target energy storage system (such as the May 2022 segment of B15), the server adopts the same process: extracting multi-dimensional battery fingerprints - generating feature chains - input encoder matching, fusion, and mapping, and finally obtains the target battery state characteristics of each segment (such as the May 2022 characteristics of B15 are [0.95, 0.04, 0.15]). Through the above process, the server completes the target battery state feature extraction of the target battery and each operating segment, providing key state representation for subsequent degradation trajectory similarity analysis and capacity decision-making.

[0034] In the embodiments of the present invention, the following implementation modes are also provided.

[0035] Obtaining a first battery instance feature chain and inputting the first battery instance feature chain into an initial feature compression encoder; the first battery instance feature chain is obtained based on a historical operating data set of a first training energy storage system, and the initial feature compression encoder includes the high-dimensional time series feature set;

[0036] By means of the initial feature compression encoder, in the high-dimensional time series feature set, high-dimensional time series features corresponding to each multi-dimensional battery fingerprint preceding the first time series node in the first battery instance feature chain are matched, cross-domain feature fusion and health factor mapping are performed on each matched high-dimensional time series feature to obtain a degradation deduction battery state feature corresponding to the first time series node, and based on the degradation deduction battery state feature corresponding to the first time series node, a first health state credible value corresponding to the multi-dimensional battery fingerprint at the first time series node in the first battery instance feature chain is obtained; the first time series node is determined from each multi-dimensional battery fingerprint time series node in the first battery instance feature chain, and the first health state credible value is used to indicate a degradation correlation between the battery corresponding to the first time series node and a time series preceding operation data set corresponding to the first time series node;

[0037] Obtaining a first degradation representation cost based on first health state credible values ​​corresponding to the multi-dimensional battery fingerprint at each first time sequence node in the first battery instance feature chain;

[0038] The encoder parameters of the initial feature compression encoder are iteratively optimized based on the first degradation representation cost to obtain the target feature compression encoder.

[0039] In an embodiment of the present invention, for example, in the practical application of the intelligent online capacity verification method for energy storage batteries, the server needs to optimize the initial feature compression encoder using historical training data to obtain a target feature compression encoder that can accurately extract the target battery state features. The following details the server training process using historical data from a retired energy storage power station (i.e., the first training energy storage system) as training samples: the server first obtains the full life cycle operating data of 100 retired batteries (e.g., numbered C01-C100) from the BMS of the first training energy storage system, including time series records of each battery's voltage, current, temperature, SOC, internal resistance, etc. (a span of 5 years, a total of 60 months of data). Based on this data, the server generates a "first battery instance feature chain" for each battery: Taking battery C17 as an example, its feature chain is composed of a series of monthly "multi-dimensional battery fingerprints" (such as fingerprint F1_201801 for January 2018, fingerprint F1_201802 for February 2018, and fingerprint F1_202212 for December 2022, a total of 60 fingerprint nodes). Each fingerprint contains features in 15 dimensions, including time dimension (such as voltage platform length), statistical dimension (such as the average monthly growth rate of internal resistance), and event dimension (such as overcharge mark). The server then inputs C17's first battery instance feature chain into the "initial feature compression encoder." In its initial state, the encoder already contains a preset "high-dimensional time series feature set" (such as a 200-dimensional degradation feature template pre-extracted from 50,000 battery data, covering typical degradation modes such as "normal aging," "overcharge damage," and "thermal runaway precursors"). The server needs to use this encoder to deduce the battery status of each "first time series node" and calculate the credibility to optimize the parameters. The server first selects the "first time series node" from the C17 feature chain - for example, the end of each quarter (March 2018, June 2018, ..., December 2022, a total of 20 nodes). For each first time series node (such as the March 2018 node), the server extracts all multi-dimensional battery fingerprints before this node (i.e., fingerprints from January 2018 and February 2018) and inputs them into the initial encoder for processing: High-dimensional time series feature matching: The encoder searches the high-dimensional time series feature set for the template that is most similar to the fingerprint from January to February 2018 (for example, it matches the "normal aging initial stage template", whose typical characteristics are: stable voltage platform length, average monthly internal resistance increase of 0.03mΩ, and no abnormal operating condition mark). ; Cross-domain feature fusion: The matched 200-dimensional high-dimensional features (such as capacity retention rate, internal resistance growth rate, temperature sensitivity, etc.) are fused through the fully connected layer and the BatchNorm layer to eliminate redundancy (compressed to 50-dimensional intermediate features); Health factor mapping: The 50-dimensional intermediate features are mapped to the health factor space, and the "degradation deduction battery status characteristics" are output (such as the deduction characteristics of the node in March 2018 are [SOH deduction value 0.97, internal resistance growth rate 0.03mΩ / month, thermal runaway risk 0.1]).The server further calculates the "first health state credibility value" of the node: the inferred SOH value (0.97) is compared with the actual disassembly measurement value of the node (for example, the actual capacity of C17 in March 2018 is 97.5% of the initial capacity, corresponding to an actual SOH of 0.975), and the credibility value is calculated by the mean square error (credible value = 1-|inferred SOH-actual SOH|, here 1-0.005=0.995). The higher the credibility value, the higher the accuracy of the encoder in inferring the current state based on the previous data. Repeating the above process, the server calculates the credibility value of all 20 first time series nodes in the C17 feature chain (such as the credibility value of the node in June 2018 is 0.98, the node in September 2018 is 0.97, etc.), and summarizes these credibility values ​​as the "first degradation characterization cost" (for example, the total cost is the sum of the squares of the errors of each node, and the total error here is 0.005). 2 +0.02 2 +…=0.0012). If the cost is high (such as exceeding the preset threshold of 0.001), the server iteratively optimizes the parameters of the initial encoder through the back-propagation algorithm: adjusting the weights of the fully connected layer of the health factor mapper, the convolution kernel parameters of the degradation trajectory fusion device, and the loss function coefficients of the credibility quantizer until the total cost is reduced to 0.0008 (below the threshold). Finally, the server repeatedly trains the feature chains of 100 batteries in the first training energy storage system (50 rounds of iterations for each battery) to obtain the "target feature compression encoder" with optimized parameters. This encoder can more accurately extract the target battery state features from the feature chain of the battery instance to be evaluated (such as reducing the error between the deduced SOH and the actual value from the initial ±2% to ±0.5%), providing a reliable state characterization basis for subsequent core capacity analysis.

[0040] In an embodiment of the present invention, the battery status characteristics corresponding to the degradation of the first timing node are deduced to obtain the first health status credible value corresponding to the multi-dimensional battery fingerprint at the first timing node in the first battery instance feature chain, which can be implemented through the following example.

[0041] Projecting the degradation trajectory of the degradation deduction battery state feature corresponding to the first timing node to obtain an initial battery state space representation corresponding to the first timing node; the initial battery state space representation includes the health state value corresponding to each battery to be evaluated in the core capacity candidate pool;

[0042] quantifying the health state of the initial battery state space representation to obtain a target battery state space representation corresponding to the first time series node; the target battery state space representation includes a first health state credible value corresponding to each battery to be evaluated in the core capacity candidate pool, and the core capacity candidate pool includes batteries corresponding to each multi-dimensional battery fingerprint in the first battery instance feature chain;

[0043] A first health state credible value corresponding to the multi-dimensional battery fingerprint at the first time sequence node in the first battery instance feature chain is determined from the target battery state space representation.

