Energy storage system power distribution method based on battery health state prediction

Abstract

The invention provides an energy storage system power distribution method based on battery pack health state prediction, and the method comprises the steps: constructing a hybrid neural network model for predicting the health state of a lithium ion battery through combining a convolutional neural network and a long short-term memory network, and optimizing the parameters of the hybrid neural network model through employing a Bayesian algorithm. The battery pack health state prediction precision of the model is improved; in combination with the provided lithium ion battery prediction model, a battery pack charging / discharging priority ranking rule based on an entropy weight method is provided; a battery pack power optimization distribution strategy is established by taking the consistency of the state of charge of the battery pack as an objective function and taking the state of charge and the state of health of the battery pack as evaluation indexes, so that energy optimization distribution of the energy storage system is realized, the service life of an old battery is prolonged, overcharge and overdischarge of a single battery are avoided, and balance of the state of charge of the battery pack is realized.

Description

Technical Field

[0001] The present application relates to the field of energy storage optimization scheduling, and in particular to a power allocation method for an energy storage system based on battery health status prediction. Background Art

[0002] Power systems are gradually moving toward a high proportion of renewable energy integration and the use of power electronic equipment. Given the uncertainty and randomness of wind and solar power output, utilizing energy storage devices is an effective means of addressing the high uncertainty and peak-shaving characteristics of renewable energy output. Battery energy storage systems (BESS) operate under complex conditions and the lifespan degradation of each battery pack is inconsistent, making BESS health management challenging. Therefore, it is necessary to study BESS health status prediction and charge / discharge power allocation strategies to extend the lifespan of energy storage and improve its economic efficiency.

[0003] Currently, most research on BESS energy management is based on the battery pack's state of charge (SOH) or the overall state of health (SOH) of the BESS. However, few energy management methods consider the real-time SOC and SOH changes of each battery pack in the BESS. This raises two issues that require further research: First, it is difficult to detect the SOH of the energy storage battery in real time while the energy storage system is operating. Second, the SOC and SOH of the energy storage battery must be considered when managing the energy storage system to reduce the rate of battery loss. Summary of the Invention

[0004] This application provides a power distribution method for an energy storage system based on battery health status prediction to further improve the existing technology to effectively distribute loads, reduce excessive loads on individual batteries, balance battery usage, and extend the service life of battery packs.

[0005] Specifically, the present application provides a method for allocating power to an energy storage system based on battery health status prediction, the method comprising:

[0006] A hybrid neural network model based on a convolutional neural network and a long short-term memory network uses a data set obtained from lithium battery tests. The battery's charge / discharge voltage, current, cycle number, and battery level are input as feature quantities into the model training to obtain a lithium battery health status prediction model. The Bayesian optimization algorithm is then used to improve the accuracy of the hybrid neural network model.

[0007] Based on the established lithium battery health status prediction model, the health status of each battery pack is predicted in real time. Based on the entropy weight method, the battery pack's charge and health status are used as evaluation indicators to sort the battery pack's charge / discharge priority, select the appropriate battery pack, and determine the optimal power allocation for each battery pack with SOC consistency as the objective function;

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] In a first aspect, the present invention provides a hybrid neural network model based on a convolutional neural network and a long short-term memory network to predict the health status of a lithium battery, comprising:

[0010] Collecting power parameters from battery operation, the power parameters including at least SOC, SOH, voltage, and current data;

[0011] The power parameters are normalized and the calculation formula is:

[0012]

[0013] Where X is the input data after normalization; x is the original input data; x max and x min Represent the maximum and minimum values ​​respectively;

[0014] Input multi-dimensional time series data X: power parameters;

[0015] Perform one-dimensional convolution on the input multi-dimensional time series data X. The calculation formula is:

[0016] Z t =f(W conv *X t:t+k-1 +b conv );

[0017] Where Z t Represents the output features after convolution, extracting the local features of time step t, W conv Represents the convolution kernel, the size is k*d, where k is the size of the convolution kernel and represents the length of the convolution window; d is the feature dimension of the input data; X t:t+k-1 represents the local time segment starting from time step t, that is, the input data subsequence from t to t+k-1; b conv represents the bias term of the convolution layer, which is used to adjust the convolution output; f(·) is the activation function, which introduces nonlinearity to enhance the expressive power of the model.

