Method for improving and optimizing operation efficiency of energy storage power station based on big data analysis

Abstract

The invention discloses an energy storage power station operation efficiency improving and optimizing method based on big data analysis, and relates to the technical field of energy storage power station operation management.The method comprises the steps that full life cycle data of batteries of the same model are collected and clustered based on a GS-ADC algorithm, three final clusters are obtained, and the stages are determined; for different stage clusters, constructing piecewise linear regression, an adaptive kernel function RBF neural network and a grey prediction GM (1, 1) model respectively; sOH, capacity attenuation acceleration and internal resistance growth rate are collected in real time through a battery management system, real-time feature vectors are formed, the weighted mahalanobis distance between the real-time feature vectors and each cluster center is calculated, stage judgment and model calling are carried out accordingly, and a battery adjustment strategy is set based on the residual life predicted at different stages. The method improves the stage division precision and prediction precision, and achieves the collaborative optimization of the operation efficiency of the energy storage power station and the service life of the battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage power station operation management, in particular to an energy storage power station operation efficiency improvement and optimization method based on big data analysis. BACKGROUND

[0002] As a key component of new power systems, the operation efficiency of energy storage power stations is closely related to battery performance management. The division of battery aging stages in the whole life cycle, the prediction of remaining life, and the dynamic adjustment strategy are the core links that affect the operation efficiency of energy storage power stations. The main defects of the existing technology are as follows: Firstly, in the division of battery aging stages, the existing technology relies on fixed experience thresholds or traditional clustering algorithms, without adaptive clustering based on the dynamic correlation of battery aging core features, resulting in fuzzy stage boundaries and easy clustering of different aging stage samples. Secondly, the design of the remaining life prediction model lacks stage adaptability. The existing technology often uses a single model to cover the whole life cycle of the battery, ignoring the different aging mechanisms of stable growth of the initial SEI film, accelerated attenuation in the middle stage, and sudden performance decline in the final stage, resulting in low prediction accuracy. Thirdly, the real-time operation adjustment strategy lacks dynamicity. The existing strategy is mostly based on fixed charging and discharging parameters or a single state indicator, without coordinated adjustment based on real-time feature vectors and the current aging stage, making it difficult to balance operation efficiency and battery life. Therefore, there is an urgent need for an energy storage power station operation efficiency improvement and optimization method based on big data analysis. SUMMARY

[0003] To overcome the shortcomings of the prior art, the present application provides an energy storage power station operation efficiency improvement and optimization method based on big data analysis, which solves the problems of insufficient battery stage division, life prediction and strategy adjustment of energy storage power stations.

[0004] To achieve the above purpose, the present application realizes the following technical scheme: an energy storage power station operation efficiency improvement and optimization method based on big data analysis, comprising: Step 1: Collecting full life cycle data of the same type of battery, clustering the data based on the gravitational search-adaptive density clustering GS-ADC algorithm to obtain three final clusters Q1, Q2 and Q3, and determining the stage of the final cluster, which is the initial stage, the middle stage and the final stage. Step 2: For clusters in the initial stage, the middle stage and the final stage, respectively, build a piecewise linear regression, an adaptive kernel function RBF neural network and a grey prediction GM(1,1) model to predict the remaining life of the battery in each stage. Step 3, real-time acquisition of SOH, capacity attenuation acceleration, and internal resistance growth rate by the battery management system BMS, formation of a real-time feature vector, calculation of the weighted Mahalanobis distance of the real-time feature vector and the clustering centers of Q1, Q2, and Q3, stage judgment and prediction model calling according to the distance, and setting of different adjustment strategies for the battery based on the predicted remaining useful life RUL in different stages.

[0005] As a further scheme of the application, the specific steps of clustering the full life cycle data of the same type of battery based on the gravitational search-adaptive density clustering GS-ADC algorithm are as follows: Taking SOH value, capacity attenuation acceleration, and internal resistance growth rate as the three-dimensional features for clustering, and performing min-max standardization on the obtained three-dimensional features to map them to the interval [0, 1]; Defining the potential energy of the sample, and the specific formula is where w j is the feature weight of sample j, d ij is the Euclidean distance between samples i and j, and ɑ is the attenuation coefficient; Determining the density threshold of each sample where is the average density of the 10 nearest samples around sample i, and max(E) is the maximum gravitational potential energy of all samples; Calculating the local density of each sample , and screening samples with and E i ≥ 0.6×max(E) as initial core samples, where d th is the distance threshold; For non-core sample i, calculating the gravitational attraction strength of each initial core sample k , and assigning the sample to the cluster to which the initial core sample with the strongest gravitational attraction belongs to form an initial cluster; Calculating the intra-cluster feature fluctuation coefficient CV of the initial cluster according to the formula CV=std(intra-cluster feature value) / mean(intra-cluster feature value) SOH , CV ron , and if the CV SOH of a cluster is greater than CV th or CV ron is greater than CV th1 , then the cluster is split according to the local minimum of the gravitational potential energy, where std() is the standard deviation function, mean() is the mean function, CV SOH is the SOH value fluctuation coefficient, CV ron is the capacity attenuation acceleration fluctuation coefficient, and CV th , CV th1 are the fluctuation coefficient thresholds; If the distance between the centers of two clusters is less than 0.2×max(d allIf the distance between the two clusters is less than the maximum distance between the cluster centers, the two clusters are merged into one cluster, and finally three clusters are retained, wherein max(d all ) is the maximum distance between all cluster centers; The intra-cluster distance proportion of each sample is calculated If r i > r th , it is determined as an isolated point, and it is attributed to the cluster closest to it and whose CV meets the requirements, wherein d cu is the distance set between the single sample and all other samples in the cluster; Based on the existing GS-ADC model, the cluster purity is verified, and after verification, the GS-ADC model based on the changed parameters is used to cluster the full life cycle data of the same type of battery again, and finally three clusters Q1, Q2 and Q3 are obtained.

[0006] As a further scheme of the application, the specific process of splitting the cluster according to the local minimum of gravitational potential energy is as follows: The sample set S of the cluster to be split is determined, and the potential energy values {E1, E2,..., En} of all samples are calculated, wherein n is the number of samples in the cluster; The samples in the cluster are sorted in descending order according to the SOH values of the samples, and an ordered sample sequence Sord={sq1, sq2,..., sqn} is obtained; For each sample sqk in the ordered sample sequence Sord except the first and last samples, if Eqk < Eq(k-1) and Eqk < Eq(k+1), i.e. the potential energy of the sample is less than that of the adjacent samples before and after it, then sqk is a local minimum point, and this point is taken as a splitting boundary, wherein 2≤k≤n-1; If there are multiple local minimum points, the one with the smallest potential energy is selected as the final splitting point; Taking the finally located local minimum point sqk as a boundary, the ordered sample sequence Sord is split into two parts: sub-cluster S1={sq1,..., sq(k-1)} and sub-cluster S2={sq(k+1),..., sqn}, and the minimum point sqk is removed; The CV SOH and CV ron of the two sub-clusters are calculated respectively, and if the two CV values of the two sub-clusters are not greater than the preset threshold, the splitting is effective; If a sub-cluster still does not meet the CV threshold, the splitting steps are repeated until the characteristic fluctuations of all sub-clusters meet the requirements.