[0044] In an embodiment of the present invention, exemplarily, during the encoder training process of the first training energy storage system, the server needs to calculate the first health state credible value corresponding to the node based on the degradation deduction battery state characteristics of the first time series node. The following takes the March 2018 node (first time series node) of the retired battery C17 as an example to explain the specific implementation process in detail: First, the server "projects the degradation trajectory" of C17 at the March 2018 node (such as the deduction result is [SOH deduction value 0.97, internal resistance growth rate 0.03mΩ / month, thermal runaway risk 0.1]). The projection process maps the deduction characteristics to a pre-constructed "battery state space" - the space is formed with the SOH value as the horizontal axis (0-1), the internal resistance growth rate as the vertical axis (0-0.2mΩ / month), and the thermal runaway risk as the height (0-1), forming a three-dimensional state space. The historical lifecycle data for each battery in the core capacity candidate pool (i.e., the 100 batteries C01-C100 participating in the training) has been pre-projected into this space (for example, the state point of C05 in June 2019 was (0.92, 0.05, 0.2), and the state point of C23 in March 2020 was (0.85, 0.1, 0.3)). The server uses a linear transformation to map C17's inferred features [0.97, 0.03, 0.1] into the state space, resulting in an "initial battery state space representation": This representation contains the health state values ​​of all 100 batteries in the core capacity candidate pool at the corresponding time points (for example, the inferred points of C17 are (0.97, 0.03, 0.1), the inferred points of C01 during the same period are (0.98, 0.02, 0.05), and the inferred points of C100 during the same period are (0.96, 0.04, 0.15)). Next, the server performs "health state quantification" on the initial state space representation. The core of quantification is to compare the inferred value with the actual measured value to calculate a credible value. For the C17 node in March 2018, the server obtained its actual health state data through battery disassembly experiments: the actual capacity was 97.5% of the initial capacity (corresponding to an actual SOH value of 0.975), the actual internal resistance growth rate was 0.032mΩ / month, and the actual thermal runaway risk was 0.08 (confirmed by electrolyte composition analysis).The server uses the SOH value as the main quantitative indicator (weight 0.7), and the internal resistance growth rate (weight 0.2) and thermal runaway risk (weight 0.1) as auxiliary indicators to calculate the comprehensive error: [{comprehensive error} = 0.7×|0.97-0.975|+0.2×|0.03-0.032|+0.1×|0.1-0.08|=0.7×0.005+0.2×0.002+0.1×0.02=0.0035+0.0004+0.002=0.0059] Then, the server maps the comprehensive error to a credible value range of 0-1 (credible value = 1-comprehensive error), and obtains the "first health state credible value" of C17 at the node in March 2018 as (1-0.0059=0.9941) (close to 1, indicating that the deduction result is highly consistent with the actual result). For other batteries in the core capacity candidate pool (such as C01 and C100), the server repeats the above projection and quantization process: extract the deduced features of each battery at the corresponding first time series node - project it to the state space - compare it with the actual measured value to calculate the comprehensive error - obtain the first health state credible value of each battery (for example, the credible value of C01 is 0.985, and the credible value of C100 is 0.972). Finally, the server extracts the credible value of C17 at the March 2018 node of 0.9941 from the quantized "target battery state space representation" as the first health state credible value of the node. This value directly reflects the accuracy of the encoder's deduction of the current state based on the previous operation data, and provides a key basis for the subsequent optimization of the encoder parameters through the first degradation representation cost (for example, if the value is lower than 0.95, the server will adjust the encoder's health factor mapping layer weight to reduce the deduction error).

[0045] In an embodiment of the present invention, the initial feature compression encoder includes a feature encoder, a health factor mapper, a degradation trajectory fusion device, and a credibility quantizer. The feature encoder is used to match high-dimensional time series features, the health factor mapper is used for health factor mapping, the degradation trajectory fusion device is used for cross-domain feature fusion, and the credibility quantizer is used to output a first health state credibility value;

[0046] The iterative optimization of the encoder parameters of the initial feature compression encoder based on the first degradation representation cost to obtain the target feature compression encoder can be implemented through the following examples.

[0047] The encoder parameters of the health factor mapper, the degradation trajectory fuser and the credibility quantizer in the initial feature compression encoder are iteratively optimized based on the first degradation representation cost to obtain the target feature compression encoder.

[0048] In an embodiment of the present invention, exemplarily, in the encoder training of the first training energy storage system, the server needs to optimize the parameters of the health factor mapper, degradation trajectory fusion device and credibility quantizer of the initial feature compression encoder based on the "first degradation representation cost". The following takes the training process of retired battery C17 as an example to explain in detail the optimization logic of each component: the initial feature compression encoder includes four modules: feature encoder (fixed high-dimensional time series feature matching rules), degradation trajectory fusion device (fully connected layer and convolution layer responsible for cross-domain feature fusion), health factor mapper (linear layer that maps fused features to health factor space), credibility quantizer (calculate the error function and weight parameters of the first health state credibility value). The optimization goal of the server is to reduce the first degradation representation cost (i.e., the sum of the credibility value errors of each first time series node) by adjusting the parameters of the latter three. Taking the C17 node from March 2018 as an example, during initial training, the first health state confidence value for this node was 0.9941 (with an error of 0.0059). However, the confidence values ​​for other nodes (such as the node from June 2018) were only 0.95 (with an error of 0.05), resulting in a total first degradation representation cost of 0.008 (exceeding the preset threshold of 0.005). The server initiated the parameter optimization process: The core of the degradation trajectory fuser is a set of fully connected layers (e.g., 200-dimensional input and 50-dimensional output) and a BatchNorm layer. The server calculated the gradients of this module through backpropagation. For example, among the inferred features for the June 2018 node, the "internal resistance growth rate" had a large error from the actual value (inferred 0.04mΩ / month, actual 0.05mΩ / month), indicating that the internal resistance-related features were not fully extracted during the fusion process. The server adjusted the weight matrix of the fully connected layer (for example, increasing the weight coefficient for the input dimension corresponding to "internal resistance growth rate" from 0.3 to 0.4) and updated the mean and variance parameters of the BatchNorm layer (for example, adjusting the normalized mean of the internal resistance feature from 0.03 to 0.035) to make the fused intermediate features more prominent in the impact of internal resistance growth. The health factor mapper includes a linear transformation layer (50-dimensional input, 3-dimensional output: SOH, internal resistance rate, and thermal risk) and an activation function (such as Sigmoid). The server discovered that the estimated SOH value (0.94) for the node in June 2018 was lower than the actual value (0.95), primarily due to the bias parameters of the linear layer. By calculating the gradient, the server adjusted the bias vector of this layer (for example, changing the bias of the SOH output dimension from -0.01 to 0) and slightly increased the corresponding weight (for example, from 0.85 to 0.87), bringing the estimated SOH value closer to the actual measured value. The trustworthiness quantifier is responsible for calculating the composite error (e.g., SOH weight 0.7, internal resistance 0.2, thermal risk 0.1). Server analysis found that the thermal risk inference errors of some nodes were large (e.g., in September 2018, the inferred thermal risk of a node was 0.15, but the actual thermal risk was 0.1), but due to the low weight (0.1), it was not fully constrained.The server adjusts the weight parameters (such as increasing the thermal risk weight to 0.15 and reducing the internal resistance weight to 0.15), and optimizes the smoothing coefficient of the error function (such as changing the absolute error to the mean square error to amplify the impact of small errors). After the adjustment, the comprehensive error of the node in September 2018 dropped from 0.008 to 0.005, and the credibility value increased from 0.92 to 0.95. The server iteratively optimizes the above parameters with a learning rate of 0.001 using the stochastic gradient descent (SGD) algorithm (processing the feature chain of 10 batteries per round). After 50 rounds of iterations, the total first degradation characterization cost of C17 dropped from 0.008 to 0.003 (below the threshold of 0.005), and the average credibility value of each node increased from 0.96 to 0.985 (deduction error ≤ 0.015). Finally, the server fixed the optimized parameters to obtain the "target feature compression encoder." The optimized degradation trajectory fuser can more accurately extract key degradation features such as internal resistance and voltage (redundant features are reduced by 30%). The health factor mapper's SOH inference error has been reduced from ±2% to ±0.8%, and the correlation coefficient between the credibility quantizer's comprehensive error calculation and actual disassembly data has been improved from 0.89 to 0.95. When subsequently applied to feature extraction of target batteries, this encoder can output more reliable target battery state characteristics, providing a solid characterization foundation for capacity reduction decisions.

[0049] In the embodiments of the present invention, the following implementation modes are also provided.

[0050] Obtain multi-dimensional operating condition information corresponding to each battery to be evaluated in the core capacity candidate pool;

[0051] Performing feature extraction processing on the multi-dimensional operating condition information through a multimodal time series convolutional network to obtain high-dimensional time series features corresponding to each of the batteries to be evaluated;

[0052] The high-dimensional time series feature set is obtained based on the high-dimensional time series features corresponding to each of the batteries to be evaluated.