[0018] After the convolution operation, in order to reduce the dimension of the convolution data while retaining the main features, maximum pooling is used for downsampling. The formula is:

[0019] Pt =max(Z t:t+p-1 );

[0020] P t is the output feature of time step t after pooling, indicating the result of the pooling operation; Z t:t+p-1 represents the convolution feature subsequence starting at time step t, with a length of p; max(·) is the maximum pooling operation, which selects the maximum value within a given time window.

[0021] The output features of the convolutional neural network are flattened and then input into the long short-term memory network;

[0022] Read the current input information and the unit state information at the previous moment, pass the information to the Sigmoid function at the same time, and decide whether to discard or retain the information based on the output value. The calculation formula is:

[0023] f t =σ(W f [x t ,h t-1 ]+b f );

[0024] Where σ(·) is the sigmoid function with an output interval of (0,1), W f is the weight matrix of the forget gate structure; b f is the bias matrix of the forget gate structure; x t Represents the input of the network at the current moment; h t-1 Represents the output of the network at the previous moment;

[0025] Calculate the input gate i through the sigmoid function t The output value of determines which information is important. The calculation formula is:

[0026] i t =σ(W i [x t ,h t-1 ]+b i );

[0027] Where W i is the weight matrix of the input gate structure;

[0028] Using c t 'Indicates that the unit state input at the current moment is the unit state at the previous moment. The tanh function is used to filter the extracted valid information. The calculation formula is;

[0029] c t '=tanh(W c [x t ,h t-1 ]+bc );

[0030] Where b i is the bias matrix of the input gate structure; W c is the weight matrix of the unit; b c is the bias matrix of the unit.

[0031] Using c t Represents the unit state at the current moment, and the calculation formula is;

[0032] c t =f t ⊙c t-1 +i t ⊙c t ';

[0033] Where ⊙ is the product of vector elements, c t It consists of two parts: the first part is based on the forget gate f t and the cell state c at the previous moment t-1 Multiplication, the second part is composed of input gate i t with c t 'Multiply to form. In this way, the current memory c t ' and long-term memory c t-1 Combined. ;

[0034] The output of the network at the previous moment and the current input are passed to the sigmoid function, and the unit state at the previous moment is passed to the tanh function; then the output of tanh is multiplied by the output of sigmoid to determine the output h of the network at the next moment t The information carried will eventually pass the new unit state and new output to the next time step. The calculation formula is:

[0035] o t =σ(W o [x t ,h t-1 ]+b o );

[0036] Where W o is the weight matrix of the output gate structure; b o is the bias matrix of the output gate structure;

[0037] Multiply the output of tanh by the output of sigmoid to determine the output h of the network at the next moment t The information carried will eventually pass the new unit state and new output to the next time step. The calculation formula is:

[0038] h t =o t⊙tanh(c t );

[0039] Output prediction value after passing through the fully connected layer The calculation formula is;

[0040]

[0041] Where W fc The weight matrix of the fully connected layer; b fc is the bias of the fully connected layer;

[0042] The training set is used to optimize the BO hyperparameters. The calculation formula is:

[0043] z * =argmaxf(z),z∈Z;

[0044] Where Z is the sample space; z is a randomly selected sample point in the sample space; z * is a globally optimal solution including real, integer, or categorical eigenvalues;

[0045] The root mean square (RMES) and determination coefficient (R) of the predicted and true values 2 The two indicators are used to verify the prediction effect, and the calculation formula is:

[0046]

[0047] Among them, the predicted value y'={y'1,y'2,...,y' n}, the true value y={y1,y2,...,y n}, n represents the number of predicted data.

[0048] In a second aspect, the present invention provides a method for allocating power to an energy storage system based on an entropy weight method, comprising:

[0049] The energy storage system receives power dispatch instructions;

[0050] The corresponding number of battery pack actions is calculated using the following formula:

[0051]

[0052] Where: N t P is the number of BESS battery packs charged / discharged at time t; t is the power that needs to be dispatched by BESS at time t; is the rated power of the battery pack; λ is the redundancy factor; [] indicates rounding up.

