Lithium battery energy storage system non-linear capacity recession trajectory prediction method

Abstract

The invention belongs to the technical field of energy storage, and relates to a lithium battery energy storage system nonlinear capacity recession trajectory prediction method, which comprises the following steps: collecting discharge data of a first cycle period and an mth cycle period of a lithium battery, and extracting a voltage-capacity curve of the mth cycle period from the discharge data; the method comprises the following steps: extracting a voltage-capacity curve of an mth cycle period, extracting a capacity increment curve and a capacity difference curve of the mth cycle period on the basis of the voltage-capacity curve, respectively constructing GASF (Gramian Angle Summation Field) graphs on the basis of the voltage-capacity curve, the capacity increment curve and the capacity difference curve, and combining to form a multi-channel format as the input of a prediction model. The method comprises the following steps: firstly, acquiring a degradation track, outputting a degradation track key node on the basis of a prediction model, then integrating the predicted degradation track key node with battery state of health (SOH) information of an mth cycle period, realizing the reconstruction of a capacity degradation track through a three-time Hermite Interpolation Polynomial (PCHIP), realizing data information mining through the collection of early data, and finally realizing the reconstruction of the capacity degradation track. Therefore, the model obtains sufficient degradation information, precise trajectory prediction is realized, the degradation process information of the lithium battery is mastered in the early stage, and time margin is provided for operation and maintenance of an energy storage system.

Description

Technical Field

[0001] The present application belongs to the field of energy storage technology, and more specifically, relates to a method for predicting nonlinear capacity decay trajectory of a lithium battery energy storage system. Background Art

[0002] As an important component of modern power systems, lithium battery energy storage systems (BESS) play a key role in renewable energy grid connection, power dispatching and distributed energy management. With the large-scale grid connection of fluctuating new energy sources such as hydropower, wind power, and solar power, the application demand for BESS is increasing. It can serve as a dispatching bridge to integrate fluctuating new energy sources and stably connect them to the power system, and can provide a variety of auxiliary services.

[0003] In grid-connected application scenarios, BESS not only performs conventional energy storage charging and discharging operations, but also often undertakes auxiliary services such as frequency regulation, load balancing, and peak-valley filling. These auxiliary services have different requirements for the operating conditions of the battery, resulting in irregular battery usage patterns under different load conditions. This frequent change in operating conditions makes the aging process of lithium batteries present complex nonlinear changes, making it difficult to grasp degradation information.

[0004] Existing research shows that the aging paths of lithium batteries vary greatly under different charging and discharging rates, temperatures and other working conditions. In particular, when providing auxiliary services that require rapid response, such as frequency regulation, the battery needs to perform high-frequency charging and discharging operations in a short period of time, which intensifies the electrochemical reaction process inside the battery and leads to rapid loss of active materials. When providing load balancing or peak-valley filling services, the battery works for a long time at a lower rate and exhibits slow capacity decay.

[0005] Due to the irregularity of these ancillary services, the capacity decay process of BESS is complex and difficult to predict, and it is difficult to grasp the health status, which can easily lead to safety hazards. For example, the capacity decay and internal resistance increase of the battery may lead to safety problems such as overheating and overcharging. In severe cases, it may cause fire or explosion, endangering the stability of the entire power system and the safety of personnel. In addition, the uncertainty of battery status will also affect the response service of the power grid. Once the actual performance of the energy storage system is not as expected, it may lead to insufficient response of the ancillary services, further affecting the frequency stability and power quality of the power grid.

[0006] Therefore, accurately predicting the capacity degradation trajectory of lithium batteries and timely maintenance and replacement have become key technical challenges to ensure the stable operation of BESS and the safety of the power grid. If the capacity degradation trajectory can be obtained at an early stage, sufficient time margin can be provided for the operation and maintenance of BESS to ensure the stable operation of the power grid. Summary of the invention

[0007] The present invention provides a method for predicting the nonlinear capacity decay trajectory of a lithium battery energy storage system, which aims to obtain information on the lithium battery degradation process at an early stage and provide a time margin for the operation and maintenance of the system.