[0007] As a further scheme of the application, the specific steps of verifying the cluster purity based on the existing GS-ADC model are as follows: Select 10 groups of batteries consistent with the battery model used in the energy storage power station to be optimized, and obtain data from each group of batteries according to the "full life cycle aging experiment": from a new battery, record a group of SOH, capacity decay acceleration and internal resistance growth rate data every 100 cycles until the battery is scrapped; Each group of batteries is labeled with initial, intermediate and final stage labels, and finally 10 verification subsets are formed, each containing 50 initial samples, 50 intermediate samples and 50 final samples, and the total verification set contains 1500 labeled samples; Put the 1500 labeled verification samples into the GS-ADC model with parameters completely consistent with the actual clustering to obtain three verification clusters, denoted as V1, V2 and V3; Statistically count the number of labeled samples belonging to the initial, intermediate and final stages in each verification cluster, calculate the cluster purity according to the formula (the number of samples of a certain stage with the most samples in the cluster / the total number of samples in the cluster), and take the certain stage with the most samples in the cluster as the stage where the cluster is located. Filter out clusters with purity < 90% as target clusters for this parameter adjustment; For the target cluster, calculate the difference between the core sample and the misclassified sample, if the difference < 0.2 x average potential energy difference, then adjust a, if a increases to 1.0 and the purity is still < 90%, then adjust the core sample density threshold , the average potential energy difference is the average potential energy difference between the core sample and the samples in the same cluster.

[0008] As a further scheme of the present application, the specific rules for adjusting a are: Set the initial a to 0.5, increase it in steps of "every +0.1", and the adjusted a should not exceed 1.0, recalculate the purity of the target cluster after each adjustment, if ≥ 90%, stop adjusting a; The specific rules for adjusting the core sample density threshold are as follows: For the core sample of the target cluster, extract its original adaptive density threshold ; Statistically count the samples with neighbor stage sample ratio > 20% in the core sample, calculate the and the average threshold of the core sample in the cluster , calculate the ratio of samples to total samples, if the ratio ≥ 80%, it is determined that the adaptive threshold is low, increase the original dynamic coefficient 0.3 in the threshold formula in steps of "every +0.1", and the maximum should not exceed 0.7; Recalculate the adaptive density threshold of all samples based on the new coefficient , reselect the core sample, and calculate the purity of the target cluster, if the purity ≥ 90%, lock the current dynamic coefficient and the corresponding adaptive threshold formula.

[0009] As a further scheme of the present application, after obtaining the final clusters Q1, Q2 and Q3, it is necessary to determine the stage of each cluster, i.e. initial stage, middle stage and final stage: calculate the distance dis(i,j) between the cluster Qi and the cluster Vj, find out the Vj corresponding to min(dis(i,j)) for Qi, and take the stage of Vj as the corresponding stage of Qi, wherein i, j∈[1, 3].

[0010] As a further scheme of the present application, the specific steps for constructing the piecewise linear regression model are as follows: Collect the single charge-discharge DOD value, cycle number and capacity retention rate of the target battery in the initial stage; Eliminate two types of abnormal samples: the single charge-discharge DOD value fluctuates more than the instruction value ΔD; the capacity retention rate fluctuates more than the previous cycle ΔR, wherein ΔD is the upper limit of the DOD fluctuation allowed by the charge-discharge system, and ΔR is the upper limit of the detection equipment accuracy; Set two DOD threshold values D1 and D2, and divide the effective data into three intervals: if the single charge-discharge DOD is less than or equal to D1, it is classified into the low DOD interval; if D1 < single charge-discharge DOD ≤ D2, it is classified into the medium DOD interval; if the single charge-discharge DOD is greater than D2, it is classified into the high DOD interval; wherein D1 < D2; For each DOD interval, take the cumulative cycle number of the samples in the interval as the independent variable, take the capacity retention rate after the corresponding cycle as the dependent variable, and use the least squares method to fit a linear equation to obtain three "cycle number-capacity retention rate" fitting curves in different DOD intervals; Substitute the current cycle number into the fitting equation of the corresponding DOD interval to obtain the current capacity retention rate, calculate the decay amount = current capacity retention rate - retirement threshold, and then obtain the predicted remaining useful life RUL = decay amount / fitting equation slope through the fitted decay rate, i.e. the slope of the fitting equation; Calculate the deviation rate of the predicted value and the measured value, and if the deviation rate is greater than the threshold value, adjust D1 and D2 again.

[0011] As a further scheme of the present application, the specific steps for constructing the adaptive kernel function RBF neural network are as follows: Collect the voltage platform pressure drop rate, temperature fluctuation amplitude and remaining useful life data of the target battery in the middle stage; Divide the samples into three groups according to the temperature fluctuation amplitude: if the temperature fluctuation amplitude is less than or equal to T1, it is classified into small fluctuation; if T1 < temperature fluctuation amplitude ≤ T2, it is classified into medium fluctuation; if the temperature fluctuation amplitude is greater than T2, it is classified into large fluctuation; wherein T1 and T2 are the lower limit and upper limit of the fluctuation amplitude; Each group of samples corresponds to a voltage platform pressure drop rate sequence, and the length is denoted as m; Determine the RBF neural network structure: the number of input layer neurons is the length m of the pressure drop rate sequence, the number of hidden layer neurons is h, and the number of output layer neurons is 1; The kernel function adopts a Gaussian kernel function, and the specific formula is Wherein, c is the mean value of each group of pressure drop rates, and sigma is dynamically adjusted according to temperature fluctuations: small fluctuations correspond to sigma1, medium fluctuations correspond to sigma2, and large fluctuations correspond to sigma3, wherein sigma3 < sigma2 < sigma1; Taking the grouped pressure drop rate sequence as input and the remaining life as output, the gradient descent method is used to minimize the loss function, and if the MSE of all groups is less than the threshold MSE0, it indicates that all training models reach the fitting accuracy; Test the trained model with several new medium-term battery data: input a group of pressure drop rate sequence, compare the deviation of the predicted remaining life and the measured value, if the deviation of a fluctuation group is greater than △L, adjust the sigma value or the number of hidden layer neurons h of the group, and retrain the model until the deviation is less than △L, wherein △L is the preset life deviation upper limit.