[0053] Evaluation indicators are established for multiple samples and multiple indicators, and the calculation formula is:

[0054]

[0055] Where: Q xy is the value of the xth evaluated object under the yth indicator, where y∈[1,b] means the evaluation method has b evaluation indicators; x∈[1,a] means there are a samples;

[0056] The data were normalized and standardized, and the calculation formula was:

[0057]

[0058] Where: Q' xy1 is a positive indicator, Q' xy2 It is a negative indicator; max(Q y )、min(Q y ) are the maximum and minimum values ​​of all objects under index y; Q' xy is the index value after standardization, and its value is 0≤Q' xy ≤1;

[0059] Calculate the entropy value of each evaluation indicator using the following formula:

[0060]

[0061] Where: S y represents information entropy, S y The larger the value of , the greater the amount of information contained in the data;

[0062] Calculate the weight of each evaluation indicator, the calculation formula is;

[0063]

[0064] A power optimization allocation model with SOC consistency as the objective function is established, and the calculation formula is:

[0065]

[0066] Where: SOC i,t is the SOC of battery energy storage unit i at the end of period t; SOC R is the reference value of SOC;

[0067] In the energy storage system, during time period t, the total charge / discharge power of each battery group must be equal to the dispatch instruction. The constraint is as follows:

[0068]

[0069] Where: is the discharge power of battery pack i in period t; is the charging power of battery pack i in period t;

[0070] Charge / discharge power constraint, the constraint formula is as follows;

[0071]

[0072] Where: and The maximum and minimum values ​​of the battery pack discharge power; and The maximum and minimum charging power for the battery pack;

[0073] The change constraint of battery pack SOC is as follows:

[0074]

[0075] Where: SOC i,t is the SOC of battery pack i at the end of period t; S N Battery rated capacity; They are the discharge and charge amounts of battery pack i in period t; n is the number of samples per hour; are the maximum and minimum allowed SOC of battery pack i respectively;

[0076] Compared with the prior art, the present invention has the following technical effects:

[0077] The power allocation method for energy storage systems based on battery health status prediction provided by the present invention uses the number of battery cycles obtained by the rain flow counting method as a feature quantity, which solves the problem of invalid data and unclear feature relationships in the data set. In addition, a hybrid neural network model combining a recurrent neural network and a convolutional neural network is proposed, which improves the model's ability to capture the correlation between features and improves the prediction accuracy. In response to the problem that hyperparameters in neural networks are difficult to determine, a Bayesian optimization algorithm is introduced to optimize the hyperparameters of the hybrid neural network model, ensuring the optimization of the overall prediction accuracy of the model. The health status of the battery pack is predicted in real time during its operation, and the SOH and SOC of each battery pack are combined to participate in the energy management of the BESS, so as to achieve effective load distribution, reduce the excessive load of a single battery, balance battery use, and extend the service life of the battery pack. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is a flow chart of the energy storage system power allocation method based on battery pack health status prediction described in this application.

[0079] Figure 2 A flow chart is formed for the battery health status prediction model described in this application.

[0080] Figure 3This is a comparison chart of the effects of the battery pack health status prediction model test set B5 described in this application.

[0081] Figure 4 This is a comparison chart of the effects of the battery pack health status prediction model test set B7 described in this application.

[0082] Figure 5 This is a flow chart of the battery pack scheduling strategy based on real-time battery pack health prediction using the entropy weight method described in this application.

[0083] Figure 6 This is a battery pack power distribution charge state trend diagram based on real-time battery pack health prediction using entropy weight method as described in this application.

[0084] Figure 7 This is a comparison chart of the battery pack degradation effect of the energy storage system power allocation method based on battery pack health status prediction described in this application.

[0085] Figure 8 This is the battery pack attenuation with different initial health status values ​​described in this application. DETAILED DESCRIPTION

[0086] The following is a further detailed description of a power allocation method for an energy storage system based on battery pack health status prediction of the present application in conjunction with specific embodiments and drawings.

[0087] See Figure 1 The present application provides a power allocation method for an energy storage system based on the prediction of the health status of a battery pack. The method collects cyclic charge and discharge experimental data of lithium-ion batteries of different models, calculates the equivalent number of cycles at each time point according to the change curve of the state of charge (SOC) thereof by using the rain flow counting method, and inputs the equivalent number of cycles, health status, voltage and current values ​​as characteristic values ​​into a hybrid neural network model of a convolutional neural network and a long short-term memory network to obtain a lithium battery health status prediction model, and optimizes the parameters of the hybrid neural network model using a Bayesian optimization algorithm.

