A kind of energy storage power station lithium battery health state prediction system and method
By combining empirical models and data-driven methods, the lithium battery capacity decay sequence is decomposed into overall trends and local fluctuations. Error compensation is performed using bidirectional long short-term memory neural networks and support vector machines, which solves the problems of accuracy and computational requirements in lithium battery SOH prediction in existing technologies, and achieves high-precision lithium battery health status prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FOSHAN UNIVERSITY
- Filing Date
- 2022-10-14
- Publication Date
- 2026-08-04
AI Technical Summary
Existing technologies for predicting the state of health (SOH) of lithium batteries are limited by experimental methods, which struggle to predict risks in real time; model-based methods exhibit poor generalization performance; and data-driven methods require substantial data and computational power, making it impossible to accurately predict SOH under various operating conditions.
A prediction system combining empirical models and data-driven methods is adopted. The system fits the overall trend of capacity decay through a bi-exponential model, uses a bidirectional long short-term memory neural network and a support vector machine for error compensation, and decomposes the capacity decay sequence into the overall trend and local fluctuations for accurate prediction.
It achieves high-precision prediction of SOH of lithium batteries under different operating conditions, reduces computational requirements, improves the accuracy and flexibility of prediction, and has the simplicity of empirical models and the flexibility of data-driven approaches.
Smart Images
Figure CN115621573B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to lithium batteries, and in particular to a system and method for predicting the health status of lithium batteries in energy storage power stations. Background Technology
[0002] With the vigorous development and research of new and renewable energy sources, and the large-scale integration of these intermittent energy sources into the power grid, the demand for grid frequency regulation and peak shaving has increased dramatically. Electrochemical energy storage systems, due to their low deployment environmental requirements and versatility, play a crucial role in addressing new energy consumption, enhancing grid stability, and improving the utilization rate of distribution systems. However, the construction of large-scale electrochemical energy storage power stations also faces many challenges, the most prominent being the safety of lithium-ion batteries. Lithium batteries inevitably undergo an aging process during use. As the number of cycles increases, lithium batteries are susceptible to risks such as overcharging, over-discharging, overheating, and internal / external short circuits, seriously jeopardizing the operational safety of both the lithium batteries and the energy storage power station. Therefore, accurately predicting the state of health (SOH) of the battery and providing safety warnings for the energy storage power station is an effective means to ensure its long-term stable operation.
[0003] Existing technologies typically employ experimental, model-based, or data-driven methods to study and predict battery state-of-the-art (SOH). Experimental methods primarily focus on studying and experimenting with lithium battery aging mechanisms, but struggle to predict the risks of real-time operating energy storage stations. Model-based methods offer computational simplicity but fail to adequately consider environmental conditions and the variations in lithium battery performance under different operating conditions, resulting in poor model generalization. Data-driven methods offer simple modeling, high flexibility, and good real-time performance, but require a large amount of complete operational data on lithium batteries in energy storage stations, along with significant computational power. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a system and method for predicting the state of health (SOH) of lithium batteries in energy storage power stations. This system has the advantages of simple calculation using empirical models and good flexibility driven by data, and can accurately predict the SOH of lithium batteries in energy storage systems.
[0005] The objective of this invention is achieved through the following technical solution: a lithium battery health status prediction system for energy storage power stations, comprising a battery pack, a current sensor, a voltage sensor, a switch control module, a SOH prediction module, and a host computer;
[0006] The switch control module includes a switching switch and a controller. The switching switch is connected in series in the power supply circuit of the battery pack, and the controller is connected to the control terminal of the switching switch to control the switching of the battery pack. The controller is also connected to a host computer to transmit the switching status of the battery pack to the host computer.
[0007] The battery pack includes multiple individual battery cells;
[0008] The current sensor is used to collect the charging and discharging current data of each battery cell and transmit it to the SOH prediction module.
[0009] The voltage sensor is used to collect the charging and discharging current data of each battery cell and transmit it to the SOH prediction module;
[0010] The SOH prediction module is used to predict the SOH of each battery cell based on the received information and transmit the prediction result to the host computer. The SOH refers to the state of health.
[0011] The host computer is used to receive the battery pack's engagement status and the SOH of each individual battery cell predicted by SOH, and to view and save the received signals in real time.