[0012] As a further scheme of the application, the specific steps for constructing the grey prediction GM(1, 1) model are: Collect the target battery's end-of-life internal resistance instantaneous increment and SOC fluctuation range data, and eliminate samples with internal resistance instantaneous increment < 0 and SOC fluctuation range < △SOC0, wherein △SOC0 is the minimum fluctuation range allowed by the charging and discharging strategy; Construct a non-negative original sequence X(0) = {x(0)(1), x(0)(2),..., x(0)(n)} of internal resistance instantaneous increments, wherein x(0)(i) is the internal resistance instantaneous increment recorded for the ith time; Once accumulate the original sequence X(0) to generate the sequence X(1)(k) = sum(X(0)(i)), wherein sum() is a summation function, and i ∈ [1, n]; Establish a grey differential equation: X(0)(k) + a × X(1)(k) = b, wherein a is a development coefficient, and b is a grey action amount; Solve the parameter vector [a, b] by the least square method T , to obtain the prediction equation of the accumulated sequence: ; Restore the original sequence by cumulative reduction to obtain the predicted value: , which is the predicted value of the internal resistance instantaneous increment for the k+1th time; Accumulate the predicted internal resistance instantaneous increment to obtain the cumulative internal resistance value Rpred, find the cycle number when Rpred first ≥ rth, and calculate the difference between the current cycle number and the cycle number, to obtain the preliminary remaining life L, and at the same time, calculate Rpred / rth; If (Rpred / rth) ≤ If L is not required to be corrected; if If (Rpred / rth)≤1, the remaining life needs to be corrected, and L×ε is obtained; if (Rpred / rth)>1, the battery needs to be immediately stopped, and L is set to 0; wherein, is a proximity threshold value, and ε is a correction coefficient.

[0013] As a further scheme of the present application, the specific operation process of stage judgment according to distance is as follows: The cluster centers of the clusters Q1, Q2 and Q3 are extracted, and are denoted as center C1: initial stage, C2: middle stage and C3: final stage. Each center contains three features: SOH, denoted as f1; capacity attenuation acceleration, denoted as f2; and internal resistance growth rate, denoted as f3. A center feature matrix is formed: ; The stage discrimination degree variance of each feature among the three centers is calculated, and the specific formula is as follows: wherein, is the mean value of the feature in the three centers, j∈[1,3]; The weight is calculated , and a weight matrix W=diag(wf1,wf2,wf3) is obtained. A real-time feature vector Y=[Y(f1),Y(f2),Y(f3)] is collected by the BMS, wherein Y(f1) is the SOH corrected by Kalman filtering, Y(f2) is the mean value of the capacity attenuation acceleration in the last 100 cycles, and Y(f3) is the mean value of the internal resistance growth rate in the day. First, the covariance matrix of the center features of the three stages is calculated , a weighted covariance matrix is constructed, and the weighted Mahalanobis distance of the real-time feature vector Y and each center is calculated, and the specific formula is as follows: , respectively, D(Y,C1), D(Y,C2) and D(Y,C3) are obtained, wherein, is the inverse matrix of the weighted covariance matrix; The minimum value among the three distances is selected, and the stage to which the center corresponding to the minimum value belongs is taken as the candidate stage. The difference △Dis between the minimum distance and the second minimum distance is calculated. If △Dis>△Dth, the candidate stage is the final stage; if △Dis≤△Dth, the stage at the previous time is maintained, wherein △Dth is a preset threshold value.

[0014] The present application provides an energy storage power station operation efficiency improvement and optimization method based on big data analysis, which has the following beneficial effects compared with the prior art: (1) The application adopts a gravity search-adaptive density clustering GS-ADC algorithm, takes the battery SOH, capacity attenuation acceleration and internal resistance growth rate as three-dimensional core features, combines adaptive density threshold setting and cluster purity verification iterative optimization, effectively solves the sample cluster problem caused by the dependence of the traditional clustering method on the fixed threshold, and ensures that the initial, intermediate and terminal cluster structures are highly matched with the actual battery aging law; (2) The application constructs a segmented linear regression, an adaptive kernel function RBF neural network and a grey prediction GM(1,1) model according to the difference mechanism of different battery aging stages, avoids the insufficient adaptability problem caused by the traditional single model covering the whole life cycle, and significantly improves the residual life prediction accuracy; (3) The application is based on the feature vectors and stage determination results collected by the BMS in real time, and combines the residual life RUL grading to develop a differential adjustment strategy, solves the contradiction between the traditional fixed parameter strategy and the "over-restriction efficiency" or "accelerated aging", and balances the operation efficiency of the energy storage power station and the whole life cycle utilization of the battery through the phased coordinated adjustment of parameters such as charge and discharge depth, voltage and rate. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The step flowchart of the application is shown in the figure; Figure 2 The step flowchart of the application is shown in the figure; Figure 3 The step flowchart of the application is shown in the figure. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0017] Embodiment 1