[0008] A method for predicting the nonlinear capacity decay trajectory of a lithium battery energy storage system comprises the following steps:

[0009] Step 1: Collect the discharge data of the 1st and mth cycles of the lithium battery, and extract the voltage-capacity curve of the mth cycle;

[0010] Step 2: extracting the capacity increment curve and the capacity difference curve of the mth cycle based on the voltage-capacity curve of the mth cycle;

[0011] Step 3: Based on the extracted voltage-capacity curve, capacity increment curve and capacity difference curve of the mth cycle, a Gramian Angular Summation Field (GASF) graph is constructed respectively. The constructed GASF graph contains capacity decay information. The GASF constructed by the voltage-capacity curve, the capacity increment curve and the capacity difference curve is input into the prediction model in a multi-channel format;

[0012] Step 4: Construct a star-shaped multi-channel dependent neural network (MCDNet) as a prediction model for key nodes of the degradation trajectory, and obtain the key trajectory nodes based on the trained prediction model;

[0013] Step 5: After integrating the key trajectory nodes predicted by the prediction model with the battery health status (SOH) information of the mth cycle, the degradation trajectory is reconstructed using the Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) to obtain the reconstructed capacity degradation trajectory.

[0014] In the present invention, the discharge data of the 1st and mth cycles of the lithium battery are collected, and the voltage-capacity curve of the mth cycle is extracted therefrom, and the capacity increment curve and the capacity difference curve of the mth cycle are extracted based on the voltage-capacity curve, and then the GASF graph is constructed based on the voltage-capacity curve, the capacity increment curve and the capacity difference curve, respectively, and a multi-channel format is formed as the input of the prediction model, and the key nodes of the degradation trajectory are output based on the prediction model, and then the predicted key nodes of the degradation trajectory are integrated with the SOH information of the mth cycle, and the reconstruction of the capacity degradation trajectory is realized through PCHIP. By collecting early data, data information mining is realized, so that the model obtains sufficient degradation information, realizes accurate trajectory prediction, grasps the lithium battery degradation process information at an early stage, and provides time margin for the operation and maintenance of the energy storage system.

[0015] Preferably, the incremental capacity (IC) curve is extracted by differential approximation to obtain the IC value, and the incremental capacity curve is obtained based on the IC values ​​of all sampling points in the mth cycle:

[0016]

[0017] Where: F2 represents the IC value; dQ is the capacity differential value; dV is the voltage differential value; Q t 、V t is the discharge capacity and voltage at the tth sampling point of the voltage-capacity curve data of the mth cycle of the battery; Q t+n 、V t+n is the discharge capacity and voltage of the t+nth sampling point of the voltage-capacity curve data of the mth cycle of the battery; dt is the time difference between the t+nth sampling point and the tth sampling point; is the current average between the t+nth sampling point and the tth sampling point; dV is usually set to be greater than 0.004V.

[0018] Preferably, the steps of extracting the capacity difference curve are as follows:

[0019]

[0020] Where: F3 is the capacity difference; is the voltage-capacity curve of the mth cycle; It represents the voltage-capacity curve of the first cycle; It is the difference in capacity curve between the mth cycle and the 1st cycle.

[0021] Preferably, the specific steps of constructing the GASF graph are as follows:

[0022] Normalization: The voltage-capacity curve, capacity increment curve and capacity difference curve all contain 100 sampling points. For a single curve, the original sequence is x=[x1, x2, ..., x 100 ], the data in the original sequence is mapped to the range [0, 1] by minimization-maximization normalization:

[0023]

[0024] Where: x′ t represents the tth data point of the normalized result sequence x′; x t represents the tth data in the original sequence; min(x) is the minimum value of the original sequence; max(x) is the maximum value of the original sequence.

[0025] The normalized value is mapped to the angle space [0, π / 2] using the inverse cosine function to obtain φ t =arccos(x′ t ), φ t is x′ t The tth value of the angle sequence obtained by arccosine transformation is used to obtain the angle sequence φ = [φ1, φ2, ..., φ n ] Where n = 100, because the number of sampling points of each curve is 100, and then the angle sequence is used to generate the GASF matrix G, where the elements in the matrix are:

[0026] G[i, j] = cos(φ i +φ j );

[0027] Where: G[i, j] represents the value of the i-th row and j-th column of the GASF matrix, which is obtained by the i-th angle sequence value φ i and the jth angle sequence value φ j Add and find the cosine to obtain;

[0028] The GASF matrices corresponding to the voltage-capacity curve, capacity increment curve and capacity difference curve are displayed through heat maps respectively, and then the GASF images constructed by the voltage-capacity curve, capacity increment curve and capacity difference curve are stacked as the input of the prediction model.