[0088] See Figure 2 Based on the established lithium battery health status prediction model, the health status of each battery pack is predicted in real time. Based on the entropy weight method, the charge state and health state of the battery pack are used as evaluation indicators to sort the charge / discharge priority of the battery pack, select the appropriate battery pack, and determine the optimal power allocation of each battery pack with SOC consistency as the objective function.

[0089] The lithium-ion battery health status prediction model based on the hybrid neural network model of the present invention includes:

[0090] Collecting power parameters from battery operation, the power parameters including at least SOC, SOH, voltage, and current data;

[0091] The power parameters are normalized and the calculation formula is:

[0092]

[0093] Where X is the input data after normalization; x is the original input data; x max and x min Represent the maximum and minimum values ​​respectively;

[0094] Input multi-dimensional time series data X: power parameters;

[0095] Perform one-dimensional convolution on the input multi-dimensional time series data X. The calculation formula is:

[0096] Z t =f(W conv *X t:t+k-1 +b conv );

[0097] Where Z t Represents the output features after convolution, extracting the local features of time step t, W conv Represents the convolution kernel, the size is k*d, where k is the size of the convolution kernel and represents the length of the convolution window; d is the feature dimension of the input data; X t:t+k-1 represents the local time segment starting from time step t, that is, the input data subsequence from t to t+k-1; b conv represents the bias term of the convolution layer, which is used to adjust the convolution output; f(·) is the activation function, which introduces nonlinearity to enhance the expressive power of the model.

[0098] After the convolution operation, in order to reduce the dimension of the convolution data while retaining the main features, maximum pooling is used for downsampling. The formula is:

[0099] P t =max(Z t:t+p-1 );

[0100] P t is the output feature of time step t after pooling, indicating the result of the pooling operation; Z t:t+p-1 represents the convolution feature subsequence starting at time step t, with a length of p; max(·) is the maximum pooling operation, which selects the maximum value within a given time window.

[0101] The output features of the convolutional neural network are flattened and then input into the long short-term memory network;

[0102] Read the current input information and the unit state information at the previous moment, pass the information to the Sigmoid function at the same time, and decide whether to discard or retain the information based on the output value. The calculation formula is:

[0103] f t =σ(W f [x t ,h t-1 ]+b f );

[0104] Where σ(·) is the sigmoid function with an output interval of (0,1), W f is the weight matrix of the forget gate structure; b f is the bias matrix of the forget gate structure; x t Represents the input of the network at the current moment; h t-1 Represents the output of the network at the previous moment;

[0105] Calculate the input gate i through the sigmoid function t The output value of determines which information is important. The calculation formula is:

[0106] i t =σ(W i [x t ,h t-1 ]+b i );

[0107] Where W i is the weight matrix of the input gate structure;

[0108] Using c′ t The cell state input at the current moment is the cell state at the previous moment. The tanh function is used to filter the extracted valid information. The calculation formula is:

[0109] c t '=tanh(W c [x t ,h t-1 ]+b c );

[0110] Where b i is the bias matrix of the input gate structure; W c is the weight matrix of the unit; b c is the bias matrix of the unit.

[0111] Using c t Represents the unit state at the current moment, and the calculation formula is;

[0112] c t =f t ⊙c t-1 +i t ⊙c t ';

[0113] Where ⊙ is the product of vector elements, c t It consists of two parts: the first part is based on the forget gate f t and the cell state c at the previous moment t-1 Multiplication, the second part is composed of input gate i t and c′ t In this way, the current memory c′ t and long-term memory c t-1 Combined.