[0012] Preferably, the battery pack contains N battery strings, each battery string contains M battery cells, the M battery cells in each battery string are connected in series, and the battery strings are connected in parallel, where M and N are both positive integers.
[0013] The SOH prediction module is connected to the current sensor via a wired connection.
[0014] The controller, SOH prediction module, and host computer are interconnected in pairs wirelessly.
[0015] The host computer is also connected to an alarm module, which is used to issue an early warning when the SOH drops below 80%.
[0016] Furthermore, the SOH prediction module includes:
[0017] The capacity decay sequence construction unit is used to calculate the capacity and construct the capacity decay sequence based on the battery current and voltage information collected by the current and voltage sensors in the system.
[0018] The empirical mode decomposition unit decomposes the capacity decay sequence to distinguish between the overall trend and local fluctuations of capacity decay;
[0019] Empirical model building unit, used to build a double exponential model of battery capacity degradation;
[0020] The data-driven unit is used to construct a bidirectional long short-term memory neural network and a support vector machine, and to train the bidirectional long short-term memory neural network and support vector machine to obtain a mature bidirectional long short-term memory neural network and support vector machine.
[0021] Bidirectional Long Short-Term Memory Neural Networks analyze forward and reverse sequences through two independent hidden layers, and their output sequence yt The output h of the forward and backward hidden layers is determined by both the forward and backward hidden layers. b h f and output y t The calculation formula is as follows:
[0022]
[0023] Where W, U, and V are the weight matrices of the bidirectional long short-term memory neural network; b and c are the biases; h t The output of the hidden layer at time t; each bidirectional LSTM layer is connected to two fully connected layers;
[0024] Support Vector Machines (SVMs) map input nonlinear data to a high-dimensional feature space, where an optimal linear regression function is constructed to achieve the nonlinear discriminant function of the original space, as shown in the following formula.
[0025]
[0026] Where y is the corresponding sample output, ω T Let be the weight vector of the hyperplane. It is a nonlinear mapping from the input space to the high-dimensional feature space, where b is the bias value.
[0027] The prediction unit uses a mature bidirectional long short-term memory neural network and support vector machine to make predictions, and then adds the predicted values together to calculate the percentage of the sum to the nominal value, which is the final prediction result.
[0028] A method for predicting the health status of lithium batteries in an energy storage power station includes the following steps:
[0029] S1: Collect battery current and voltage information using current and voltage sensors in the system;
[0030] S2: Integrating the current yields the characteristic capacity. The capacity is calculated as the integral of the current over time, i.e. The times corresponding to 3.8V and 3.3V are used as the upper and lower limits of integration, respectively.
[0031] S3: Conduct cyclic charge-discharge experiments to obtain the capacity at each charge-discharge cycle number, forming a capacity decay sequence:
[0032] Each charge-discharge cycle of the battery yields a capacity. The capacity during the battery cycle is saved, and S1-S2 is repeated to obtain data during the battery cycle. The capacity at each charge-discharge cycle number is used to form a capacity decay sequence.
[0033] S4: Construct an empirical model, namely the double-exponential model, using the capacity sequence obtained in S3:
[0034] Q' n =ae bn +ce dn
[0035] When n = 1, 2, 3, ..., N, the capacity of the nth charge-discharge cycle in the capacity decay sequence is taken as Q'. n Substituting these values into the double exponential model, the unknown parameters are identified through least squares fitting, resulting in the identified double exponential model; where N represents the total number of charge-discharge cycles; a, b, c, and d are the unknown parameters of the exponential model, and N is the total number of charge-discharge cycles.
[0036] In the identified double-exponential model, the output Q' of the double-exponential model is determined by the number of charge-discharge cycles n when n = 1, 2, 3, ..., N. n This serves as the fitting result for the double exponential model.
[0037] S5: Use the fitting results of the double exponential model to correct outliers in the battery capacity decay sequence;
[0038] The outlier correction process is as follows: 1. Set an error range and compare the fitting result of the double exponential model with the capacity in the battery degradation sequence; 2. Remove capacity values that exceed the set error range; 3. Replace the outlier values with the average of the left and right adjacent values as the interpolation value.