[0029] Preferably, the prediction model includes:

[0030] Encoder: It consists of six convolutional blocks and a pooling layer. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a rectified linear unit activation function.

[0031] The features extracted by the encoder are converted into four channels {S1, S2, S3, S4}, each of which is:

[0032] S i =[S i1 , …, S ij ];

[0033] Where: S i represents the feature sequence of the i-th channel, i is between 1 and 4; j represents the dimension of the single channel sequence, the maximum value is 128;

[0034] From S i Extract the core information sequence O with a dimension length of 32:

[0035] O=f(S1,S2,S a , S4);

[0036] Where: f represents the random pooling operation; S i The feature sequence is converted to the same dimension as O through the multi-layer perceptron module. i , i is the i-th channel, the dimension of a single channel is 32, and then each dimension c is normalized by the Softmax activation function, c is the dimension index, the maximum value is 32, p is the selection probability p of each eigenvalue, and the eigenvalue A of the c-th dimension of the i-th channel is calculated ic The probability of selection p ic :

[0037]

[0038] Where: e represents a natural constant; A kc is the eigenvalue of the cth dimension of the kth channel. Since there are only four channels, the range of k is 1 to 4.

[0039] During the training process, for a single dimension c, the core information value O of the cth dimension is extracted from the multinomial distribution based on the feature selection probability p. c :

[0040] O c =A ic where i~P(p 1c , p 2c ,P 3c , P 4c );

[0041] Where: The eigenvalue A of the cth dimension of the i-th channel ic , i is the channel index, the selection range of i is 1 to 4, and the probability of selection is based on p 1c , p 2c , p 3c , p 4c The probability distribution P(p 1c , p 2c , p3c , p 4c ). 1c Refers to the eigenvalue A of the cth dimension of the first channel 1c The probability of being selected, p 2c Refers to the eigenvalue A of the cth dimension of the second channel 2c The probability of being selected, p 3c Refers to the eigenvalue A of the cth dimension of the third channel 3c The probability of being selected, p 4c Refers to the eigenvalue A of the cth dimension of the 4th channel 4c The probability of being selected.

[0042] The initial sequence S i Combined with O to obtain SO i =[S i , O], and then through the multi-layer perceptron module, SO i The feature dimension j+c is converted to S i j dimension, and obtain the fused dependency feature S i ′, realize S i Fusion with O;

[0043] Using residual connection, the original feature S i With the dependency feature S i ′ and decode the fused features into key trajectory nodes through a fully connected layer.

[0044] Preferably, in the process of reconstructing the trajectory using PCHIP, for a given kth SOH value y k Use the kth cycle number C k and the k+1th period number C k+1 When interpolating, C is used k Derivative of point f' k , let the kth period interval h k =C k+1 -C k , then C k The slope d k =(y k+1 -y k ) / h k , in C k The formula for calculating the derivative of a point is:

[0045]

[0046] Where: d k-1 is the number C in the k-1th cycle k-1 The slope at h k-1 is the k-1th period interval, and the coefficients w1 and w2 are equal to 2h respectively. k +hk-1 and h k +2h k-1 .

[0047] The beneficial effects of the present invention include:

[0048] In the present invention, the discharge data of the 1st and mth cycles of the lithium battery are collected, and the voltage-capacity curve of the mth cycle is extracted therefrom, and the capacity increment curve and the capacity difference curve of the mth cycle are extracted based on the voltage-capacity curve, and then the Gramian Angular Summation Field (GASF) graph is constructed based on the voltage-capacity curve, the capacity increment curve and the capacity difference curve, respectively, and a multi-channel format is formed as the input of the prediction model, and the degradation trajectory key nodes are output based on the prediction model, and then the predicted degradation trajectory key nodes are integrated with the battery health state (State Of Health, SOH) information of the mth cycle, and the capacity degradation trajectory is reconstructed by the Piecewise Cubic Hermite Interpolating Polynomial (PCHIP). By collecting early data, data information mining is realized, so that the model obtains sufficient degradation information, realizes accurate trajectory prediction, grasps the lithium battery degradation process information at an early stage, and provides time margin for the operation and maintenance of the energy storage system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 A schematic diagram of the overall structure provided for an embodiment of the present invention.

[0051] Figure 2 The discharge capacity-voltage curves within the first 100 cycles provided by the embodiment of the present invention; wherein Figure (a) is the discharge voltage curve of a short-life battery; and Figure (b) is the discharge voltage curve of a long-life battery.