[0114] The output of the network at the previous moment and the current input are passed to the sigmoid function, and the unit state at the previous moment is passed to the tanh function; then the output of tanh is multiplied by the output of sigmoid to determine the output h of the network at the next moment t The information carried will eventually pass the new unit state and new output to the next time step. The calculation formula is:

[0115] o t =σ(W o [x t ,h t-1 ]+b o );

[0116] Where W o is the weight matrix of the output gate structure; b o is the bias matrix of the output gate structure;

[0117] Multiply the output of tanh by the output of sigmoid to determine the output h of the network at the next moment t The information carried will eventually pass the new unit state and new output to the next time step. The calculation formula is:

[0118] h t =o t ⊙tanh(c t );

[0119] Output prediction value after passing through the fully connected layer The calculation formula is;

[0120]

[0121] Where W fc The weight matrix of the fully connected layer; b fc is the bias of the fully connected layer;

[0122] The training set is used to optimize the BO hyperparameters. The calculation formula is:

[0123] z * =argmaxf(z),z∈Z;

[0124] Where Z is the sample space; z is a randomly selected sample point in the sample space; z * is a globally optimal solution including real, integer, or categorical eigenvalues;

[0125] The root mean square (RMES) and determination coefficient (R) of the predicted and true values 2 The two indicators are used to verify the prediction effect, and the calculation formula is:

[0126]

[0127] Among them, the predicted value y'={y'1,y'2,...,y' n}, the true value y={y1,y2,...,y n}, n represents the number of predicted data.

[0128] In a second aspect, the present invention provides a method for allocating power to an energy storage system based on an entropy weight method, comprising:

[0129] The energy storage system receives power dispatch instructions;

[0130] The corresponding number of battery pack actions is calculated using the following formula:

[0131]

[0132] Where: N t P is the number of BESS battery packs charged / discharged at time t; t is the power that needs to be dispatched by BESS at time t; is the rated power of the battery pack; λ is the redundancy factor; [] indicates rounding up;

[0133] Evaluation indicators are established for multiple samples and multiple indicators, and the calculation formula is:

[0134]

[0135] Where: Q xy is the value of the xth evaluated object under the yth indicator, where y∈[1,b] means the evaluation method has b evaluation indicators; x∈[1,a] means there are a samples;

[0136] The data were normalized and standardized, and the calculation formula was:

[0137]

[0138] Where: Q' xy1 is a positive indicator, Q' xy2 It is a negative indicator; max(Q y )、min(Qy ) are the maximum and minimum values ​​of all objects under index y; Q' xy is the index value after standardization, and its value is 0≤Q' xy ≤1;

[0139] Calculate the entropy value of each evaluation indicator using the following formula:

[0140]

[0141] Where: S y represents information entropy, S y The larger the value of , the greater the amount of information contained in the data;

[0142] Calculate the weight of each evaluation indicator, the calculation formula is;

[0143]

[0144] A power optimization allocation model with SOC consistency as the objective function is established, and the calculation formula is:

[0145]

[0146] Where: SOC i,t is the SOC of battery energy storage unit i at the end of period t; SOC R is the reference value of SOC;

[0147] In the energy storage system, during time period t, the total charge / discharge power of each battery group must be equal to the dispatch instruction. The constraint is as follows:

[0148]

[0149] Where: is the discharge power of battery pack i in period t; is the charging power of battery pack i in period t;

[0150] Charge / discharge power constraint, the constraint formula is as follows;

[0151]

[0152] Where: and The maximum and minimum values ​​of the battery pack discharge power; and The maximum and minimum charging power for the battery pack;

[0153] The change constraint of battery pack SOC is as follows:

[0154]

[0155] Where: SOC i,t is the SOC of battery pack i at the end of period t; S N Battery rated capacity; They are the discharge and charge amounts of battery pack i in period t; n is the number of samples per hour; are the maximum and minimum allowed SOC of battery pack i respectively;

[0156] In order to further demonstrate the effectiveness of the energy storage system power allocation method based on battery health status prediction proposed in this application, the following examples are further provided here:

[0157] See Figure 3-4 , based on the convolutional neural network, the battery B5 and B7 data sets released by the NASA Ames Center are processed to extract the time series features. These features capture the local patterns and trends in the data. Then, the extracted time series features are processed using the long short-term memory network, which can capture the long-term dependencies and time series characteristics in the data. Finally, the features processed by the long short-term memory network are integrated through the fully connected layer, and the output layer summarizes the output values ​​of all neurons in the fully connected layer. The Bayesian optimization algorithm is used to optimize the network hyperparameters, and finally the prediction results of the battery pack health status are obtained. The hyperparameter optimization results are given by Figure 3 As shown by Figure 4 、 Figure 5 It can be seen that the effect of this method is compared with LSTM, BP, CNN and other methods. It shows that this method has better fitting effect and accuracy.