[0039] S6: Perform empirical mode decomposition on the corrected capacity decay sequence obtained in S5. The capacity decay sequence is further decomposed into the overall trend IMF. 1 and local fluctuations IMF 2 ,imf 3 …imf p When the number of charge-discharge cycles is n, the overall trend obtained from the capacity decay sequence decomposition is denoted as . Local fluctuations are denoted as IMF. n 2 ,imf n 3 …imf n p p is a positive integer;
[0040] S7: Training the BiLSTM Neural Network: Construct a BiLSTM neural network, using the fitting results of the bi-exponential model as features, and the overall trend IMF obtained from the decomposition in S6. 1 The difference e between the fitting result Q' and the bi-exponential model is used as the label to train the neural network: when the number of charge-discharge cycles is n, the neural network output is imf. n 1 Fitting result Q' with the double exponential model n The differences, features, and labels are shown below:
[0041] Q' = [Q' n-1 Q' n-2 …Q' n-m ] T
[0042]
[0043] Where Q' is a feature, e n The label is 'm', where 'm' represents the feature length.
[0044] When n = m+1, m+2, ..., N, construct features Q' and labels e. n The neural network is trained to obtain a trained BiLSTM neural network model;
[0045] S8: When n takes any integer greater than m, use the BiLSTM neural network trained in S7 to predict the overall trend imf. n 1 Fitting result Q' with the double exponential model n The error e between n The fitting result of the bi-exponential model is input into the BiLSTM neural network, and its input is:
[0046] Q' = [Q' n-1 Q' n-2 …Q' n-m ] T
[0047] Q' represents the overall capacity decay trend value fitted by the double exponential model, n-1, n-2… represent the number of battery cycles, and x… n-1 This is the overall trend value fitted to the double exponential model in the (n-1)th iteration, and so on;
[0048] The model output is the error e n Here, the neural network is used to estimate the nth-th fitting result Q' using the double exponential fitting results from n-1 to nm iterations. n With the nth overall trend The error between;
[0049] S9: The BILSTM neural network performs error compensation on the fitting results of the double exponential model to obtain a more accurate overall trend value of capacity decay.
[0050] Q″ n =Q' n +e n
[0051] Q″ n e represents the overall trend forecast for capacity decay. nRepresenting the error compensation value obtained in S7, BiLSTM fits the volume sequence data using a double exponential model before the nth iteration, and outputs the double exponential model fitting result of the nth iteration and the true population trend (imf). 1 The error between the two values is calculated by adding the error value to the fitted value of the double exponential model to obtain the predicted value Q″. n ;
[0052] S10: Training the Support Vector Machine: Input the m local fluctuation data points of capacity decay before the nth battery cycle into the support vector machine. The local fluctuation sequence is obtained by decomposition in step S6. The input sequence is:
[0053]
[0054] X imf Given the input sequence of the SVM, imf i 2 ,imf i 3 …imf i p This is the local fluctuation sequence after empirical mode decomposition for the number of charge-discharge cycles i, where i = n-1, n-2, nm represents the number of battery cycles; p is a positive integer, and the label is the local fluctuation Y of the nth cycle. imf ;
[0055] When n = m+1, m+2, ..., N, use X imf and Y imf The support vector machine is trained to obtain a trained support vector machine;
[0056] S11: When n takes any integer greater than m, use the trained support vector machine to predict local fluctuations: Input the m capacity decay local fluctuation data before the nth battery cycle into the support vector machine, and the local fluctuation sequence IMF... 2 ,imf 3 …imf p Obtained from S6, the input sequence is X. imf The support vector output shows the local fluctuation Y of the nth cycle. imf ;
[0057]
[0058] Y imf =f(X) imf )
[0059] S12: The predicted value of the battery capacity Q is obtained by adding the predicted value of the BISTLM neural network and the predicted value of the support vector machine.
[0060] Q = Q″ + Y imf
[0061] S13: Calculate the SOH based on the capacity prediction value and output it as the prediction result:
[0062]
[0063] Q represents the current characteristic capacity of the battery. N The nominal characteristic capacity is the capacity of the battery between the discharge voltages of [3.3V, 3.8V] during the initial cycle.