[0018] As Figure 1 , the application provides an energy storage power station operation efficiency improvement and optimization method based on big data analysis, which comprises: Step 1, collect full life cycle data of the same type of battery, and cluster the data based on the gravity search-adaptive density clustering GS-ADC algorithm, the specific operation steps are as follows: Take SOH value, capacity attenuation acceleration and internal resistance growth rate as three-dimensional clustering features, collect full life cycle data of the same type of battery, covering new batteries to scrap state; SOH value is the core indicator to measure the "current state" of battery aging, which quantifies the health degree by the ratio of current capacity to rated capacity, and can intuitively correspond to the stage division of initial stage (high SOH), middle stage (medium SOH) and end stage (low SOH). At the same time, it can also avoid the "false health" misjudgment caused by only looking at the current capacity by comparing with the initial state. Capacity attenuation acceleration focuses on the "dynamic trend" of aging, which can make up for the static lag of SOH, that is, it can distinguish batteries with the same SOH but different attenuation speeds, and avoid one-size-fits-all clustering. Its core advantage is to accurately capture the turning point of the aging stage: slow attenuation in the initial stage (acceleration close to 0), acceleration in the middle stage (absolute value of acceleration increases), and cliff-like attenuation in the end stage (acceleration suddenly rises), making the stage boundary clearer. Capacity attenuation rate is calculated every 100 cycles, and the specific formula is: (this cycle capacity attenuation rate - last cycle capacity attenuation rate) / cycle difference; Internal resistance growth rate reflects the "internal source" of aging, which is a direct mapping of internal structure degradation. It can penetrate the short-term fluctuations of SOH, avoid surface judgment deviation, and distinguish different failure modes, so that batteries in the same cluster have similar aging mechanisms, which is convenient for subsequent targeted protection. In addition, the exponential rise of internal resistance growth rate can accurately identify batteries approaching failure, avoiding confusion between middle-stage low SOH and end-stage batteries. Internal resistance growth rate is calculated daily, and the specific formula is: (today's internal resistance - yesterday's internal resistance) / yesterday's internal resistance; The obtained three-dimensional features are respectively subjected to min-max standardization, which are mapped to the interval [0, 1]; In order to avoid the deviation caused by subjective allocation of weight based on experience, according to the battery aging mechanism, the correlation degree between the three features and the actual cycle life is calculated by gray correlation analysis, and the feature weight is allocated according to the proportion of correlation degree; Set the GS-ADC model parameters: define the sample potential energy, and the specific formula is The higher the potential energy, the stronger the attraction to the surrounding samples, and the more likely it is to become a core sample, where w j is the feature weight of sample j, d ij is the Euclidean distance between sample i and j, and ɑ is the attenuation coefficient; The density threshold of each sample is determined by the local density mean × dynamic coefficient , which ensures that the threshold in the high-density area is higher, avoiding over-splitting of clusters, where is the average density of the 10 nearest samples around sample i, and max(E) is the maximum gravitational potential energy of all samples; The local density of each sample is calculated , and the samples that meet the conditions of and E i ≥ 0.6 × max(E) are selected as the initial core samples, where d th is the distance threshold; For non-core samples i, calculate its gravitational strength with each initial core sample k , the sample is attributed to the cluster to which the initial core sample with the strongest gravitational strength belongs, forming an initial cluster; Calculate the intra-cluster characteristic fluctuation coefficient CV of the initial cluster according to the formula CV = std (intra-cluster characteristic value) / mean (intra-cluster characteristic value) SOH , CV ron , if the CV SOH of a cluster is greater than CV th or CV ron is greater than CV th1 , split the cluster according to the local minimum of gravitational potential energy, wherein std() is the standard deviation function, mean() is the mean function, CV SOH is the SOH value fluctuation coefficient, CV ron is the capacity attenuation acceleration fluctuation coefficient, CV th , CV th1 is the fluctuation coefficient threshold, which needs to be set according to the actual situation; CV SOH reflects the relative fluctuation of the SOH value in the cluster. The health status of the battery at the same stage should be relatively small. If CV SOH is greater than CV th , it means that samples of different SOH levels may be mixed in the cluster (such as containing initial high SOH and medium low SOH samples at the same time), which does not conform to the single stage characteristic; CV ron reflects the relative fluctuation of the aging speed in the cluster. The aging speed trend at the same stage should be consistent. If CV ron is greater than CV th1 , it means that the cluster may contain slow and fast decay samples, which violates the aging law within the stage; Therefore, when the two CVs exceed the threshold, it indicates that the samples in the cluster are not in the same aging stage and need to be split to ensure the stability of each cluster; If the distance between the centers of two clusters is less than 0.2 x max (d all ), merge the two clusters into one cluster, and finally retain three clusters, wherein max (d all ) is the maximum distance between all cluster centers; Calculate the intra-cluster distance proportion of each sample If r i is greater than r th , it is determined as an isolated point and is attributed to the cluster with the closest distance and CV that meets the requirements, wherein d cu is the distance set between a single sample and all other samples in the cluster; To further verify the accuracy of clustering, cluster purity verification needs to be performed based on the existing GS-ADC model; After the cluster purity is verified, the full life cycle data of the same type of battery collected based on the GS-ADC model after the modified parameters is clustered again to obtain three final clusters Q1, Q2 and Q3; After obtaining the three final clusters, it is necessary to determine the stages of the three clusters, i.e. initial stage, middle stage and end stage: calculate the center distance dis(i,j) of the cluster Qi and the cluster Vj, find the Vj corresponding to min(dis(i,j)) for Qi, and take the stage of Vj as the corresponding stage of Qi, wherein i, j ∈ [1, 3].

[0019] Step 2: For the clusters in the initial stage, the middle stage and the end stage, respectively, a piecewise linear regression, an adaptive kernel function RBF neural network and a grey prediction GM(1,1) model are constructed to predict the remaining life of the battery in each stage. For the cluster in the initial stage, the specific steps of constructing the piecewise linear regression model are as follows: Collect the single charge-discharge DOD value, cycle number and capacity retention rate of the target battery in the initial stage. Eliminate two types of abnormal samples: ① The single charge-discharge DOD value fluctuates more than △D (△D is the upper limit of DOD fluctuation allowed by the charge-discharge system); ② The capacity retention rate fluctuates more than △R (△R is the upper limit of the accuracy of the detection equipment) than the previous cycle. According to the conventional charge-discharge strategy of the target battery, set D1 and D2 as two DOD thresholds, and divide the effective data into three intervals: if the single charge-discharge DOD is less than or equal to D1, it is classified as a low DOD interval; if D1 < single charge-discharge DOD ≤ D2, it is classified as a medium DOD interval; if the single charge-discharge DOD is greater than D2, it is classified as a high DOD interval; wherein D1 < D2. For each DOD interval, use the cumulative cycle number of the samples in the interval as the independent variable (reflecting the cumulative aging process) and the capacity retention rate after the corresponding cycle as the dependent variable (reflecting the current aging state), and use the least squares method to fit a linear equation to obtain three "cycle number-capacity retention rate" fitting curves for different DOD intervals, which correspond to the initial capacity decay law under different single load intensities. Substitute the current cycle number into the fitting equation of the corresponding DOD interval to obtain the current capacity retention rate, calculate the decay amount = current capacity retention rate - retirement threshold, and then combine the fitted decay rate (i.e. the slope of the fitting equation) to obtain the predicted remaining useful life RUL = decay amount / fitting equation slope. Calculate the deviation rate of the predicted value and the measured value, and if the deviation rate is greater than the threshold, adjust D1 and D2 again. For the cluster in the middle stage, the specific steps of constructing the adaptive kernel function RBF neural network are as follows: Collect the voltage platform pressure drop rate, temperature fluctuation amplitude and remaining life data of the target battery in the middle stage. The samples are divided into three groups according to the temperature fluctuation amplitude: small fluctuation (temperature fluctuation amplitude ≤ T1), medium fluctuation (T1 < temperature fluctuation amplitude ≤ T2), and large fluctuation (temperature fluctuation amplitude > T2), wherein T1 and T2 are set according to the temperature range of the target battery in use environment; Each group of samples corresponds to a voltage platform pressure drop rate sequence, and the length is denoted as m; The RBF neural network structure is determined: the number of input layer neurons is the length m of the pressure drop rate sequence, the number of hidden layer neurons is h (determined by k-fold cross-validation, usually 5 or 10), and the number of output layer neurons is 1 (output remaining life); The kernel function adopts a Gaussian kernel function, and the specific formula is , wherein c is the mean of each group of pressure drop rates, and σ is dynamically adjusted according to temperature fluctuation: σ1 for small fluctuation, σ2 for medium fluctuation, and σ3 for large fluctuation, wherein σ3 < σ2 < σ1; The grouped pressure drop rate sequence is input, and the remaining life is output. The gradient descent method is used to minimize the loss function. If the MSE of all groups is ≤ threshold MSE0, it indicates that all prediction models reach the fitting accuracy; Test the model with several new groups of medium-term battery data (covering small, medium, and large temperature fluctuations): input a group of pressure drop rate sequences, compare the deviation between the model predicted remaining life and the measured value, if the deviation of a fluctuation group > △L (△L is the preset life deviation upper limit), adjust the σ value or the number of hidden layer neurons h of this group, and retrain the model until the deviation ≤ △L; For the end-stage cluster, the specific steps of constructing the gray prediction GM(1,1) model are as follows: Collect the target battery's end-stage internal resistance instantaneous increment and SOC fluctuation range data, and eliminate samples with internal resistance instantaneous increment < 0 (no negative growth of internal resistance) and SOC fluctuation range < △SOC0 (△SOC0 is the minimum fluctuation range allowed by the charging and discharging strategy); Construct a non-negative original sequence X(0) = {x(0)(1), x(0)(2),..., x(0)(n)} of internal resistance instantaneous increment, wherein x(0)(i) is the internal resistance instantaneous increment recorded for the i-th time; Perform once accumulation on the original sequence X(0) to generate a sequence X(1)(k) = sum(X(0)(i)), which weakens the randomness of the original data, wherein i ∈ [1, n]; Combine the two to establish a gray differential equation: X(0)(k) + a × X(1)(k) = b, wherein a is the development coefficient (reflecting the internal resistance growth trend), and b is the gray action (reflecting the driving strength); Solve the parameter vector [a, b] using the least squares method T , and obtain the prediction equation of the accumulated sequence: ; The predicted value of the original sequence is obtained by cumulative reduction: , which is the predicted value of the k+1th instantaneous increment of internal resistance; The cumulative internal resistance value Rpred is obtained by sequentially accumulating the predicted internal resistance instantaneous increment, the cycle number when Rpred first ≥ rth is found, and the difference between the current cycle number is obtained to obtain the preliminary remaining life L, and Rpred / rth is calculated; If (Rpred / rth)≤ , L does not need to be corrected; if < (Rpred / rth)≤1, the remaining life needs to be corrected to obtain L×ε; if (Rpred / rth) > 1, the battery needs to be immediately stopped, and L is set to 0; wherein, is the proximity threshold, and ε is the correction coefficient, both of which need to be adjusted according to the actual situation.