[0053] In an embodiment of the present invention, for example, in a certain energy storage battery data center, the server needs to construct a "high-dimensional time series feature set" for the batteries in the core capacity candidate pool for feature matching of the subsequent target feature compression encoder. The core capacity candidate pool contains 300 lithium iron phosphate batteries (numbered D01-D300) retired from a certain batch. Each battery has completed full life cycle operation data recording and disassembly verification (actual SOH, internal resistance and other parameters are known). The server completes the construction of the high-dimensional time series feature set through the following steps: First, the server obtains "multi-dimensional operating condition information" from the BMS and environmental monitoring system of each battery. Taking battery D12 as an example, its multi-dimensional information includes: timing conditions: minute-by-minute voltage (2.8V-3.6V), current (-2C to +2C, negative for discharge), and temperature (15℃-45℃) curves (data from the past five years, a total of 2,628,000 time points); statistical conditions: monthly charge and discharge cycles (average 30 times / month), deep discharge frequency (5 times per month with SOC < 20%), and monthly average growth rate of internal resistance (0.04mΩ / month); event conditions: timestamps and severity of abnormal events such as overcharge (voltage > 3.65V, 8 times in total), over-discharge (voltage < 2.7V, 12 times in total), and high-temperature storage (temperature > 40℃ for 24 hours, 5 times in total). The server preprocesses the multidimensional information of all 300 batteries by unifying timestamp accuracy (1 minute), normalizing the value range (e.g., voltage normalized to [0, 1]), and filling in missing values ​​(using linear interpolation to fill in occasional BMS communication interruptions) to obtain structured input data. The server then inputs this preprocessed multidimensional information into a "multimodal temporal convolutional network." The network is designed as a three-branch structure, which processes timing conditions, statistical conditions, and event conditions respectively: Timing branch: uses a 1D convolution layer (convolution kernel size = 60, step size = 10, extracts 1-hour voltage / current / temperature change features), connected to a maximum pooling layer (pooling size = 3, reduces time resolution), and outputs 128-dimensional timing features; Statistical branch: uses a fully connected layer (inputs 10 statistical indicators, such as number of cycles and internal resistance growth rate), extracts statistical correlation features through the ReLU activation function, and outputs 64-dimensional statistical features; Event branch: uses an embedding layer (encodes the event type into a 5-dimensional vector and the severity into a 3-dimensional vector), connected to an LSTM layer (captures the time interval and cumulative effect of the event), and outputs 64-dimensional event features. The three-branch features are merged into a 256-dimensional mixed feature through the splicing layer, and then compressed to 200 dimensions through the fully connected layer to form a "high-dimensional time series feature" (for example, the high-dimensional feature of D12 is [0.78, 0.12, 0.05, ..., 0.03], corresponding to the voltage fluctuation pattern, internal resistance growth rate, cumulative impact of over-discharge events, and other dimensions).The server repeats the above process for the 300 cells in the core capacity candidate pool, obtaining 200 high-dimensional time series features for each cell (e.g., the features for D01 are [0.82, 0.09, 0.04, ..., 0.02], and the features for D300 are [0.65, 0.15, 0.08, ..., 0.05]). Finally, the server constructs a "high-dimensional time series feature set" based on these high-dimensional features: The features of the 300 cells are clustered using the K-means clustering algorithm (setting 100 clusters), with the centroid of each cluster serving as a "typical degradation feature template" (e.g., Cluster 1 corresponds to the "normal aging" template: stable voltage fluctuations, slow internal resistance growth, and no abnormal events; Cluster 5 corresponds to the "over-discharge damage" template: shortened voltage plateau, high internal resistance growth rate, and frequent over-discharge events). The server stores these 100 templates as a high-dimensional time series feature set for subsequent feature matching in the target feature compression encoder. Through the above process, the server completes the construction of the high-dimensional time series feature set. This feature set covers the typical degradation patterns of various types of batteries in the core capacity candidate pool, providing a rich template library for extracting target battery status features, ensuring the accuracy of subsequent similarity analysis and core capacity decision-making.

[0054] In an embodiment of the present invention, determining the associated battery data set of the target battery from the historical operation data set of the target energy storage system based on the degradation trajectory similarity metric between the target battery and the target battery state characteristics of each of the operation segments can be implemented through the following example.

[0055] Based on a degradation trajectory similarity metric between the target battery and the target battery state characteristics of the operation segments, determining a plurality of associated battery operation segments of the target battery from a historical operation data set of the target energy storage system according to the degradation trajectory similarity metric from low to high;

[0056] The operation segments of the associated batteries are aligned with the charge and discharge cycle phases of the target energy storage system to obtain an associated battery data set of the target battery.

[0057] In an embodiment of the present invention, for example, in the energy storage system of a new energy power station, the server needs to screen the associated battery data set with similar degradation trajectory from the historical operation data of the target energy storage system for the target battery B32 (current cycle number 1200 times). The following is the specific implementation process: First, the server has extracted the "target battery status features" of each operation segment of the target battery B32 and the other 199 batteries in the system through the target feature compression encoder. Taking B32 as an example, its feature trajectory in the last 12 months (cycle number 1100-1300 times) is a time series sequence (S B32 =[f1,f2,...,f 12]), where each (f i ) includes SOH prediction value, internal resistance growth rate, thermal runaway risk level (such as (f1 = [0.95, 0.04, 0.1]), (f2 = [0.94, 0.045, 0.12]), and gradually degrades to (f 12 =[0.85,0.1,0.25])). The historical operation segments of other batteries in the system (such as the January 2020-December 2021 segment of B15 and the June 2019-May 2020 segment of B47) also store their own characteristic trajectories (such as (S B15 =[0.96,0.94,0.92,...,0.86]), (S B47 =[0.98,0.95,0.90,...,0.80])). The server needs to calculate (S B32 ) and each segment trajectory (S other )’s “degradation trajectory similarity”.

[0058] The server uses the Dynamic Time Warping (DTW) algorithm to measure the similarity of two time series feature curves. DTW calculates the minimum cumulative distance between two sequences by finding the optimal time alignment path (the smaller the distance, the more similar the trajectories). For example: the trajectories of B32 and B15 (S B32 ) and (S B15 ) has a DTW distance of 0.03 (highly similar, with almost identical trajectory shapes); the trajectories of B32 and B47 (S B32 ) and (S B47 ) has a DTW distance of 0.1 (moderately similar, B47 degrades faster in the early stage); the trajectories of B32 and B99 (S B32 ) and (S B99 ) has a DTW distance of 0.25 (not similar, B99 has a slow increase in internal resistance due to long-term low-temperature operation). The server sorts the DTW distance from small to large (i.e., similarity from high to low) and selects the top 20 most similar operating segments (corresponding to 15 different batteries). The DTW distances of these segments are all less than 0.08, ensuring a high correlation with the degradation pattern of B32. Next, the server needs to align these associated segments according to the "charge and discharge cycle phase" to ensure that each segment is in the same degradation stage as B32. The specific operations are as follows: The current number of cycles of the target battery B32 is 1200, and its degradation trajectory (S B32) corresponds to a cycle number of 1100-1300 times (that is, every 100 cycles correspond to a feature point of 1 month); for each associated segment (such as the January 2020-December 2021 segment of B15), the server extracts its cycle number label (the B15 segment covers a cycle number of 1000-1200 times), and maps its feature trajectory to the time axis of 1100-1300 cycles through linear interpolation (for example, the B15 cycle number of 1000 times corresponds to the B32 cycle number of 1100 times, and the cycle number of 1200 times corresponds to the B32 cycle number of 1300 times, and the middle points are aligned proportionally); for segments with missing cycle numbers (such as the B47 segment only records cycle numbers of 900-1100 times), the server uses historical data to complete the features of the subsequent 100 cycles (based on the extrapolation of its internal resistance growth rate) to ensure alignment to the 1100-1300 times interval. Ultimately, the server generated an "associated battery data set," consisting of characteristic trajectory data from 15 batteries aligned over 1100-1300 cycles. For example, the aligned trajectory for B15 is ([0.95, 0.93, 0.91, ..., 0.86]), which nearly overlaps with the trajectory for B32 ([0.95, 0.93, 0.91, ..., 0.85]). This data set is at the same degradation stage as B32, and exhibits highly similar degradation patterns. This provides reliable comparative data for subsequent timing-graph convolution hybrid network analysis, ensuring the accuracy of system-level degradation characterization and capacity decisions.