[0064] The beneficial effects of this invention are as follows: This invention combines the advantages of simple calculation of empirical models and good flexibility of data-driven approaches, enabling accurate prediction of the state of energy (SOH) of lithium batteries in energy storage systems. It fits the overall capacity decay trend of a battery using a double exponential model, and the differences caused by its operation under different conditions are compensated for by a bidirectional long short-term memory neural network, enabling the prediction of battery SOH under different operating conditions through a single double exponential model. The capacity decay sequence is decomposed into an overall trend and local fluctuations, and then estimated using BiLSTM and SVM respectively. The estimation results are then reconstructed, improving the estimation accuracy. Attached Figure Description
[0065] Figure 1 This is a schematic diagram of the system principle of the present invention;
[0066] Figure 2 This is a schematic diagram illustrating the prediction principle of the SOH prediction module;
[0067] Figure 3 This is a schematic diagram of the BiLSTM structure.
[0068] Figure 4 This is a schematic diagram of the empirical mode decomposition results. Detailed Implementation
[0069] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.
[0070] This invention addresses the State of Health (SOH) prediction of energy storage batteries by proposing a fusion of empirical models and data-driven methods. To overcome the poor performance of empirical models in predicting local fluctuations, the capacity decay sequence is decomposed into an overall trend and local fluctuations. The empirical model is used to fit the overall trend, and a data-driven method is applied for error compensation under different operating conditions, enabling the prediction of battery SOH under various operating conditions using a single empirical model. Local fluctuations are predicted using a different data-driven method. The fitted data from the empirical model can correct historical anomalies and improve data features, preventing outliers from affecting model training and thus improving the accuracy of data-driven prediction results. Furthermore, the fitted data from the empirical model has a high correlation with the battery capacity sequence, ensuring the accuracy of the data-driven model even without training with voltage, current, or temperature data, reducing significant computational costs. This invention combines the advantages of empirical models and data-driven methods, achieving high-precision SOH prediction for batteries under different operating conditions using limited data. Specifically:
[0071] like Figure 1 As shown, a lithium battery health status prediction system for an energy storage power station includes a battery pack, a current sensor, a controller, a SOH prediction module, and a host computer.
[0072] The controller is connected to the positive and negative terminals of the battery pack and is used to control the engagement of the battery pack. The controller is also connected to a host computer and is used to transmit the engagement status of the battery pack to the host computer.
[0073] The battery pack comprises M×N individual battery cells, where M and N are positive integers.
[0074] The current sensor is used to collect the charging and discharging current data of each battery cell and transmit it to the SOH prediction module.
[0075] The SOH prediction module is used to predict the SOH of each battery cell based on the received information and transmit the prediction result to the host computer. The SOH refers to the state of health.
[0076] The host computer is used to receive the battery pack's engagement status and the SOH of each individual battery cell predicted by SOH, and to view and save the received signals in real time.
[0077] In the embodiments of this application, the M×N battery cells are connected in series and parallel in sequence, and the SOH prediction module is connected to the current sensor via a wired connection;
[0078] The controller, SOH prediction module, and host computer are interconnected in pairs wirelessly.
[0079] The controller, SOH prediction module, and host computer are all wirelessly connected to each other. The SOH prediction module can automatically send instructions to switch the battery pack based on the prediction results. In addition to remotely viewing the battery status information, the host computer can also manually switch the battery pack.
[0080] The host computer is also connected to an alarm module, which is used to issue an early warning when the SOH drops below 80%.
[0081] like Figure 2 As shown, the SOH prediction module includes:
[0082] The capacity decay sequence construction unit is used to calculate the capacity and construct the capacity decay sequence based on the battery current and voltage information collected by the current and voltage sensors in the system.
[0083] The Empirical Mode Decomposition (EMD) unit decomposes the capacity decay sequence to distinguish between the overall trend of capacity decay and local fluctuations.
[0084] The empirical model building unit is used to construct a double-exponential model for battery capacity degradation, as shown in the following formula.
[0085]
[0086] Where Q is the battery capacity, n = 1, 2, 3, ... represents the number of charge-discharge cycles of the lithium-ion battery, and α n b n c n d n For the unknown parameters of the exponential model, Kalman filtering is used to identify the parameters and establish a double exponential model;
[0087] The data-driven unit is used to construct a bidirectional long short-term memory neural network and a support vector machine, and to train the bidirectional long short-term memory neural network and support vector machine to obtain a mature bidirectional long short-term memory neural network and support vector machine.