[0020] Step 3, real-time feature vectors are formed by collecting SOH, capacity attenuation acceleration, and internal resistance growth rate through the battery management system BMS, the weighted Mahalanobis distance of the real-time feature vectors and the clustering centers of Q1, Q2, and Q3 is calculated, and stage judgment and prediction model calling are performed according to the distance. The specific operation process is: The feature data of the clustering centers of Q1, Q2, and Q3 is extracted, denoted as center C1: initial stage, C2: middle stage, and C3: final stage. Each center contains three features: SOH, denoted as f1; capacity attenuation acceleration, denoted as f2; and internal resistance growth rate, denoted as f3, forming a center feature matrix: ; The stage discrimination variance of each feature among the three centers is calculated, and the specific formula is The smaller the variance, the more significant the difference between the features in different stages, and the higher the discrimination. Wherein, is the mean value of the feature in the three centers, j∈[1,3]; According to the principle that the higher the weight, the higher the discrimination, the weight is calculated to obtain the weight matrix W=diag(wf1,wf2,wf3); The real-time feature vector Y=[Y(f1),Y(f2),Y(f3)] is collected through the BMS, wherein Y(f1) is the SOH corrected by Kalman filtering, Y(f2) is the average of the capacity attenuation acceleration in the last 100 cycles, and Y(f3) is the average of the internal resistance growth rate on the same day; The covariance matrix of the three stage center features is calculated first , which is combined with W to construct the weighted covariance matrix The weighted Mahalanobis distance between the real-time feature vector Y and each center is calculated, and the specific formula is , respectively, D(Y, C1) (distance from the initial center), D(Y, C2) (distance from the middle center), and D(Y, C3) (distance from the final center), wherein, is the inverse matrix of the weighted covariance matrix; The minimum value of the three distances is selected, and the center corresponding to the minimum value belongs to the candidate stage. The difference between the minimum distance and the second minimum distance is calculated as △Dis. If △Dis > △Dth, the candidate stage is the final stage. If △Dis≤△Dth, the stage at the previous time is maintained. The preset threshold value △Dth can be determined based on historical data statistics, and the value range is [0.1, 0.2]. The specific rules for calling the prediction model are as follows: If the final stage is the initial stage, a trained piecewise linear regression model is called to predict the remaining useful life; If the final stage is the middle stage, a trained adaptive kernel function RBF neural network model is called to predict the remaining useful life; If the final stage is the final stage, a trained gray prediction GM(1, 1) model is called to predict the remaining useful life; Based on the remaining useful life RUL predicted in different stages, different adjustment strategies for the battery are set, and the specific content is as follows: When the battery is in the initial stage, if RUL > R1, the charge and discharge depth is relaxed by 5% compared with the conventional threshold value (the standard depth of the same type of battery in the initial stage), the charge cut-off voltage is maintained at the initial conventional value U1, and the discharge cut-off voltage is maintained at the initial conventional value U2. If R2 < RUL≤R1, the charge and discharge depth is tightened by 10% compared with the conventional threshold value, the charge cut-off voltage is still U1, and the discharge cut-off voltage U2 needs to be increased. If RUL≤R2, the charge and discharge depth is tightened by 15% compared with the conventional threshold value, the charge cut-off voltage needs to be reduced compared with U1 to avoid the SEI film from entering the cracking stage due to depth cycling, wherein R1 is the initial long-life threshold value, and R2 is the initial medium-life threshold value. When the battery is in the middle stage, if RUL > R3 and △T≤T0, the charge rate is maintained at the middle conventional value C1, and the discharge cut-off voltage needs to be increased compared with the middle conventional value U3 to reduce polarization accumulation. If RUL > R3 and △T > T0, the charge rate needs to be reduced compared with C1, and the discharge cut-off voltage is still maintained at the increased value. If RUL≤R3, regardless of the size of △T, the charge rate needs to be reduced compared with C1, and the discharge cut-off voltage needs to be increased compared with U3, wherein R3 is the middle long-life threshold value, T0 is the temperature stability threshold value, and △T is the temperature fluctuation value collected by the BMS in real time. When the battery is in the end stage, if R4 < RUL≤R5, the charging cutoff voltage needs to be reduced compared with the conventional value U4, the maximum discharge power needs to be reduced compared with the rated power, and the SOC operating interval needs to be narrowed by 10% (5% for each of the upper limit and the lower limit) compared with the conventional interval; If RUL≤R4, the charging cutoff voltage needs to be reduced compared with U4, the maximum discharge power needs to be reduced compared with the rated power, and the SOC operating interval needs to be narrowed by 20% (10% for each of the upper limit and the lower limit) compared with the conventional interval, wherein R4 is an end-stage safety threshold, and R5 is an end-stage critical threshold.