[0052] Figure 3 A 100×100 pixel feature map is provided for an embodiment of the present invention, wherein Figure (a) represents the F1 graph; Figure (b) represents the F2 graph; Figure (c) represents the F3 graph; and Figure (d) represents the three-channel graph feature obtained by stacking F1, F2 and F3.

[0053] Figure 4 A detailed structural diagram of MCDNet provided in an embodiment of the present invention.

[0054] Figure 5 A schematic diagram of the capacity decay trajectory of the battery in the MIT dataset provided in an embodiment of the present invention.

[0055] Figure 6 A schematic diagram of the degradation trajectory prediction results provided in an embodiment of the present invention, wherein Figure (a) represents the MAE and RMSE of the degradation trajectory reconstruction results; Figure (b) represents the node prediction results; and Figure (c) represents the degradation trajectory reconstruction results of batteries A1, A2, and A3. DETAILED DESCRIPTION

[0056] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0057] The preferred embodiment of the present invention will be further described with reference to the accompanying drawings;

[0058] Step 1: In the early stage of the use of lithium battery energy storage systems (BESS), it is relatively simple and easy to collect voltage and capacity data during the discharge process for feature extraction; Figure 2 The discharge capacity-voltage curve of the lithium battery in the first 100 cycles is shown. As the lithium battery starts to cycle, the discharge capacity gradually decreases, and as the lithium battery ages, the knee of the capacity-voltage curve gradually moves to the left. In addition, the degradation patterns of batteries with different lifespans are also very different. For example, Figure 2 The capacity-voltage curve of the short-life battery with a cycle life of 299 times in (a) changes significantly, but Figure 2 The early capacity change of the long-life battery with a cycle life of 2236 times in (b) is quite subtle; it can be observed that during the early use of the battery, these cycles have certain changes relative to the initial cycle. Therefore, the difference between the mth cycle and the 1st cycle in the initial stage can be combined to mine degradation information and obtain the full life capacity degradation trajectory; this can reduce the data collection requirements and difficulty. In the example of the present invention, only the voltage-capacity curve data of the 100th cycle is used to predict the battery capacity degradation trajectory.

[0059] Step 2: Although the requirements for collecting voltage-capacity data of two cycles in the early use stage of BESS are relatively low, the characteristic information it embodies is also relatively limited. Therefore, it is necessary to mine more degradation information from single-cycle data to accurately predict the capacity degradation trajectory; therefore, the present invention sets the first feature F1 as the voltage-capacity curve of the 100th cycle, and then derives the incremental capacity (IC) curve from F1 as the second feature F2; the IC curve is the derivative of the voltage-capacity curve. As a non-destructive in-situ detection method, it can truly reflect the electrochemical reaction process inside the battery and is widely used in aging mechanism analysis. Therefore, using it as F2 improves the dimension of characteristic information; in practical applications, the IC value is often obtained by differential approximation calculation, and the calculation formula is as follows:

[0060]

[0061] Where: F2 represents the IC value; dQ is the capacity differential value; dV is the voltage differential value; Q t 、V t is the discharge capacity and voltage at the tth sampling point of the voltage-capacity curve data of the mth cycle of the battery; Q t+n 、V t+n is the discharge capacity and voltage of the t+nth sampling point of the voltage-capacity curve data of the mth cycle of the battery; dt is the time difference between the t+nth sampling point and the tth sampling point; is the current average between the t+nth sampling point and the tth sampling point. Therefore, the capacity difference in the discharge process is roughly equivalent to the trapezoidal approximation of the current integral in the time interval between the two sampling points; in order to reduce the numerical differential noise caused by high-frequency sampling, dV is set to be greater than 0.004V.

[0062] Research from the Massachusetts Institute of Technology (MIT) shows that the cycle life is related to the capacity difference curve ΔQ of the 100th cycle relative to the 10th cycle. 100,10 The correlation coefficient between the logarithmic variances is -0.93, which reveals the huge application potential of the capacity difference ΔQ curve. Therefore, the present invention extracts the capacity difference curve from F1 as F3 as follows:

[0063] F3=ΔQ 100,1 (V) = Q 100 (V)-Q1(V).