[0088] When constructing a Bidirectional Long Short-Term Memory (BiLSTM) neural network and a Support Vector Machine (SVM), the BiLSTM neural network is as follows: Figure 3 As shown, the forward and reverse sequences are analyzed through two independent hidden layers, and the output sequence y is obtained. t The output h of the forward and backward hidden layers is determined by both the forward and backward hidden layers. b h f and output y tThe calculation formula is as follows:
[0089]
[0090] Where W, U, and V are the weight matrices of the bidirectional long short-term memory neural network; b and c are the biases; h t This represents the output of the hidden layer at time t. Each bidirectional LSTM layer is connected to two fully connected layers.
[0091] Support Vector Machines (SVMs) map input nonlinear data to a high-dimensional feature space, where an optimal linear regression function is constructed to achieve the nonlinear discriminant function of the original space, as shown in the following formula.
[0092]
[0093] Where y is the corresponding sample output, ω T Let be the weight vector of the hyperplane. It is a nonlinear mapping from the input space to the high-dimensional feature space, where b is the bias value.
[0094] The prediction unit uses a mature bidirectional long short-term memory neural network and support vector machine to make predictions, and then adds the predicted values together to calculate the percentage of the sum to the nominal value, which is the final prediction result.
[0095] A method for predicting the health status of lithium batteries in an energy storage power station includes the following steps:
[0096] S1: Collect battery current and voltage information using current and voltage sensors in the system;
[0097] S2: Integrating the current yields the characteristic capacity. The capacity is calculated as the integral of the current over time, i.e. The times corresponding to 3.8V and 3.3V are used as the upper and lower limits of integration, respectively.
[0098] Because the batteries in an energy storage system are not fully charged and discharged during actual operation, the actual battery capacity cannot be directly measured. The capacity here is taken as the discharge amount of the lithium battery between [3.3V, 3.8V]. However, this discharge amount is not the actual capacity of the battery when fully charged and discharged. When the actual battery capacity decays, the discharge amount between [3.3V, 3.8V] will also decay, and its decay is highly correlated with the actual capacity decay. Therefore, the battery's SOH can be calculated based on the discharge amount between [3.3V, 3.8V].
[0099] S3: Conduct cyclic charge-discharge experiments to obtain the capacity at each charge-discharge cycle number, forming a capacity decay sequence:
[0100] Each charge-discharge cycle of the battery yields a capacity value. This capacity is saved during the cycle, and the process of S1-S2 is repeated to acquire the data from the battery cycle. A capacity decay sequence is formed from the capacity values at each charge-discharge cycle. (The SOH prediction module has the function of storing data, saving the capacity decay sequence within the SOH prediction module.)
[0101] S4: Construct an empirical model, namely the double-exponential model, using the capacity sequence obtained in S3:
[0102] Q' n =ae bn +ce dn
[0103] When n = 1, 2, 3, ..., N, the capacity of the nth charge-discharge cycle in the capacity decay sequence is taken as Q'. n Substituting these values into the double exponential model, the unknown parameters are identified through least squares fitting, resulting in the identified double exponential model; where N represents the total number of charge-discharge cycles; a, b, c, and d are the unknown parameters of the exponential model, and N is the total number of charge-discharge cycles.
[0104] In the identified double-exponential model, the output Q' of the double-exponential model is determined by the number of charge-discharge cycles n when n = 1, 2, 3, ..., N. n This serves as the fitting result for the double exponential model.
[0105] The double-exponential model is a polynomial, serving as a mathematical model. After least-squares parameter identification, a, b, c, and d can be determined as constants, while n represents the number of cycles of the lithium-ion battery. The value of n is not fixed. Q' n As the value of n changes, Q' n This is the fitting result of the double exponential model, which is the value of a certain cycle in the overall trend of capacity decay.
[0106] S5: Use the fitting results of the double exponential model to correct outliers in the battery capacity decay sequence;
[0107] The outlier correction process is as follows: 1. Set an error range and compare the fitting result of the double exponential model with the capacity in the battery degradation sequence; 2. Remove capacity values that exceed the set error range; 3. Replace the outlier values with the average of the left and right adjacent values as the interpolation value.