[0021] Embodiment 2

[0022] This embodiment continues to disclose a method for splitting an existing cluster according to local minima of gravitational potential energy on the basis of Embodiment 1, as shown in FIG. 2, the specific contents include: Figure 2 determining a sample set S={s1, s2,...,sn} of the cluster to be split (n is the number of samples in the cluster), and calculating potential energy values {E1, E2,...,En} of all samples, wherein a larger potential energy value indicates that a sample is more strongly associated with other samples in the cluster (the gravitational force of surrounding samples on the sample is more concentrated); sorting the samples in the cluster in descending order according to SOH values of the samples, because SOH is a core representation of the aging stage, and the aging degrees of adjacent samples after sorting are more similar, thereby obtaining an ordered sample sequence Sord={sq1, sq2,...,sqn}; for each sample sqk (2≤k≤n-1) in the ordered sample sequence Sord except the first and last samples, if Eqk < Eq(k-1) and Eqk < Eq(k+1), i.e., the potential energy of the sample is smaller than that of the adjacent samples before and after the sample, then sqk is a local minimum point, the point is a weakly associated part in the cluster, the gravitational force of the sample on the adjacent samples before and after the sample is weak, and the sample is suitable as a splitting boundary; if there are multiple local minimum points, the point with the smallest potential energy is selected as the final splitting point, so as to ensure that the internal association of the two sub-clusters after splitting is the strongest; taking the finally located local minimum point sqk as a boundary, the ordered sample sequence Sord is split into two parts: a sub-cluster S1={sq1,...,sq(k-1)} (high SOH end, light aging degree) and a sub-cluster S2={sq(k+1),...,sqn} (low SOH end, heavy aging degree), and the local minimum point sqk is removed (if the potential energy is too low, the local minimum point sqk is determined as a noise sample in the cluster); CV SOH and CV ron of the two sub-clusters are calculated, respectively, if both CV SOH and CV ron of the two sub-clusters are not greater than a preset threshold, the splitting is valid. If a sub-cluster still does not meet the CV threshold, repeat the above splitting step until the characteristic fluctuations of all sub-clusters meet the requirements.

[0023] Embodiment 3

[0024] This embodiment continues to disclose a method for verifying the cluster purity of the GS-ADC model on the basis of Embodiments 1 and 2, as shown in the following specific process: Figure 3 Select 10 groups of batteries identical to the battery model used in the energy storage power station to be optimized, and obtain data from each group of batteries according to the "full life cycle aging experiment": from a new battery (SOH = 100%), record a group of SOH, capacity decay acceleration and internal resistance growth rate data every 100 cycles until the battery is scrapped (SOH ≤ 40%) ; label each group of batteries with initial (SOH ≥ 80%), intermediate (60% ≤ SOH < 80%) and final (SOH < 60%) stage labels, and finally form 10 verification subsets, each containing 50 initial samples, 50 intermediate samples and 50 final samples, for a total of 1500 labeled samples in the verification set; Put the 1500 labeled verification samples into the GS-ADC model with parameters completely consistent with the actual clustering to obtain three verification clusters, denoted as V1, V2 and V3; Statistically count the number of labeled samples in each verification cluster belonging to the initial, intermediate and final stages, calculate the cluster purity according to the formula (the number of samples in a certain stage with the most samples in the cluster / the total number of samples in the cluster), and take the stage with the most samples in the cluster as the stage of the cluster. Filter out clusters with purity < 90% as target clusters for this parameter adjustment; For the target cluster, calculate the difference in gravitational potential energy between the core sample and the misclassified sample. If the difference is < 0.2 x average potential energy difference (average potential energy difference is the average potential energy difference between the core sample and the samples in the same cluster), it is determined that the gravitational discrimination is insufficient. At this time, ɑ needs to be adjusted. The specific adjustment rule is: set the initial ɑ to 0.5, increase it by "0.1" every time in a stepwise manner, and the adjusted ɑ should not exceed 1.0. After each adjustment, the purity of the target cluster needs to be recalculated. If it is ≥ 90%, stop adjusting ɑ. If ɑ increases to 1.0 and the purity is still < 90%, adjust the core sample density threshold , which is specifically operated as follows: Extract the original adaptive density threshold of the core sample of the target cluster; Statistically count the samples with a neighbor ratio of > 20% of the samples in the core sample, calculate the of these samples , and calculate ​If the ratio of the sample to the total sample is ≥80%, it is determined that the adaptive threshold is too low, resulting in insufficient representation of the core sample, and the threshold formula needs to be adjusted. The original dynamic coefficient is 0.3 (0.3 in the formula), which is increased by a step of +0.1 each time, and the maximum is not more than 0.7. The adaptive density threshold of all samples is recalculated based on the new coefficient The core sample is re-screened, and the purity of the target cluster is calculated. If the purity is ≥90%, the current dynamic coefficient and the corresponding adaptive threshold formula are locked.

[0025] Some data in the above formula are dimensionless numerical calculations, and the contents not described in detail in the specification are all prior art known to those skilled in the art.

[0026] The above examples are only used to illustrate the technical method of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis, characterized in that, include: Step 1: Collect full life cycle data of the same type of battery, and cluster them based on the Gravity Search-Adaptive Density Clustering (GS-ADC) algorithm to obtain three final clusters Q1, Q2, and Q3. Then determine the stage of the final clusters, which are the initial stage, the middle stage, and the final stage. Step 2: For clusters in the initial, middle and final stages, construct piecewise linear regression, adaptive kernel function RBF neural network and grey prediction GM(1,1) model respectively to predict the remaining life of the battery in each stage. Step 3: Real-time data collection of SOH, capacity decay acceleration, and internal resistance growth rate is performed through the Battery Management System (BMS) to form a real-time feature vector. The weighted Mahalanobis distance between the vector and the cluster centers of Q1, Q2, and Q3 is calculated. Based on the distance, stage judgment and prediction model are called. Based on the predicted remaining life (RUL) of different stages, different adjustment strategies for the battery are set.

2. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, The specific steps for clustering the entire lifecycle data of batteries of the same model based on the Gravity Search-Adaptive Density Clustering (GS-ADC) algorithm are as follows: The SOH value, capacity decay acceleration, and internal resistance growth rate were used as the three-dimensional clustering features, and the obtained three-dimensional features were min-max normalized and mapped to the [0,1] interval. The sample potential is defined by the following formula: , where w j Let d be the feature weights of sample j. ij Let be the Euclidean distance between samples i and j, and α be the attenuation coefficient; Determine the density threshold for each sample. ,in, Let be the average density of the 10 nearest samples around sample i, and max(E) be the maximum gravitational potential energy of all samples. Calculate the local density for each sample ,filter And E i Samples with a ≥0.6×max(E) value are used as the initial core samples, where d th Distance threshold; For a non-core sample i, calculate the gravitational interaction strength between it and each initial core sample k. The samples are assigned to the cluster of the initial core sample with the strongest gravitational influence, thus forming the initial cluster; The intra-cluster eigenvalue CV of the initial cluster is calculated using the formula: CV = std(intra-cluster eigenvalue) / mean(intra-cluster eigenvalue). SOH CV ron If a certain cluster CV SOH >CV th or CV ron >CV th1 The cluster is then partitioned according to the local minimum of gravitational potential energy, where std() is the standard deviation function, mean() is the mean function, and CV is the mean value. SOH CV is the fluctuation coefficient of SOH value. ron CV is the capacity decay acceleration fluctuation coefficient. th CV th1 The threshold for the volatility coefficient; If the distance between the centers of the two clusters is less than 0.2 × max(d) all If the two clusters are merged into one, then three clusters are ultimately retained, where max(d) all ) represents the maximum distance between all cluster centers; Calculate the intra-cluster distance percentage for each sample If r i >r th If it is an isolated point, it is classified as an isolated point and assigned to the cluster with the nearest CV that meets the requirements, where d cu It is the set of distances between a single sample and all other samples in the cluster; Cluster purity is verified based on the existing GS-ADC model. After the verification is successful, the collected full life cycle data of the same type of battery needs to be clustered again based on the GS-ADC model with modified parameters to obtain three final clusters Q1, Q2 and Q3.

3. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 2, characterized in that, The specific process for splitting the cluster based on the local minimum of gravitational potential energy is as follows: Determine the sample set S={s1,s2,...,sn} of the cluster to be split, and calculate the potential energy values ​​{E1,E2,...,En} of all samples, where n is the number of samples in the cluster; The samples within a cluster are sorted in descending order according to their Soh values ​​to obtain an ordered sample sequence Sord={sq1,sq2,...,sqn}; For each sample sqk in the ordered sample sequence Sord, excluding the first and last samples, if Eqk < Eq(k-1) and Eqk < Eq(k+1), that is, the potential energy of the sample is less than the potential energy of the previous and next adjacent samples, then sqk is a local minimum point. This point is used as the split boundary, where 2 ≤ k ≤ n-1. If there are multiple local minima, the point with the lowest potential energy is selected as the final split point. Using the finally located local minimum point sqk as the boundary, the ordered sample sequence Sord is split into two parts: sub-cluster S1={sq1,...,sq(k-1)} and sub-cluster S2={sq(k+1),...,sqn}, and the local minimum point sqk is removed; Calculate the CV of the two subclusters separately. SOH CV ron If both CV values ​​of the two subclusters are not greater than the preset threshold, then the split is valid; If a sub-cluster still does not meet the CV threshold, repeat the above splitting steps until the feature fluctuations of all sub-clusters meet the requirements.

4. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, The specific steps for verifying cluster purity based on the existing GS-ADC model are as follows: Ten groups of batteries with the same model as the batteries used in the energy storage power station to be optimized were selected. Data for each group of batteries was obtained according to the "full life cycle aging test": starting from the new battery, data on SOH, capacity decay acceleration, and internal resistance growth rate were recorded every 100 cycles until the battery was scrapped. Each group of batteries is labeled with stage tags for the initial, middle and final stages, ultimately forming 10 validation subsets, each containing 50 samples from the initial stage, 50 samples from the middle stage, and 50 samples from the final stage, for a total of 1500 labeled samples in the validation set. Substituting 1500 labeled validation samples into the GS-ADC model with parameters completely consistent with the actual clustering, we obtained 3 validation clusters, denoted as V1, V2, and V3. The number of labeled samples belonging to the initial, middle and final stages in each validation cluster is counted. The cluster purity is calculated according to the formula: (number of samples in the stage with the largest number in the cluster / total number of samples in the cluster). The stage with the largest number in the cluster is taken as the stage of the cluster. Clusters with purity <90% are selected as the target clusters for this parameter adjustment. For the target cluster, calculate the gravitational potential energy difference between its core sample and misclustered sample. If the difference is less than 0.2 × the average potential energy difference, then α needs to be adjusted. If the purity is still less than 90% after increasing α to 1.0, then adjust the core sample density threshold. The average potential energy difference is the average potential energy difference between the core sample and the samples in the same cluster.

5. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 4, characterized in that, The specific rules for adjusting α are as follows: Set the initial α to 0.5, and increase it stepwise in increments of 0.1 each time. The maximum adjusted α should not exceed 1.

0. After each adjustment, the purity of the target cluster needs to be recalculated. If it is ≥90%, stop adjusting α. Adjusting the core sample density threshold The specific rules are as follows: For the core samples of the target cluster, extract their original adaptive density threshold. ; In the core statistical sample, for samples where the proportion of samples from different stages in the neighborhood is greater than 20%, calculate the percentage of these samples. Average threshold of core samples within the cluster ,calculate The ratio of the sample to the total sample is used. If the ratio is ≥80%, the adaptive threshold is judged to be too low. The original dynamic coefficient of 0.3 in the threshold formula is increased stepwise by "+0.1 each time", with a maximum of 0.

7. The adaptive density threshold for all samples was recalculated based on the new coefficients. The core samples are re-screened, and the purity of the target cluster is calculated. If the purity is ≥90%, the current dynamic coefficient and the corresponding adaptive threshold formula are locked.

6. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, After obtaining the final clusters Q1, Q2, and Q3, it is necessary to determine the stage of each cluster, namely the initial stage, the middle stage, and the final stage: calculate the center distance dis(i,j) between cluster Qi and cluster Vj respectively. For Qi, find the Vj corresponding to min(dis(i,j)) and take the stage of Vj as the stage corresponding to Qi, where i and j ∈ [1,3].

7. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, The specific steps for constructing a piecewise linear regression model are as follows: Collect the DOD value, cycle number, and capacity retention rate of the target battery during the initial charge-discharge cycle. Two types of abnormal samples were removed: the fluctuation of the DOD value in a single charge / discharge cycle compared to the command value was greater than ΔD; and the fluctuation of the capacity retention rate compared to the previous cycle was greater than ΔR. Here, ΔD is the upper limit of the allowable DOD fluctuation of the charge / discharge system, and ΔR is the upper limit of the accuracy of the detection equipment. Two DOD thresholds, D1 and D2, are set, and the valid data is divided into three intervals: if the DOD of a single charge-discharge cycle is less than or equal to D1, it is classified as the low DOD interval; if D1 is less than the DOD of a single charge-discharge cycle and less than or equal to D2, it is classified as the medium DOD interval; if the DOD of a single charge-discharge cycle is greater than or equal to D2, it is classified as the high DOD interval; where D1 is less than D2. For each DOD interval, the cumulative number of cycles for the samples within that interval is used as the independent variable, and the capacity retention rate after the corresponding cycle is used as the dependent variable. The least squares method is used to fit the linear equation, resulting in three fitting curves for the "number of cycles - capacity retention rate" of different DOD intervals. Substitute the current cycle number into the fitting equation of the corresponding DOD interval to obtain the current capacity retention rate, calculate the decay amount = current capacity retention rate - retirement threshold, and then obtain the predicted remaining lifetime RUL = decay amount / fitting equation slope through the fitted decay rate, i.e. the slope of the fitted equation. Calculate the deviation rate between the predicted value and the measured value. If the deviation rate is greater than the threshold, readjust D1 and D2.

8. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, The specific steps for constructing an adaptive kernel function RBF neural network are as follows: Collect data on the voltage plateau rate of voltage drop, temperature fluctuation amplitude, and remaining lifetime of the target battery during its mid-term operation. The samples are divided into three groups according to the temperature fluctuation range: if the temperature fluctuation range is ≤ T1, it is classified as small fluctuation; if T1 < temperature fluctuation range ≤ T2, it is classified as medium fluctuation; if the temperature fluctuation range is > T2, it is classified as large fluctuation; where T1 and T2 are the lower limit and upper limit of the fluctuation range, respectively. Each sample group corresponds to a voltage plateau rate of drop sequence, with a length denoted as m; The RBF neural network structure is determined as follows: the number of neurons in the input layer is the length m of the voltage drop rate sequence, the number of neurons in the hidden layer is h, and the number of neurons in the output layer is 1. The kernel function used is the Gaussian kernel function, and the specific formula is as follows: Where c is the average pressure drop rate of each group, and σ is dynamically adjusted with temperature fluctuations: small fluctuations correspond to σ1, medium fluctuations correspond to σ2, and large fluctuations correspond to σ3, where σ3 < σ2 < σ1; Using the grouped pressure drop rate sequence as input and remaining lifetime as output, the loss function is minimized using gradient descent. If the MSE of all groups is less than or equal to the threshold MSE0, it indicates that all trained models have reached the fitting accuracy. Test the trained model with several sets of new mid-term battery data: Input a set of voltage drop rate sequences and compare the deviation between the model's predicted remaining lifetime and the measured value. If the deviation of a certain fluctuation group is greater than ΔL, adjust the σ value or the number of hidden layer neurons h of that group and retrain the model until the deviation is less than or equal to ΔL, where ΔL is the preset upper limit of lifetime deviation.

9. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, The specific steps for constructing the grey prediction GM(1,1) model are as follows: Collect data on the instantaneous increase in internal resistance and the range of SOC fluctuation at the end of the target battery period, and remove samples with an instantaneous increase in internal resistance < 0 and a range of SOC fluctuation < ΔSOC0, where ΔSOC0 is the minimum fluctuation range allowed by the charging and discharging strategy. Construct a non-negative original sequence X(0)={x(0)(1),x(0)(2),...,x(0)(n)} of instantaneous internal resistance increments, where x(0)(i) is the instantaneous internal resistance increment of the i-th record; The original sequence X(0) is summed once to generate the sequence X(1)(k)=sum(X(0)(i)), where sum() is the summation function and i∈[1,n]. Establish the gray differential equation: X(0)(k)+a×X(1)(k)=b, where a is the development coefficient and b is the gray action quantity; Solve for the parameter vector [a, b] using the least squares method. T The prediction equation for the cumulative sequence is obtained as follows: ; The predicted value of the original sequence is obtained by cumulative subtraction: That is, the predicted value of the instantaneous increment of internal resistance in the (k+1)th iteration; The cumulative internal resistance value Rpred is obtained by successively accumulating the predicted instantaneous increment of internal resistance. The number of cycles when Rpred first ≥ rth is found, and the difference between Rpred and the current number of cycles is calculated to obtain the preliminary remaining lifetime L. At the same time, Rpred / rth is calculated. If (Rpred / rth)≤ If L does not need to be corrected; If (Rpred / rth) ≤ 1, the remaining lifespan needs to be adjusted to obtain L × ε; if (Rpred / rth) > 1, the battery needs to be immediately deactivated, and L should be set to 0; where, To approximate the threshold, ε is a correction coefficient.

10. The method for improving and optimizing the operating efficiency of energy storage power stations based on big data analysis according to claim 1, characterized in that, The specific operational procedure for determining the stage based on distance is as follows: Extract the cluster centers of clusters Q1, Q2, and Q3, denoted as C1 (early stage), C2 (middle stage), and C3 (late stage). Each center contains three features: SOH (saturation loss), denoted as f1; capacity decay acceleration, denoted as f2; and internal resistance growth rate, denoted as f3, forming a center feature matrix. ； The stage discrimination variance of each feature among the three centers is calculated using the following formula: ,in, Let j be the mean of this feature at the three centers, j∈[1,3]; Calculate weights The weight matrix W = diag(wf1, wf2, wf3) is obtained. The real-time feature vector Y = [Y(f1), Y(f2), Y(f3)] is collected by BMS, where Y(f1) is the SOH after Kalman filtering correction, Y(f2) is the average capacity decay acceleration of the most recent 100 cycles, and Y(f3) is the average internal resistance growth rate of the day. First, calculate the covariance matrix of the central features of the three stages. Construct a weighted covariance matrix Calculate the weighted Mahalanobis distance between the real-time feature vector Y and each center, using the following formula: We obtain D(Y,C1), D(Y,C2), and D(Y,C3) respectively, where, It is the inverse of the weighted covariance matrix; Select the minimum value among the three distances and take the corresponding center stage as the candidate stage. Calculate the difference ΔDis between the minimum distance and the second smallest distance. If ΔDis > ΔDth, the candidate stage is the final stage. If ΔDis ≤ ΔDth, the stage at the previous moment is maintained. ΔDth is a preset threshold.

Images