[0064] Where: Q 100 (V) is the voltage-capacity curve of the 100th cycle, ΔQ 100,1(V) is the capacity difference between the 100th cycle and the first cycle, reflecting the capacity change of the 100th cycle relative to the initial state; represents the voltage-capacity curve of the first cycle; F3 represents the capacity difference; after realizing the mining of feature information, the information obtained is still three curves, which is not conducive to feature learning by deep learning methods. Therefore, the sequence problem is linked to the computer vision problem through the Gramian Angular Summation Field (GASF), so that image recognition technologies such as CNN can be used to process the sequence problem. For details, see step 3.

[0065] Step 3: The three characteristic curves obtained by the present invention are all 100 sampling points (using voltage V as the index sampling point in each cycle, and collecting 100 points in a voltage cycle of 2.0-3.5). For a single curve, the original sequence is x=[x1, x2, ..., x 100 ], and mapped it to the range [0, 1] by min-max normalization:

[0066]

[0067] Where: x′t represents the tth data point of the normalized result sequence x′; x t represents the tth data in the original sequence; min(x) is the minimum value of the original sequence; max(x) is the maximum value of the original sequence.

[0068] The normalized value is mapped to the angle space [0, π / 2] using the inverse cosine function to obtain φ t =arccos(x′ t ), φ t is x′ t The tth value of the angle sequence obtained by arccosine transformation is used to obtain the angle sequence φ = [φ1, φ2, ..., φ n ] Where n = 100, because the number of sampling points of each curve is 100, and then the angle sequence is used to generate the GASF matrix G, where the elements in the matrix are:

[0069] G[i, j] = cos(φ i +φ j );

[0070] Where: G[i, j] represents the value of the i-th row and j-th column of the GASF matrix, which is obtained by the i-th angle sequence value φ i and the jth angle sequence value φ jIt is obtained by adding and taking the cosine, which can represent the similarity between two points in the sequence; the dynamic features of the sequence are converted into geometric structures in the image, which is convenient for analysis with the help of visual technologies such as convolutional neural networks; because GASF retains the timing information and local similarities in the sequence, it can reflect the overall trend and local pattern of the time series in the converted image, which can make subsequent recognition more effective.

[0071] The GASF graph constructed with the F1, F2 and F3 curve data of a battery is as follows: Figure 3 As shown, Figure 3 (a)-(c) represent F1, F2 and F3 respectively, which are all 100×100 pixel images. The constructed GASF image contains some potential capacity decay information. These three images are combined into a multi-channel format as shown in Figure 3 As shown in (d), the input model can provide rich information for the prediction of lithium battery capacity degradation trajectory.

[0072] Step 4: After constructing the graphical features through data collection, construct a multi-channel dependent neural network (MCDNet) such as Figure 4 As a prediction model for the key nodes of the degradation trajectory, it can effectively extract relevant features and form an independent channel sequence; the constructed model adopts an innovative STarAggregate-Redistribute (STAR) module, which can redistribute and fuse the feature information of each channel. This process enhances the interdependence of the information of each channel, thereby facilitating the accurate prediction of the key nodes of the degradation trajectory.

[0073] The encoder of MCDNet mainly consists of six convolution blocks and a pooling layer. Each convolution block consists of a convolution layer, a batch normalization (BN) layer and a rectified linear unit (ReLU) activation function. The present invention uses a convolution layer with a kernel size of 11 in the initial layer of the encoder, which is conducive to the model to extract features between image rows. Since the BN layer and ReLU do not have unique parameter designs, the specific parameter configurations of the convolution layer and the pooling layer are shown in Table 1, where the hyperparameters are kernel size, convolution step size, and padding dimension. The convolution operation gradually reduces the dimension of a single feature map, filtering out irrelevant information while retaining the main features; then, the pooling layer is used to reduce the mixed information and concentrate the information.

[0074] Table 1 Configuration of MCDNet encoder

[0075]

[0076] exist Figure 4 In the example, the features extracted by the encoder are converted into four channels {S1, S2, S3, S4}, each of which is:

[0077] S i =[S i1 , …, S ij ];

[0078] Where: S i represents the feature sequence of the i-th channel, i is between 1 and 4; j represents the dimension of the single channel sequence, the maximum value is 128;

[0079] From S i Extract the core information sequence O with a dimension length of 32:

[0080] O = f(S1, S2, S3, S4);

[0081] Where: f represents the random pooling operation; S i The feature sequence is converted to the same dimension as O through the multi-layer perceptron module. i , i is the i-th channel, the dimension of a single channel is 32, and then each dimension c is normalized by the Softmax activation function, c is the dimension index, the maximum value is 32, p is the selection probability p of each eigenvalue, and the eigenvalue A of the c-th dimension of the i-th channel is calculated ic The probability of selection p ic :

[0082]

[0083] Where: e represents a natural constant; A kc is the eigenvalue of the cth dimension of the kth channel. Since there are only four channels, the range of k is 1 to 4.