[0059] In the embodiments of the present invention, the following implementation modes are also provided.

[0060] Obtaining a second battery instance feature chain, and inputting the second battery instance feature chain into an original hybrid network; wherein the second battery instance feature chain is obtained based on a set of historical operating data of a second training energy storage system, and the original hybrid network includes the preset battery state feature set;

[0061] By means of the original hybrid network, in the preset battery state feature set, the preset battery state features corresponding to each multidimensional battery fingerprint before the second time series node in the second battery instance feature chain are matched, and a multi-source feature collaborative analysis based on a multi-head graph attention mechanism is performed on each matched preset battery state feature to obtain a degradation deduction battery state feature corresponding to the second time series node, and based on the degradation deduction battery state feature corresponding to the second time series node, a second health state credible value corresponding to the multidimensional battery fingerprint at the second time series node in the second battery instance feature chain is obtained; the second time series node is determined from each multidimensional battery fingerprint time series node in the second battery instance feature chain, and the second health state credible value is used to indicate the degradation correlation between the battery corresponding to the second time series node and the time series preceding operation data set corresponding to the second time series node;

[0062] Obtaining a second degradation representation cost based on the second health state credible value corresponding to each of the multi-dimensional battery fingerprints at each second time series node in the second battery instance feature chain;

[0063] The encoder parameters of the original hybrid network are iteratively optimized based on the second degradation representation cost to obtain the temporal-graph convolutional hybrid network.

[0064] In an embodiment of the present invention, for example, in a photovoltaic energy storage power station (i.e., the second training energy storage system), the server needs to train the "original hybrid network" through historical operating data and optimize its parameters to obtain a "time series-graph convolution hybrid network" that can accurately output system-level degradation characterization. The power station contains 200 lithium iron phosphate batteries (numbered E01-E200), each of which has 5 years of full life cycle operating data (voltage, current, temperature, SOC, etc.) and actual health status data after decommissioning and disassembly (actual SOH, internal resistance, electrolyte composition, etc.). The following takes the training process of battery E12 as an example to explain in detail the specific operation of the server: the server first extracts the historical operating data of each battery from the BMS of the second training energy storage system, and generates a "second battery instance feature chain" for each battery. Taking the E12 as an example, its feature chain is composed of a series of monthly "multi-dimensional battery fingerprints" (e.g., fingerprint F2_201801 for January 2018, fingerprint F2_201802 for February 2018, and fingerprint F2_202212 for December 2022, for a total of 60 fingerprint nodes). Each fingerprint contains features in 15 dimensions, including time (e.g., voltage plateau length), statistical dimensions (e.g., monthly average internal resistance growth rate), and event dimensions (e.g., overcharge flags). For example, F2_201801 contains information such as "voltage plateau for 45 minutes, internal resistance monthly increase of 0.03 mΩ, and no abnormal events." The server inputs the feature chain of the second battery instance of E12 into the "original hybrid network." In its initial state, the network already includes a "preset battery state feature set" (expanded from the high-dimensional time series feature set of the first training energy storage system, containing 1,000 typical degradation templates, such as "normal aging," "overcharge damage," and "thermal runaway precursors"). It also possesses infrastructure such as time series convolutional layers, graph convolutional layers, and multi-head graph attention layers. The server processes each "second time series node" in the feature chain of the second battery instance (for example, selecting 10 nodes for each semi-annual time point: June 2018, December 2018, ..., December 2022) using the original hybrid network. For the June 2018 node (second time series node) as an example, the fingerprint preceding it is the fingerprint from January to May 2018 (F2_201801 to F2_201805). Preset feature matching: The original hybrid network first searches the preset battery state feature set for the degradation template that is most similar to the fingerprint from January to May 2018. For example, the five fingerprints of E12 all show "stable voltage platform (45-48 minutes), monthly internal resistance increase of 0.03-0.035mΩ, and no abnormal events", which matches the "normal early aging template" (the typical characteristics of this template are: voltage platform 40-50 minutes, monthly internal resistance increase ≤0.04mΩ, and no abnormal events).Multi-head graph attention collaborative analysis: The network inputs matched preset features (such as the 200-dimensional features of the "normal aging" template) into the graph convolution layer to construct a correlation graph between batteries (nodes are E01-E200, and edge weights are the cosine similarity of each battery's features). Subsequently, a multi-head graph attention mechanism with eight attention heads is used to analyze the collaborative degradation relationships between batteries. For example, due to the physical proximity of E12 and its clustered counterparts E11 and E13, the coupled temperature fields result in a 5% higher internal resistance growth rate than a single battery. Furthermore, due to the different charge and discharge strategies of E12 and E50 (E50 participates in more deep discharge), the voltage plateau length of E12 decreases 10% faster than that of E12. Outputting degradation-inferred battery state characteristics: By fusing a temporal convolutional layer (capturing temporal dependencies) with a graph attention layer (capturing spatial associations), the network outputs the "degradation-inferred battery state characteristics" for the node in June 2018 (e.g., [SOH inferred value 0.96, internal resistance growth rate 0.035mΩ / month, thermal runaway risk 0.08]). The server needs to verify the consistency of the inferred characteristics with the actual data and calculate the "Second Health State Credibility Value." The actual disassembly data for the E12 in June 2018 showed: the actual capacity was 96.5% of the initial capacity (corresponding to actual SOH = 0.965), the actual internal resistance growth rate was 0.037mΩ / month, and the actual thermal runaway risk was 0.07 (confirmed by electrolyte composition analysis). For servers, SOH is used as the primary metric (weight 0.7), with internal resistance growth rate (0.2) and thermal risk (0.1) as auxiliary indicators. The comprehensive error is calculated as follows: [{comprehensive error} = 0.7 × |0.96-0.965| + 0.2 × |0.035-0.037| + 0.1 × |0.08-0.07| = 0.7 × 0.005 + 0.2 × 0.002 + 0.1 × 0.01 = 0.0035 + 0.0004 + 0.001 = 0.0049]. The credible value = 1 - comprehensive error = 0.9951, indicating that the deduced results are highly consistent with actual results. The server repeats the above calculation for all 10 second time series nodes in the E12 feature chain (e.g., the node credibility value in December 2018 is 0.98, the node credibility value in June 2019 is 0.97, etc.), and summarizes them as the "second degradation representation cost" (for example, the total cost is the sum of the squares of the errors of each node, where the total error is 0.0049). 2 +0.02 2+…=0.0015). If the cost exceeds the preset threshold of 0.001, the server starts parameter optimization. The original hybrid network contains a feature encoder (fixed preset feature matching rules), a feature coupler (multi-head graph attention layer and graph convolution layer), and a credibility quantizer (error calculation layer and weight parameters). The server optimizes the parameters of the latter two through backpropagation: Feature coupler optimization: It was found that the error in the deduction of the internal resistance growth rate of the node in June 2019 was large (deduced 0.04mΩ / month, actual 0.05mΩ / month). The server adjusted the attention weight of the multi-head attention layer (such as increasing the attention coefficient of the "same cluster battery temperature" feature from 0.2 to 0.3) to make the network pay more attention to the impact of temperature coupling on internal resistance; at the same time, the adjacency matrix parameters of the graph convolution layer were updated (such as increasing the edge weight of the same cluster battery from 0.8 to 0.9) to enhance the ability to capture spatial associations. Credibility quantifier optimization: The thermal risk deduction error of some nodes was not fully constrained (for example, the node deduced a thermal risk of 0.12 in December 2020, but the actual risk was 0.1). The server adjusted the weight of the error function (increasing the thermal risk weight from 0.1 to 0.15) and optimized the loss function to smooth L1 loss (reducing the impact of outliers). After the adjustment, the comprehensive error of the node dropped from 0.008 to 0.005, and the credibility value increased from 0.92 to 0.95. The server iteratively trained 200 rounds using the Adam optimizer (learning rate 0.0001), processing the feature chain of 50 batteries in each round. In the end, the total second degradation characterization cost of E12 dropped from 0.0015 to 0.0007 (below the threshold), and the average credibility value of each node increased from 0.96 to 0.985 (deduction error ≤ 0.015). In the optimized "time series-graph convolutional hybrid network," the feature coupler more accurately captures the coordinated degradation relationship between batteries (the error in inferring the internal resistance growth rate of batteries in the same cluster has been reduced from ±8% to ±3%), and the correlation coefficient between the comprehensive error of the credibility quantizer and actual teardown data has increased from 0.85 to 0.93. When this network is subsequently applied to capacity analysis of target batteries, it can output more reliable system-level degradation characterization and attribution reports, providing key support for capacity reduction decisions.