[0108] S6: Perform empirical mode decomposition on the corrected capacity decay sequence obtained in S5. The capacity decay sequence is further decomposed into the overall trend IMF. 1 and local fluctuations IMF 2 ,imf 3 …imf pWhen the number of charge-discharge cycles is n, the overall trend obtained from the capacity decay sequence decomposition is denoted as . Local fluctuations are denoted as IMF. n 2 ,imf n 3 …imf n p p is a positive integer;
[0109] Empirical mode decomposition (EMD) can be applied to capacity decay sequences to separate them into overall trends and local fluctuations, such as... Figure 4 As shown, the capacity decay curve is ultimately decomposed into one overall trend curve and three local fluctuation curves. Therefore, each battery cycle corresponds to an overall capacity decay trend value x. n ;
[0110] Each charge and discharge cycle of the battery pack constitutes one cycle. The number of cycles gradually increases during operation. This does not mean that the battery needs to be cycled for every SOH prediction; it only requires recording the data from normal battery pack operation. The overall trend can be quantified as follows: Figure 4 As shown in curve 2, curve 1 represents the capacity decay sequence. Decomposing curve 1 using empirical modes yields several curves, where the residuals represent the overall trend. (Empirical mode decomposition yields several curves, one of which is the residual, such as...) Figure 4 As shown in curve 2, the residual is typically a smooth curve; the rest are modal function curves, such as curves 3, 4, and 5, which usually exhibit significant fluctuations. The residual and modal function are expressions of empirical mode decomposition. This patent represents the decomposed curves as the overall trend and local fluctuations of capacity decay.
[0111] S7: Training the BiLSTM Neural Network: Construct a BiLSTM neural network, using the fitting results of the bi-exponential model as features, and the overall trend IMF obtained from the decomposition in S6. 1 The difference e between the fitting result Q' and the bi-exponential model is used as a label to train the neural network: when the number of charge-discharge cycles is n, the neural network output is... Fitting result Q' with the double exponential model n The differences, features, and labels are shown below:
[0112] Q' = [Q' n-1 Q' n-2 ...Q' n-m ] T
[0113]
[0114] Where Q' is a feature, e nThe label is 'm', where 'm' represents the feature length.
[0115] When n = m+1, m+2, ..., N, construct features Q' and labels e. n The neural network is trained to obtain a trained BiLSTM neural network model;
[0116] S8: When n takes any integer greater than m, use the BiLSTM neural network trained in S7 to predict the overall trend. Fitting result Q' with the double exponential model n The error e between n The fitting result of the bi-exponential model is input into the BiLSTM neural network, and its input is:
[0117] Q' = [Q' n-1 Q' n-2 …Q' n-m ] T
[0118] Q' represents the overall capacity decay trend value fitted by the double exponential model, n-1, n-2… represent the number of battery cycles, and x… n-1 This is the overall trend value fitted to the double exponential model in the (n-1)th iteration, and so on;
[0119] The model output is the error e n Here, the neural network is used to estimate the nth-th fitting result Q' using the double exponential fitting results from n-1 to nm iterations. n With the nth overall trend The error between;
[0120] S9: The BILSTM neural network performs error compensation on the fitting results of the double exponential model to obtain a more accurate overall trend value of capacity decay.
[0121] Q″ n =Q' n +e n
[0122] Q″ n e represents the overall trend forecast for capacity decay. n Representing the error compensation value obtained in S7, BiLSTM fits the volume sequence data using a double exponential model before the nth iteration, and outputs the double exponential model fitting result of the nth iteration and the true population trend (imf). 1 The error between the two values is calculated by adding the error value to the fitted value of the double exponential model to obtain the predicted value Q″. n ;
[0123] The above is the overall trend (IMF). 1 The following is an estimate, showing local fluctuations:
[0124] S10: Training the Support Vector Machine: Input the m local fluctuation data points of capacity decay before the nth battery cycle into the support vector machine. The local fluctuation sequence is obtained by decomposition in step S6. The input sequence is:
[0125]
[0126] X imf Given the input sequence of the SVM, imf i 2 ,imf i 3 …imf i p This is the local fluctuation sequence after empirical mode decomposition for the number of charge-discharge cycles i, where i = n-1, n-2, nm represents the number of battery cycles; p is a positive integer, and the label is the local fluctuation Y of the nth cycle. imf ;
[0127] When n = m+1, m+2, ..., N, use X imf and Y imf The support vector machine is trained to obtain a trained support vector machine;
[0128] S11: When n takes any integer greater than m, use the trained support vector machine to predict local fluctuations: Input the m capacity decay local fluctuation data before the nth battery cycle into the support vector machine, and the local fluctuation sequence IMF... 2 ,imf 3 …imf p Obtained from S6, the input sequence is X. imf The support vector output shows the local fluctuation Y of the nth cycle. imf ;
[0129]
[0130] Y imf =f(X) imf )
[0131] S12: The predicted value of the battery capacity Q is obtained by adding the predicted value of the BISTLM neural network and the predicted value of the support vector machine.