[0084] During the retraining process, for a single dimension c, the core information value O of the cth dimension is extracted from the multinomial distribution based on the feature selection probability p. c :

[0085] O c =A ic where i~P(p 1c , p 2c , p 3c , p 4c );

[0086] Where: The eigenvalue A of the cth dimension of the i-th channel ic , i is the channel index, the selection range of i is 1 to 4, and the probability of selection is based on p 1c , p2c , p 3c , p 4c The probability distribution P(p 1c , p 2c , p 3c , p 4c ). 1c Refers to the eigenvalue A of the cth dimension of the first channel 1c The probability of being selected, p 2c Refers to the eigenvalue A of the cth dimension of the second channel 2c The probability of being selected, p 3c Refers to the eigenvalue A of the cth dimension of the third channel 3c The probability of being selected, p 4c Refers to the eigenvalue A of the cth dimension of the 4th channel 4c The probability of being selected.

[0087] The initial sequence S i Combined with O to obtain SO i =[ S i, O], and then through the multi-layer perceptron module, SO i The feature dimension j+c is converted to S i j dimension, and obtain the fused dependency feature S i ′, realize S i Fusion with O;

[0088] Using residual connection, the original feature S i With the dependency feature S i ′, and the fused features are decoded into key trajectory nodes C1, C2, C3 and C4 through a fully connected layer.

[0089] Step 5: The present invention selects the SOH index sequence as [95%, 90%, 85%, 80%] to select the corresponding cycle number to form the key node sequence; this is because the range of the SOH operation is larger and the reconstructed trajectory is closer to the actual situation. By integrating the nodes predicted by MCDNet in step 4 with the known data, that is, the SOH information of the 100th cycle, the degraded trajectory can be reconstructed using the Piecewise Cubic Hermite Interpo1ating Polynomial (PCHIP), such as Figure 1 As shown in Figure 2, PCHIP can ensure the monotonicity of the interpolation points and is consistent with the actual battery degradation trajectory characteristics; during the PCHIP interpolation process, for a given k-th SOH value y k Use the kth cycle number C k and the k+1th period number C k+1 When interpolating, C is used k Derivative of point f'k , let the kth period interval h k =C k+1 -C k , then C k The slope d k =(y k+1 -y k ) / h k , in C k The formula for calculating the derivative of a point is:

[0090]

[0091] Where: d k-1 is the number C in the k-1th cycle k-1 The slope at h k-1 is the k-1th period interval, and the coefficients w1 and w2 are equal to 2h respectively. k +h k -1 and h k +2h k-1 By using PCHIP, the capacity degradation trajectory can be perfectly reconstructed.

[0092] Based on the above technical solution, it is finally possible to collect data information of the initial cycle and early single cycle in the early stage of BESS to obtain capacity degradation trajectory information. This method has low requirements for data collection and only requires the collection of simple voltage-capacity curve data, and the prediction effect of the trajectory is accurate.

[0093] In this example, the MIT data set was selected, which contains 124 APR18650M1A LiFePO4 / C batteries (nominal voltage of 3.3V, nominal capacity of 1.1Ah, voltage range between 3.6V and 2.0V), which were charged in a 30°C thermostat according to a fast charging policy; all batteries were fast charged to 80% state of charge (State Of Charge, SOC) in one or two steps, and then charged to 100% SOC in 1C constant voltage mode. The discharge process uses a 4C current for discharge and stops when it reaches 2V; Figure 5 The capacity decay trajectories of the batteries used are shown. The lifespan of these batteries varies greatly, ranging from a few hundred cycles to more than two thousand cycles, and includes a variety of trajectory degradation modes.

[0094] In this embodiment, when evaluating the prediction accuracy, the evaluation indicators selected include Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). It is worth noting that we sampled the SOH from the 100th cycle to the end of life (EOL) in units of 0.1%, and used the cycle values ​​corresponding to these SOH points to calculate the error:

[0095]

[0096]

[0097] in is the predicted number of cycles, C i is the true value, is the average of the true values.