[0065] In an embodiment of the present invention, the original hybrid network includes a feature encoder, a feature coupler, and a credibility quantizer. The feature encoder is used to match preset battery status features. The feature coupler is used for collaborative analysis of multi-source features based on a multi-head graph attention mechanism. The credibility quantizer is used to output a second health status credibility value.

[0066] The iterative optimization of the encoder parameters of the original hybrid network based on the second degradation representation cost to obtain the timing-graph convolution hybrid network can be implemented through the following examples.

[0067] The encoder parameters of the feature coupler and the credibility quantizer in the original hybrid network are iteratively optimized based on the second degradation representation cost to obtain the timing-graph convolution hybrid network.

[0068] In an embodiment of the present invention, exemplarily, during network training of the second training energy storage system, the server optimizes the parameters of the feature coupler and credibility quantizer of the original hybrid network to enhance its collaborative analysis capabilities for battery degradation. The following uses the training process of retired battery E12 as an example to explain the optimization logic in detail: The original hybrid network consists of three modules: a feature encoder (fixed with preset battery state feature matching rules), a feature coupler (including a multi-head graph attention layer and a graph convolution layer, responsible for collaborative analysis of multi-source features), and a credibility quantizer (calculating the error function and weight parameters of the second health state credibility value). The server's optimization goal is to reduce the "second degradation representation cost" (i.e., the sum of the credibility value errors of each second time series node) by adjusting the parameters of the latter two. The core of the feature coupler is the multi-head graph attention layer (8 attention heads) and the graph convolution layer (the adjacency matrix (A) represents the correlation weights between batteries). Taking the June 2019 node of E12 as an example, during initial training, the "internal resistance growth rate" in the deduced feature of this node had a large error (deduced 0.04mΩ / month, actual 0.05mΩ / month). The server calculated the gradients through backpropagation and discovered that the error stemmed from the attention head's insufficient attention to the "same-cluster temperature coupling" feature. Multi-head attention head weight adjustment: The server adjusted the query ((W^Q)), key ((W^K)), and value ((W^V)) matrix parameters of each attention head. For example, the weight of the "temperature standard deviation" dimension in the (W^Q) matrix of the third attention head responsible for "temperature field correlation" was increased from 0.2 to 0.3, making this head pay more attention to the impact of temperature fluctuations on the internal resistance of batteries in the same cluster. After this adjustment, the attention weight output by this head increased the correlation between E12 and its cluster members E11 and E13 from 0.7 to 0.85, and the inferred internal resistance growth rate was revised from 0.04 to 0.048 (closer to the actual 0.05). Graph convolution layer adjacency matrix optimization: The server updated the graph convolution layer's adjacency matrix (A), increasing the edge weights of batteries in the same cluster from 0.8 to 0.9 (the weights of batteries outside the cluster remained at 0.5). This move strengthens the degradation correlation between physically adjacent batteries, making the inferred characteristics of E12 more consistent with the actual degradation trend of the same cluster of batteries (for example, the internal resistance growth rate of E11 is 0.049mΩ / month, that of E13 is 0.051mΩ / month, and the inferred value of E12 is adjusted to 0.049, with the error reduced to 0.001). The credibility quantifier includes an error function (such as mean square error) and weight parameters for each feature (SOH weight 0.7, internal resistance 0.2, thermal risk 0.1). The server found that in the December 2020 node of E12, the error between the thermal risk inferred value (0.12) and the actual value (0.1) was not fully constrained (because the weight was only 0.1). Weight parameter adjustment: The server increased the thermal risk weight from 0.1 to 0.15 and reduced the internal resistance weight to 0.15 (the SOH weight remained at 0.7).After the adjustment, the node's comprehensive error calculation is: [0.7×|0.92-0.93|+0.15×|0.05-0.052|+0.15×|0.12-0.1|=0.007+0.0003+0.003=0.0103]. The credibility value increased from 0.92 (error 0.08) to 0.9897 (error 0.0103), more accurately reflecting the inference deviation of thermal risk. Loss function optimization: The server switched the error function from mean square error (MSE) to smooth L1 loss (SmoothL1Loss) to reduce the impact of outliers on training. For example, when abnormal BMS data caused the actual internal resistance growth rate of a node to be 0.1mΩ / month (much higher than the normal 0.05mΩ / month), the smoothed L1 loss only calculated the linear error (0.05) rather than the squared error (0.0025), avoiding over-penalization of occasional anomalous data and improving network generalization. After 100 epochs of iterative training using the Adam optimizer (learning rate 0.0001), the total second-order degradation representation cost of the E12 was reduced from 0.0015 to 0.0006 (below the threshold of 0.001), and the average credibility of each node increased from 0.96 to 0.985. The optimized feature coupler more accurately captures the co-degradation relationship between batteries (the error in the inferred internal resistance growth rate of batteries in the same cluster was reduced from ±8% to ±3%), and the correlation coefficient between the comprehensive error of the credibility quantizer and the actual teardown data was improved from 0.85 to 0.93. Finally, the server fixes the optimized parameters to obtain a "time series-graph convolution hybrid network." When this network is subsequently applied to core capacity analysis of the target battery, it can output more reliable system-level degradation characterization and attribution reports.

[0069] In an embodiment of the present invention, the multi-source feature collaborative analysis based on the multi-head graph attention mechanism is performed on each matched preset battery status feature to output the system-level degradation representation of the target battery data set, which can be implemented through the following examples.

[0070] combining the matched preset battery state features according to the battery sequence of the target battery data set to obtain a preset battery state feature set;

[0071] Performing a multi-source feature collaborative analysis based on a multi-head graph attention mechanism on the preset battery state feature set to obtain a battery coupling feature set; the battery coupling feature set includes battery coupling features consistent with the number of batteries in the target battery data set;

[0072] A battery coupling feature at a battery timing node of the target battery is obtained from the battery coupling feature set as a system-level degradation representation of the target battery data set.

[0073] In an embodiment of the present invention, for example, in the core capacity analysis of the energy storage system of a new energy power station, the server performs multi-source feature collaborative analysis on the target battery B32 and its associated 15 similar batteries (a total of 16 batteries, constituting the target battery data set) through a time series-graph convolution hybrid network, and the specific process of outputting the system-level degradation characterization is as follows: the server first combines the preset battery state features matched to each battery into a "preset battery state feature set" according to the battery sequence of the target battery data set (sorted by number: B32, B15, B47, B08, ..., B92). For example: B32 matches the "SEI film damage caused by high-rate discharge" template (feature vector (F B32 =[0.85,0.1,0.25]), corresponding to SOH 85%, internal resistance growth rate 0.1mΩ / week, and thermal runaway risk 25%); B15 (same cluster as B32) matches the same template ((F B15 =[0.86,0.095,0.23]); B47 (located at the vent) matches the "normal aging" template ((F B47 =[0.88,0.07,0.15]); the remaining 13 batteries are matched to the templates of “thermal coupling accelerated aging” and “deep discharge damage” respectively. Finally, the preset battery status feature set is ({F B32 ,F B15 ,F B47 ,...,F B92}), containing 16 feature vectors. The server inputs the preset battery status feature set into the multi-head graph attention layer for multi-source feature collaborative analysis. This layer contains 8 attention heads, each of which independently learns different association patterns between batteries (such as temperature coupling, charge and discharge strategy association, physical location proximity, etc.). Take the first attention head (focusing on temperature field association) as an example:

[0074] Calculate the attention weight between batteries: B32 and B15 in the same cluster have a temperature standard deviation of only 2°C (the standard deviation of other batteries is > 5°C), so the attention weight (α B32,B15 =0.8)(much higher than that of B47 (α B32,B47 =0.3)); for the characteristics of B32 (F B32 ), combined with the characteristics of adjacent cells (such as B15 (F B15 )) and perform weighted summation to obtain the coupling characteristics of the head (C (1) B32 = 0.8 × F B15 +0.2×F B32 =[0.858,0.097,0.236]). The second attention head (focusing on the charge-discharge strategy association) found that the charge-discharge rate of B32 and B08 (both belonging to the high-rate discharge group) is 1.5C (other batteries ≤ 1.0C), so (αB32,B08 =0.7), coupling characteristics (C (2) B32 =0.7×F B08 +0.3×F B32 =[0.853,0.099,0.241]). After concatenation and linear transformation of the coupling features output by the 8 attention heads, the final “battery coupling feature” of each battery is obtained (e.g., the coupling feature of B32 (C B32 =[0.855,0.098,0.238]), which integrates multi-dimensional correlations such as temperature and strategy. Finally, the server obtains a "battery coupling feature set" containing 16 coupling features. The server extracts the coupling feature (C) of the target battery B32 at the current time sequence node (cycle number 1200) from the battery coupling feature set. B32 ), as a "system-level degradation characterization". This characterization not only includes the degradation status of B32 itself (SOH85%, internal resistance growth of 0.1mΩ / week), but also integrates the collaborative degradation information with associated batteries (such as the thermal coupling of cluster B15 increases the internal resistance growth rate by 5%, and the strategy impact of the high-rate group B08 increases the thermal risk by 3%). For example, the system-level degradation characterization shows: "The system-level health of B32 is 82% (its own SOH85%, but it is reduced by 3% due to the thermal coupling of batteries in the same cluster), the internal resistance growth rate is 0.098mΩ / week (8% higher than a single battery), and the thermal runaway risk is 23.8% (increased due to the synergistic effect of the high-rate strategy group)". This characterization directly reflects the adaptability of B32 to the health status of the entire energy storage system (such as the average system health of 83%, B32 is slightly lower but not seriously deviated), providing a key basis for core capacity decision-making from a system-level perspective. Through the above process, the server conducts multi-source collaborative analysis based on the multi-head graph attention mechanism, deeply integrating the degradation characteristics of a single battery with the related information of other batteries in the system, outputting a more comprehensive and accurate system-level degradation representation, effectively supporting the scientific and targeted nature of core capacity decision-making.

[0075] In an embodiment of the present invention, the preset battery state feature in the preset battery state feature set is a target battery state feature of the battery, and the target battery state feature of the battery is obtained by encoding the high-dimensional time series feature of the battery through an autoencoder.

[0076] In an embodiment of the present invention, for example, in the intelligent core capacity system of a large energy storage power station, the "preset battery status feature set" constructed by the server is directly composed of the "target battery status features" of the battery, and these target features are obtained by encoding high-dimensional time series features through the autoencoder. The following takes the historical training data of the power station as an example to explain the specific implementation process in detail: the server first collects multi-dimensional operating condition information from the retired energy storage system (training sample), including the full life cycle data of 200 batteries (numbered F01-F200) (voltage, current, temperature, SOC and other time series data, as well as overcharge / over-discharge and other event records). After processing this information through a multimodal time series convolutional network, the "high-dimensional time series features" of each battery are obtained. Taking the F12 battery as an example, its high-dimensional time series features are fused together by three parts: time series features: the 1-hour voltage / current fluctuation pattern extracted by the 1D convolutional layer (e.g., "the voltage plateau lasts for 45 minutes during charging"); statistical features: the monthly average growth rate of internal resistance (0.04 mΩ / month) and the number of cycles (1200 times) extracted by the fully connected layer; and event features: the cumulative impact of over-discharge events (a total of 10 over-discharges, with an average interval of 3 months) extracted by the LSTM layer. Ultimately, the high-dimensional time series features of F12 are a 200-dimensional vector (H F12 =[0.78,0.12,0.05,...,0.03])(each dimension corresponds to a quantitative indicator of different degradation modes). The server inputs the high-dimensional time series features (H) of each battery into the trained "target feature compression encoder" (autoencoder) for encoding. The encoder includes an encoder (dimensionality reduction) and a decoder (reconstruction verification). The core is to extract low-dimensional features that can characterize the nature of battery degradation through encoding. F12 ) as an example, the encoder process is as follows: Encoding layer: Through the fully connected layer and ReLU activation function, the 200-dimensional (H F12 ) compressed to a 50-dimensional intermediate feature (Z F12 =[0.92,0.04,0.15,...,0.08])(corresponding to health factors such as SOH prediction value, internal resistance growth rate, thermal risk level, etc.); Health factor mapping: (Z F12 ) is mapped to an interpretable 3D “target battery state feature” (S F12 =[0.95,0.04,0.1])(representing SOH 95%, internal resistance monthly increase 0.04mΩ, thermal runaway risk 10% respectively). The server will train the target battery status features of all 200 batteries in the sample (such as (S F12 )、(S F05 =[0.93,0.05,0.12]),(S F30=[0.88,0.08,0.2]) etc.) are collected and sorted to form a "preset battery status feature set". Each feature in this set corresponds to a typical battery degradation state, for example: (S F12 ) corresponds to the “normal early stage of aging” (high SOH, slow growth of internal resistance, and no high-risk events); (S F30 ) corresponds to the "mid-stage of deep discharge damage" (SOH drops significantly, internal resistance increases rapidly, and thermal risk increases). To ensure the representativeness of the feature set, the server clusters the 200 target features using the K-means clustering algorithm (setting 100 clusters), and the centroid of each cluster is used as the "typical preset battery state feature". For example, the centroid of cluster 1 (S cluster1 =[0.94,0.042,0.11]) represents the average state of “normal aging”, and the centroid of cluster 5 (S cluster5}=[0.85,0.09,0.25]) represents the average state of “high rate discharge damage”. Finally, the preset battery state feature set contains 100 typical target battery state features, each of which is generated by encoding high-dimensional time series features by the autoencoder. Later in the core capacity analysis, when the feature chain of the target battery is input into the hybrid network, the network will match the most similar preset features in the feature set (such as the feature of the target battery B32 (S B32}=[0.85,0.1,0.25]) matches to cluster 5 (S cluster5 )), thereby quickly identifying degradation patterns and generating system-level characterizations. Through this process, a pre-set battery state feature set directly links high-dimensional time series features with interpretable health factors, providing a standardized degradation template for collaborative analysis of hybrid networks and ensuring the accuracy and interpretability of capacity decisions.

[0077] In the embodiments of the present invention, the following implementation modes are also provided.

[0078] Acquiring battery status characteristics of the target battery and energy storage system characteristics of the target energy storage system;

[0079] The system-level degradation characterization, the battery state characteristics of the target battery, and the energy storage system characteristics of the target energy storage system are input into a target health state adaptation network to obtain a health state adaptation degree between the target battery and the target energy storage system.