[0132] Q = Q″ + Y imf
[0133] S13: Calculate the SOH based on the capacity prediction value and output it as the prediction result:
[0134]
[0135] Q represents the current characteristic capacity of the battery. NThe nominal characteristic capacity is the capacity of the battery between the discharge voltages of [3.3V, 3.8V] during the initial cycle.
[0136] The above description represents preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technical or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A system for predicting the state of health of a lithium battery in an energy storage plant, the system comprising: It includes a battery pack, current sensor, voltage sensor, switch control module, SOH prediction module, and host computer; The switch control module includes a switching switch and a controller. The switching switch is connected in series in the power supply circuit of the battery pack, and the controller is connected to the control terminal of the switching switch to control the switching of the battery pack. The controller is also connected to a host computer to transmit the switching status of the battery pack to the host computer. The battery pack includes multiple individual battery cells; The current sensor is used to collect the charging and discharging current data of each battery cell and transmit it to the SOH prediction module. The voltage sensor is used to collect the charge and discharge voltage data of each battery cell and transmit it to the SOH prediction module. The SOH prediction module is used to predict the SOH of each battery cell based on the received information and transmit the prediction result to the host computer. The SOH refers to the state of health. The host computer is used to receive the battery pack's engagement status and the SOH of each individual battery cell predicted by SOH, and to view and save the received signals in real time. The prediction process of the system includes: S1: Collect battery current and voltage information using current and voltage sensors in the system; S2: the current is integrated to obtain the characteristic capacity, the calculation formula of the capacity is the integral of current to time, that is , 3.8V, 3.3V corresponding time as the upper and lower limit of integration; S3: Conduct cyclic charge-discharge experiments to obtain the capacity at each charge-discharge cycle number, forming a capacity decay sequence: Each charge-discharge cycle of the battery yields a capacity. The capacity during the battery cycle is saved, and S1-S2 is repeated to obtain data during the battery cycle. The capacity at each charge-discharge cycle number is used to form a capacity decay sequence. S4: Construct an empirical model, namely the double-exponential model, using the capacity sequence obtained in S3: ； When n = 1, 2, 3, ……N, the capacity of the n-th charge and discharge cycle in the capacity attenuation sequence is taken as , substituted into the double exponential model, the unknown parameters are identified by least square fitting, and the identified double exponential model is obtained; wherein N represents the total number of charge and discharge; a, b, c, d are unknown parameters of the exponential model, and N is the total number of cycle charge and discharge experiments. In the identified double-exponential model, the output of the double-exponential model is determined by the number of charge-discharge cycles n when n=1,2,3,...N. This serves as the fitting result for the double exponential model. S5: Use the fitting results of the double exponential model to correct outliers in the battery capacity decay sequence; The outlier correction process is as follows:
1. Set an error range and compare the fitting result of the double exponential model with the capacity in the battery degradation sequence; 2. Remove capacity values that exceed the set error range; 3. Replace the outlier values with the average of the left and right adjacent values as the interpolation value. S6: Perform empirical mode decomposition on the corrected capacity decay sequence obtained in S5. The capacity decay sequence is further decomposed into the overall trend. and local fluctuations When the number of charge-discharge cycles is n, the overall trend obtained from the capacity decay sequence decomposition is denoted as . Local fluctuations are denoted as p is a positive integer; S7: Training the BiLSTM Neural Network: Construct a BiLSTM neural network, using the fitting results of the bi-exponential model as features, and the overall trend obtained from the decomposition in S6. Fitting results with the double exponential model The difference As labels, the neural network is trained: when the number of charge-discharge cycles is n, the neural network output is... Fitting results with the double exponential model The differences, features, and labels are shown below: ； in, As a feature, The label is 'm', where 'm' represents the feature length. When n = m+1, m+2, ..., N, construct features and tags The neural network is trained to obtain a trained BiLSTM neural network model; S8: When n takes any integer