[0098] Assuming that only the single cycle data of the 100th cycle and the initial cycle data of each battery are available, the present invention uses 50% of the MIT data set as a training set and the other 50% as a test set. The error of the test data set is as follows: Figure 6 (a) shows that only five batteries have a capacity degradation trajectory prediction result with a MAE greater than 100 cycles. In most cases, the trajectory prediction error is small, which shows that the present invention can obtain more accurate results in the capacity degradation trajectory prediction task. Figure 6 As shown in (b), it can be seen from the figure that the proposed model can accurately predict the key trajectory nodes within the battery life range, which verifies the high efficiency of the proposed model for key node prediction tasks. In order to intuitively demonstrate the effect of the capacity decay trajectory obtained by the present invention, Figure 6 (c) shows three batteries (A1, A2, A3) in the test data set, whose lifespans are 470, 1013, and 1433 cycles, respectively. It can be found that the error between the obtained trajectory results and the true results is small, which proves the excellent effect of the trajectory reconstruction method and verifies the accuracy of the present invention in obtaining the capacity degradation trajectory at an early stage.

[0099] Based on the above, it can be seen that the present invention only uses the initial cycle and single cycle data in the early stage of BESS, that is, the IC curve is extracted from the single voltage-capacity curve and the capacity difference curve is obtained in combination with the initial cycle, and these three are constructed as graphic features through the Gram angle and field, thereby realizing data information mining, enabling the model to obtain sufficient degradation information, and then realizing accurate trajectory prediction, solving the problems of the existing methods requiring pre-classification and high data collection requirements.

[0100] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the logic and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting the nonlinear capacity decay trajectory of a lithium battery energy storage system, characterized in that: The following steps are involved: Step 1: Collect the discharge data of the 1st and mth cycles of the lithium battery, and extract the voltage-capacity curve of the mth cycle; Step 2: extracting the capacity increment curve and the capacity difference curve of the mth cycle based on the voltage-capacity curve of the mth cycle; Step 3: Based on the extracted voltage-capacity curve, capacity increment curve and capacity difference curve of the mth cycle, a Gramian Angular Summation Field (GASF) graph is constructed respectively. The constructed GASF graph contains capacity decay information. The GASF graph constructed by the voltage-capacity curve, the capacity increment curve and the capacity difference curve is formed into a multi-channel format and input into the prediction model; Step 4: Construct a multi-channel dependent neural network (MCDNet) of a star module as a prediction model for the key nodes of the degradation trajectory, and obtain the key trajectory nodes based on the trained prediction model; Step 5: After integrating the key trajectory nodes predicted by the prediction model with the battery health status (SOH) information of the mth cycle, the degradation trajectory is reconstructed using the Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) to obtain the reconstructed capacity degradation trajectory.

2. A method for predicting nonlinear capacity decay trajectory of a lithium battery energy storage system according to claim 1, characterized in that: The incremental capacity (IC) curve is extracted by differential approximation to obtain the IC value, and the incremental capacity curve is obtained based on the IC values ​​of all sampling points in the mth cycle: Where: F2 represents the IC value; dQ is the capacity differential value; dV is the voltage differential value; Q t 、V t is the discharge capacity and voltage at the tth sampling point of the voltage-capacity curve data of the mth cycle of the battery; Q t+n 、V t+n is the discharge capacity and voltage of the t+nth sampling point of the voltage-capacity curve data of the mth cycle of the battery; dt is the time difference between the t+nth sampling point and the tth sampling point; is the current average between the t+nth sampling point and the tth sampling point; dV is set to be greater than 0.004V.

3. A method for predicting nonlinear capacity decay trajectory of a lithium battery energy storage system according to claim 1, characterized in that: The steps of extracting the capacity difference curve are as follows: Where: F3 is the capacity difference; is the voltage-capacity curve of the mth cycle; It represents the voltage-capacity curve of the first cycle; It is the difference in capacity curve between the mth cycle and the 1st cycle.