[0080] In an embodiment of the present invention, for example, in the core capacity analysis of the energy storage system of a new energy power station, after the server completes the system-level degradation characterization output, it is necessary to further calculate the compatibility of the target battery with the overall health status of the system. The following takes the target battery B32 and the energy storage system (200 batteries) as an example to describe the specific implementation process in detail: the server first extracts key data from the previous processing results: the battery status features of the target battery B32: the current state features (S B32}=[0.85,0.1,0.25])(SOH 85%, internal resistance growth rate 0.1mΩ / week, thermal runaway risk 25%); Energy storage system characteristics of the target energy storage system: The overall characteristics extracted by the server from the system-level operating data, including: system average SOH: 83% (the average SOH value of 200 batteries); system maximum allowable discharge rate: 1.2C (current policy limit); temperature field standard deviation: 5°C (temperature difference between each battery cluster); inter-cluster capacity consistency: 90% (the deviation rate between each cluster capacity and the system average capacity). The server will "system-level degradation characterization" ((C B32 =[0.82,0.098,0.238]), including collaborative degradation information), "Battery status characteristics of B32 (S B 32 )”, “System Characteristics (S sys=[0.83,1.2,5,90])" is input into the "target health state adaptation network". This network is a fully connected neural network, consisting of an input layer (3+3+4=10 dimensions), a hidden layer (two 64-dimensional fully connected layers, ReLU activation), and an output layer (1 dimension, Sigmoid activation, outputting a degree of fitness of 0-1). The network fuses multi-source information and calculates the degree of fitness through the following logic: Feature normalization: The input layer first normalizes each feature (such as SOH normalized to [0,1], discharge rate normalized to [0,2], and temperature standard deviation normalized to [0,10]) to ensure balanced weights of each dimension. For example, the SOH0.85 of B32 is normalized to 0.85, and the system average SOH0.83 is normalized to is 0.83, and the temperature standard deviation of 5°C is normalized to 0.5 (5 / 10). Hidden layer feature fusion: The first hidden layer extracts the "individual-system" correlation features through full connection operations. For example, the difference between the SOH of B32 and the average SOH of the system is calculated (0.85-0.83=0.02) to analyze whether it deviates from the system level; combined with the maximum discharge rate of the system (1.2C) and the internal resistance growth rate of B32 (0.1mΩ / week), the impact of high-rate discharge on its degradation is evaluated (such as the internal resistance growth may accelerate by 5% under 1.2C). Output layer fitness calculation: The second hidden layer further integrates collaborative information (such as "thermal coupling increases the thermal risk of B32 by 3%" in the system-level degradation characterization) through the Sigmoid activation function. Output adaptability. For example, network calculations found that: the SOH of B32 is slightly higher than the system average (+2%), but the internal resistance growth rate is 25% higher than the system average (0.08mΩ / week); the system temperature standard deviation of 5°C causes thermal coupling between B32 and batteries in the same cluster, accelerating the internal resistance growth by 8%; if the current 1.2C discharge strategy is maintained, the thermal runaway risk of B32 (23.8%) may exceed the system safety threshold (25%) within 3 months. Taking all the above factors into consideration, the network output health status adaptability is 0.72 (0.72≤0.8 is the "need to pay attention" level). This value indicates that the health status adaptability of B32 to the system is average. Although it does not seriously slow down the system, there are potential risks caused by excessive internal resistance growth and thermal coupling. The server is based on the adaptability The adaptability (0.72) and system thresholds (such as 0.8 for "good" and 0.7 for "warning") generate targeted suggestions: If the adaptability is greater than 0.8: no adjustment is required, normal operation; 0.7 < adaptability ≤ 0.8: the internal resistance and temperature of B32 need to be closely monitored, and its discharge rate should be adjusted to 1.0C to reduce thermal risks; adaptability ≤ 0.7: capacity must be checked immediately and replacement should be considered. In this case, the adaptability of 0.72 prompted the operation and maintenance personnel to adopt the strategy of "reduced rate operation + enhanced monitoring" for B32 to avoid its accelerated degradation affecting the overall life of the system. Through the above process, the server quantified the health adaptability of the target battery and the system based on multi-source feature fusion and neural network analysis, providing a more accurate system-level basis for capacity decision-making.

[0081] In the embodiments of the present invention, the following implementation modes are also provided.

[0082] Acquire training features configured with sample target values; the training features include battery state features of a training battery, energy storage system features of a training energy storage system, and integrated features obtained based on a set of historical operating data of the training battery and the training energy storage system;

[0083] Based on the training features and the sample target values, reinforcement learning of degradation stage perception is performed on the initial health state adaptation network to obtain the target health state adaptation network.

[0084] In an embodiment of the present invention, for example, in an algorithm optimization project of a certain energy storage technology laboratory, the server needs to perform "degradation stage perception reinforcement learning" on the "initial health state adaptation network" through historical training data to obtain a "target health state adaptation network" that can accurately evaluate the health compatibility of the battery and the system. The following takes the retired energy storage system data collected by the laboratory (including 100 training batteries G01-G100 and corresponding system data) as an example to explain the training process in detail: the server first extracts "training features" from the full life cycle operation data of the retired energy storage system and marks the "sample target value": battery state features of the training battery: target battery state features of each battery (such as G12) (obtained by encoding high-dimensional time series features by autoencoder), for example (S G12=[0.92,0.04,0.1])(SOH 92%, internal resistance monthly increase 0.04 mΩ, thermal risk 10%); training energy storage system characteristics: overall system characteristics (such as average SOH 90%, maximum discharge rate 1.0C, temperature standard deviation 3°C, capacity consistency 95%); integrated characteristics: collaborative degradation information based on historical operating data (such as internal resistance growth rate deviation of cluster batteries, thermal risk correlation of high-rate discharge group, etc.), for example, the "temperature coupling coefficient" of the cluster where G12 is located is 0.8 (indicating the proportion of its internal resistance growth affected by batteries in the same cluster). The sample target value is obtained in two ways: actual adaptation results: the "healthy fitness" verified by disassembly after retirement (such as the actual fitness of G12 and the system is 0.85, because the SOH is consistent with the system average and the thermal risk is low); expert annotation: for some data (such as no disassembly data in the early operation stage), battery experts will annotate the fitness level according to the operation log (such as "good" corresponds to 0.8-1.0, "needs attention" corresponds to 0.7-0.8). The server adopts the "degradation stage perception" reinforcement learning framework to divide the battery life cycle into three stages (initial: number of cycles <500 times; mid-term: 500-1500 times; final stage: >1500 times), and different reward weights are set for different stages. The intelligent agent is the initial health state adaptation network (fully connected neural network), the environment is the historical operation data of the energy storage system, the state is the training feature, and the action is the output fitness prediction value. The reward (R) is dynamically adjusted based on the error between the predicted value and the sample target value (A*) and the degradation stage. Taking the training battery G12 (800 cycles, mid-term stage) as an example, the server performs the following training steps: State input: The battery state feature (S G12 ), system characteristics (S sys )(average SOH 90%, discharge rate 1.0C), integrated features (temperature coupling coefficient 0.8) are input into the initial network to obtain the fitness prediction Reward calculation: Basic reward: (If the sample target value (A*=0.85), then (R base =1-0.03=0.97); Stage weight adjustment: The mid-term stage focuses more on "internal resistance growth and system consistency" (weight 0.6), so the reward is adjusted to (R=R base ×0.6+R base×0.4×text{(matching degree between internal resistance growth rate and system average)})(G12 internal resistance growth rate is 0.04mΩ / month, system average is 0.035mΩ / month, matching degree is 0.9, so (R=0.97×0.6+0.97×0.4×0.9=0.582+0.349=0.931)). Policy update: The server updates the network parameters through the PPO (Proximal Policy Optimization) algorithm: calculates the policy gradient, adjusts the weight matrix of the hidden layer (such as enhancing the connection weights of neurons corresponding to the "matching degree of internal resistance growth rate"), and makes the predicted value Closer to (A*). Multi-stage generalization training: The server repeats the above process for batteries in the early stage (such as G05, 300 cycles, reward weights tilted towards "capacity consistency") and the late stage (such as G99, 2000 cycles, reward weights tilted towards "thermal risk control") to ensure that the network can perceive the fitness evaluation focus at different degradation stages. The server has undergone 500 rounds of reinforcement learning training (each round contains training data for 100 batteries), and the prediction error of the initial network has dropped from the initial ±0.12 to ±0.05. For example, the prediction value of the late stage battery G99 (2000 cycles, sample target value (A*=0.65)) was corrected from the initial 0.72 to 0.66, with an error of only 0.01. Finally, the server obtained a "target health state adaptation network" whose average fitness prediction accuracy on the test set (20 batteries that did not participate in the training) reached 92% (error ≤ 0.05). Through the above process, the server uses reinforcement learning based on degradation stage perception to enable the target health status adaptation network to dynamically adjust the evaluation focus at different stages, accurately quantify the health compatibility between the battery and the system, and provide a basis for core capacity decision-making that is more in line with actual degradation laws.