greater than m, use the BiLSTM neural network trained in S7 to predict the overall trend. Fitting results with the double exponential model Error between The fitting result of the double exponential model is input into the BiLSTM neural network, and its input is: ； This represents the overall trend value of capacity decay fitted by the bi-exponential model. …represents the number of battery cycles. This is the overall trend value fitted to the double exponential model in the (n-1)th iteration, and so on; The model output is the error. Here, the neural network is used to estimate the nth fitting result from the double exponential fitting results of n-1 to nm iterations. With the nth overall trend The error between; S9: The BILSTM neural network performs error compensation on the fitting results of the double exponential model to obtain a more accurate overall trend value of capacity decay. ； This represents the predicted overall trend of capacity decay. Representing the error compensation value obtained in S7, BiLSTM fits the volume sequence data using a double exponential model before the nth iteration, and outputs the fitting result of the double exponential model in the nth iteration and the true population trend. The error between the two values is used to obtain the predicted value by adding the error value to the fitted value of the double exponential model. ; S10: Training the Support Vector Machine: Input the m local fluctuation data points of capacity decay before the nth battery cycle into the support vector machine. The local fluctuation sequence is obtained by decomposition in step S6. The input sequence is: ； The input sequence for SVM, This is the local fluctuation sequence after empirical mode decomposition for a charge-discharge cycle number i. Represents the number of battery cycles; p is a positive integer, and the label is the local fluctuation Y of the nth cycle. imf ; When n = m+1, m+2, ..., N, use and Y imf The support vector machine is trained to obtain a trained support vector machine; S11: When n takes any integer greater than m, use the trained support vector machine to predict local fluctuations: Input the m capacity decay local fluctuation data before the nth battery cycle into the support vector machine, and the local fluctuation sequence... Obtained from S6, the input sequence is X. imf Support vector output shows the local fluctuations in the nth cycle. ; ； ； S12: The predicted value of the battery capacity is obtained by adding the predicted value of the BISTLM neural network and the predicted value of the support vector machine. ; ； S13: Calculate the SOH based on the capacity prediction value and output it as the prediction result: ； Q is the current battery characteristic capacity, Q N is the nominal battery characteristic capacity, i.e. the capacity between the discharge voltages [3.3 V, 3.8 V] at the initial cycle of the battery.
2. The lithium battery health status prediction system for an energy storage power station according to claim 1, characterized in that: The battery pack contains N battery strings, each battery string contains M battery cells, the M battery cells in each battery string are connected in series, and the battery strings are connected in parallel, where M and N are both positive integers.
3. The lithium battery health status prediction system for energy storage power stations according to claim 1, characterized in that: The SOH prediction module is connected to the current sensor via a wired connection. The controller, SOH prediction module, and host computer are interconnected in pairs wirelessly.
4. The lithium battery health status prediction system for an energy storage power station according to claim 1, characterized in that: The host computer is also connected to an alarm module, which is used to issue an early warning when the SOH drops below 80%.
5. The lithium battery health status prediction system for an energy storage power station according to claim 1, characterized in that: The SOH prediction module includes: The capacity decay sequence construction unit is used to calculate the capacity and construct the capacity decay sequence based on the battery current and voltage information collected by the current and voltage sensors in the system. The empirical mode decomposition unit decomposes the capacity decay sequence to distinguish between the overall trend and local fluctuations of capacity decay; Empirical model building unit, used to build a double exponential model of battery capacity degradation; The data-driven unit is used to construct a bidirectional long short-term memory neural network and a support vector machine, and to train the bidirectional long short-term memory neural network and support vector machine to obtain a mature bidirectional long short-term memory neural network and support vector machine. The prediction unit uses a mature bidirectional long short-term memory neural network and support vector machine to make predictions, and then adds the predicted values together to calculate the percentage of the sum to the nominal value, which is the final prediction result.