4. A method for predicting nonlinear capacity decay trajectory of a lithium battery energy storage system according to claim 1, characterized in that: The specific steps of constructing the GASF graph are as follows: Normalization: The voltage-capacity curve, capacity increment curve and capacity difference curve all contain 100 sampling points. For a single curve, the original sequence is x=[x1, x2, ..., x 100 ], the data in the original sequence is mapped to the range [0, 1] by minimization-maximization normalization: Where: x′ t represents the tth data point of the normalized result sequence x′; x t represents the tth data in the original sequence; min(x) is the minimum value of the original sequence; max(x) is the maximum value of the original sequence. The normalized value is mapped to the angle space [0, π / 2] using the inverse cosine function to obtain φ t =arccos(x′ t ), φ t is x′ t The tth value of the angle sequence obtained by arccosine transformation is used to obtain the angle sequence φ = [φ1, φ2, ..., φ n ] Where n = 100, because the number of sampling points of each curve is 100, and then the angle sequence is used to generate the GASF matrix G, where the elements in the matrix are: G[i,j]=cos(φ i +φ j ); Where: G[i, j] represents the value of the i-th row and j-th column of the GASF matrix, which is obtained by the i-th angle sequence value φ i and the jth angle sequence value φ j Add and find the cosine to obtain; The GASF matrices corresponding to the voltage-capacity curve, capacity increment curve and capacity difference curve are displayed through heat maps respectively, and then the GASF images constructed by the voltage-capacity curve, capacity increment curve and capacity difference curve are stacked as the input of the prediction model.

5. A method for predicting nonlinear capacity decay trajectory of a lithium battery energy storage system according to claim 1, characterized in that: The prediction model includes: Encoder: It consists of six convolutional blocks and a pooling layer. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a rectified linear unit activation function. The features extracted by the encoder are converted into {S1, S2, S i , S4} four channels, each channel is: S i =[S i1 ,…,S ij ]; Where: S i represents the feature sequence of the i-th channel, i is between 1 and 4; j represents the dimension of the single channel sequence, the maximum value is 128; From each channel feature S i Extract the core information sequence O with a dimension length of 32: O = f(S1, S2, S3, S4); Where: f represents the random pooling operation; S i The feature sequence is converted to the same dimension as O through the multi-layer perceptron module. i , i is the i-th channel, the dimension of a single channel is 32, and then each dimension c is normalized by the Softmax activation function, c is the dimension index, the maximum value is 32, p is the selection probability p of each eigenvalue, and the eigenvalue A of the c-th dimension of the i-th channel is calculated ic The probability of selection p ic : Where: e represents a natural constant; A kc is the eigenvalue of the cth dimension of the kth channel. Since there are only four channels, the range of k is 1 to 4. During the training process, for a single dimension c, the core information value O of the cth dimension is extracted from the multinomial distribution based on the feature selection probability p. c : O c =A ic where i~P(p 1c ,p 2c ,p 3c ,p 4c ); Where: The eigenvalue A of the cth dimension of the i-th channel ic , i is the channel index, the selection range of i is 1 to 4, and the probability of selection is based on p 1c , p 2c , p 3c , p 4c The probability distribution P(p 1c , p 2c , p 3c , p 4c ). 1c Refers to the eigenvalue A of the cth dimension of the first channel 1c The probability of being selected, p 2c Refers to the eigenvalue A of the cth dimension of the second channel 2c The probability of being selected, p 3c Refers to the eigenvalue A of the cth dimension of the third channel 3c The probability of being selected, p 4c Refers to the eigenvalue A of the cth dimension of the 4th channel 4c The probability of being selected. The initial sequence S i Combined with O to obtain SO i =[S i , O], and then through the multi-layer perceptron module, SO i The feature dimension j+c is converted to S i j dimension, and obtain the fused dependency feature S i ′, realize S i Fusion with O; Using residual connection, the original feature S i With the dependency feature S i ′ and decode the fused features into key trajectory nodes through a fully connected layer.

6. A method for predicting nonlinear capacity decay trajectory of a lithium battery energy storage system according to claim 1, characterized in that: In the process of reconstructing the trajectory using PCHIP, for a given k-th SOH value y k Use the kth cycle number C k and the k+1th period number C k+1 When interpolating, C is used k Derivative of point f' k , let the kth period interval h k =C k+1 -C k , then C k The slope d k =(y k+1 -y k ) / h k , in C k The formula for calculating the derivative of a point is: Where: d k-1 is the number C in the k-1th cycle k-1 The slope at h k-1 is the k-1th period interval, and the coefficients w1 and w2 are equal to 2h respectively. k +h k-1 and h k +2h k-1 .

Images