[0085] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent online capacity verification method for energy storage batteries based on deep learning. Figure 2 As shown, Figure 2A block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. In order to achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly. For illustrative purposes, the above description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Based on the above teachings, numerous modifications and variations are possible. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. An intelligent online capacity verification method for energy storage batteries based on deep learning, characterized in that: include: Obtain a set of historical operating data of the target battery and target energy storage system; Extracting target battery state features of the battery from an operating segment in a historical operating data set of the target battery and the target energy storage system; the target battery state features are obtained by encoding high-dimensional time series features of the battery through an autoencoder; determining, based on a degradation trajectory similarity metric between the target battery and the target battery state characteristics of the respective operation segments, a battery data set associated with the target battery from the historical operation data set of the target energy storage system; Based on the associated battery data set and the target battery, a target battery data set is obtained, and the target battery data set is input into a time series-graph convolution hybrid network; the time series-graph convolution hybrid network includes a preset battery state feature set, and the time series-graph convolution hybrid network is trained based on a historical operation data set of a training energy storage system; Through the time series-graph convolution hybrid network, the preset battery state features corresponding to each battery in the target battery data set are matched in the preset battery state feature set, and a multi-source feature collaborative analysis based on the multi-head graph attention mechanism is performed on each matched preset battery state feature to output a system-level degradation representation and degradation attribution report of the target battery data set; the system-level degradation representation is used to determine the health state compatibility between the target battery and the target energy storage system, and combined with the degradation attribution report to generate capacity decision recommendations.

2. The method according to claim 1, characterized in that The step of extracting target battery state characteristics of the battery from the operation segments in the historical operation data set of the target battery and the target energy storage system includes: Generate a feature chain of the battery instance to be evaluated based on the multi-dimensional battery fingerprint of the battery instance to be evaluated, and input the feature chain of the battery instance to be evaluated into a target feature compression encoder; the battery instance to be evaluated is an operation segment in a historical operation data set of the target battery or the target energy storage system, and the target feature compression encoder includes a high-dimensional time series feature set; Through the target feature compression encoder, in the high-dimensional time series feature set, the high-dimensional time series features corresponding to the multi-dimensional battery fingerprint of the battery instance to be evaluated are matched, cross-domain feature fusion and health factor mapping are performed on the matched high-dimensional time series features, and the target battery state features of the battery instance to be evaluated are output.

3. The method according to claim 2, characterized in that The method further comprises: Obtaining a first battery instance feature chain and inputting the first battery instance feature chain into an initial feature compression encoder; the first battery instance feature chain is obtained based on a historical operating data set of a first training energy storage system, and the initial feature compression encoder includes the high-dimensional time series feature set; Using the initial feature compression encoder, in the high-dimensional time series feature set, the high-dimensional time series features corresponding to each multi-dimensional battery fingerprint before the first time series node in the first battery instance feature chain are matched, cross-domain feature fusion and health factor mapping are performed on each matched high-dimensional time series feature to obtain the degradation deduction battery state feature corresponding to the first time series node; Projecting the degradation trajectory of the degradation deduction battery state feature corresponding to the first timing node to obtain an initial battery state space representation corresponding to the first timing node; the initial battery state space representation includes the health state value corresponding to each battery to be evaluated in the core capacity candidate pool; quantifying the health state of the initial battery state space representation to obtain a target battery state space representation corresponding to the first time series node; the target battery state space representation includes a first health state credible value corresponding to each battery to be evaluated in the core capacity candidate pool, and the core capacity candidate pool includes batteries corresponding to each multi-dimensional battery fingerprint in the first battery instance feature chain; Determining, from the target battery state space representation, a first health state credible value corresponding to the multidimensional battery fingerprint at the first time series node in the first battery instance feature chain; the first time series node is determined from each multidimensional battery fingerprint time series node in the first battery instance feature chain, and the first health state credible value is used to indicate a degradation correlation between the battery corresponding to the first time series node and a time series predecessor operation data set corresponding to the first time series node; Obtaining a first degradation representation cost based on first health state credible values ​​corresponding to the multi-dimensional battery fingerprint at each first time sequence node in the first battery instance feature chain; The encoder parameters of the initial feature compression encoder are iteratively optimized based on the first degradation representation cost to obtain the target feature compression encoder.

4. The method according to claim 3, characterized in that The initial feature compression encoder includes a feature encoder, a health factor mapper, a degradation trajectory fusion device and a credibility quantizer. The feature encoder is used to match high-dimensional time series features, the health factor mapper is used for health factor mapping, the degradation trajectory fusion device is used for cross-domain feature fusion, and the credibility quantizer is used to output a first health state credibility value; The iteratively optimizing the encoder parameters of the initial feature compression encoder based on the first degradation representation cost to obtain the target feature compression encoder includes: The encoder parameters of the health factor mapper, the degradation trajectory fuser and the credibility quantizer in the initial feature compression encoder are iteratively optimized based on the first degradation representation cost to obtain the target feature compression encoder.

5. The method according to claim 2, characterized in that The method further comprises: Obtain multi-dimensional operating condition information corresponding to each battery to be evaluated in the core capacity candidate pool; Performing feature extraction processing on the multi-dimensional operating condition information through a multimodal time series convolutional network to obtain high-dimensional time series features corresponding to each of the batteries to be evaluated; The high-dimensional time series feature set is obtained based on the high-dimensional time series features corresponding to each of the batteries to be evaluated.

6. The method according to claim 1, characterized in that The determining, based on the degradation trajectory similarity metric between the target battery and the target battery state characteristics of the operation segments, of an associated battery data set of the target battery from the historical operation data set of the target energy storage system comprises: Based on a degradation trajectory similarity metric between the target battery and the target battery state characteristics of the operation segments, determining a plurality of associated battery operation segments of the target battery from a historical operation data set of the target energy storage system according to the degradation trajectory similarity metric from low to high; The operation segments of the associated batteries are aligned with the charge and discharge cycle phases of the target energy storage system to obtain an associated battery data set of the target battery.

7. The method according to claim 1, characterized in that The method further comprises: Obtaining a second battery instance feature chain, and inputting the second battery instance feature chain into an original hybrid network; wherein the second battery instance feature chain is obtained based on a set of historical operating data of a second training energy storage system, and the original hybrid network includes the preset battery state feature set; By means of the original hybrid network, in the preset battery state feature set, the preset battery state features corresponding to each multidimensional battery fingerprint before the second time series node in the second battery instance feature chain are matched, and a multi-source feature collaborative analysis based on a multi-head graph attention mechanism is performed on each matched preset battery state feature to obtain a degradation deduction battery state feature corresponding to the second time series node, and based on the degradation deduction battery state feature corresponding to the second time series node, a second health state credible value corresponding to the multidimensional battery fingerprint at the second time series node in the second battery instance feature chain is obtained; the second time series node is determined from each multidimensional battery fingerprint time series node in the second battery instance feature chain, and the second health state credible value is used to indicate the degradation correlation between the battery corresponding to the second time series node and the time series preceding operation data set corresponding to the second time series node; Obtaining a second degradation representation cost based on the second health state credible value corresponding to each of the multi-dimensional battery fingerprints at each second time series node in the second battery instance feature chain; The encoder parameters of the original hybrid network are iteratively optimized based on the second degradation representation cost to obtain the temporal-graph convolutional hybrid network.

8. The method according to claim 7, characterized in that The original hybrid network includes a feature encoder, a feature coupler, and a credibility quantizer, wherein the feature encoder is used to match preset battery state features, the feature coupler is used for multi-source feature collaborative analysis based on a multi-head graph attention mechanism, and the credibility quantizer is used to output a second health state credibility value; The iteratively optimizing the encoder parameters of the original hybrid network based on the second degradation representation cost to obtain the time series-graph convolution hybrid network includes: The encoder parameters of the feature coupler and the credibility quantizer in the original hybrid network are iteratively optimized based on the second degradation representation cost to obtain the timing-graph convolution hybrid network.

9. The method according to claim 1, characterized in that The multi-source feature collaborative analysis based on the multi-head graph attention mechanism is performed on each matched preset battery state feature to output a system-level degradation representation of the target battery data set, including: combining the matched preset battery state features according to the battery sequence of the target battery data set to obtain a preset battery state feature set; Performing a multi-source feature collaborative analysis based on a multi-head graph attention mechanism on the preset battery state feature set to obtain a battery coupling feature set; the battery coupling feature set includes battery coupling features consistent with the number of batteries in the target battery data set; A battery coupling feature at a battery timing node of the target battery is obtained from the battery coupling feature set as a system-level degradation representation of the target battery data set.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images