[0158] The battery pack prediction model based on BO-CNN-LSTM hybrid neural network is applied to the power distribution of energy storage system based on entropy weight method. The process is as follows: Figure 6 As shown in the figure, the charge state change trend of the battery pack with different initial health state values ​​obtained by using the actual operation data of the energy storage system is given by Figure 7 As shown, the battery pack attenuation for different initial values ​​of health status is given by Figure 8 As shown in the figure, the results show that the power allocation strategy based on the entropy weight method makes the charge state of each battery pack gradually differentiate during operation, and is divided into 5 categories according to the 5 different initial health state values ​​of the battery pack; and it can be seen that the action frequency of the battery pack is directly related to the initial health state value. Although the use of different data sets makes the average number of actions of each battery pack slightly different, the overall trend of the average number of actions increasing significantly with the improvement of the initial health state value is completely consistent. Compared with the average division method, the number of actions of all battery packs has decreased, among which the groups with an initial health state value lower than 0.9 have reduced by more than 90%, which is in line with the expected goal of reducing the number of battery pack actions. Figure 8It can be seen that when the battery health is poor, the optimized allocation strategy can significantly slow down the battery degradation rate and extend its service life. This is because the strategy can avoid excessive load on degraded batteries and reasonably allocate power demand.

[0159] The various component embodiments of the present application can be implemented in hardware, or in a software module running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all of the functions of some modules according to the embodiments of the present application. The application can also be implemented as a part or all of a device program (e.g., a computer program and a computer program product) for performing the method described herein. Such a program implementing the present application can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0160] It should be noted that, in this article, relational terms such as first and second are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variations thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements. Although the description of the present application is carried out in conjunction with the above specific embodiments, it is obvious that those skilled in the art can make many substitutions, modifications and changes based on the above content. Therefore, all such substitutions, improvements and changes are included within the spirit and scope of the appended claims.

Claims

1. A power allocation method for an energy storage system based on battery pack health status prediction, characterized in that: include: Based on the hybrid neural network model of convolutional neural network and long short-term memory network, the data set obtained from the lithium battery test is used to input the battery charge / discharge voltage, current and cycle number as feature quantities into the model training to obtain the lithium battery health status prediction model. The hybrid neural network model is optimized using the Bayesian optimization algorithm, and the root mean square RMES and determination coefficient R 2 Two indicators were used to verify the prediction effect; Based on the lithium battery health status prediction model, the health status of each battery pack is predicted in real time. Based on the entropy weight method, the charge state and health state of the battery pack are used as evaluation indicators to sort the charge / discharge priorities of the battery packs, select the appropriate battery packs, and determine the optimal allocated power of each battery pack with the charge state consistency of the battery pack as the objective function.

2. A method for allocating power of an energy storage system based on battery health status prediction according to claim 1, characterized in that: The lithium battery health status prediction model specifically includes: Collecting power parameters from battery operation, wherein the power parameters include at least SOC, SOH, voltage and current data; The power parameters are normalized and the calculation formula is: In the formula, X is the input data obtained after normalization; x is the original input data; x max and x min Respectively represent the maximum and minimum values; Input multi-dimensional time series data X: power parameters; Perform one-dimensional convolution on the input multi-dimensional time series data X, and the calculation formula is: Z t =f(W conv *X t:t+k-1 +b conv ); In the formula, Z t Represents the output features after convolution, extracting the local features of time step t, W conv represents the convolution kernel, the size is k*d, where k is the size of the convolution kernel, represents the length of the convolution window; d is the feature dimension of the input data; X t:t+k-1 represents the local time segment starting from time step t, that is, the input data subsequence from t to t+k-1; b conv represents the bias term of the convolution layer, which is used to adjust the convolution output; f(·) is the activation function, which introduces nonlinearity to enhance the expressiveness of the model. After the convolution operation, in order to reduce the dimension of the convolutional data while retaining the main features, maximum pooling is used for downsampling. The formula is: P t =max(Z t:t+p-1 ); P t is the output feature of time step t after pooling, indicating the result of the pooling operation; Z t:t+p-1 represents the convolution feature subsequence starting at time step t, with a length of p; max(·) is the maximum pooling operation, that is, selecting the maximum value within a given time window. The output features of the convolutional neural network are flattened and input into the long short-term memory network; Read the current input information and the unit state information at the previous moment, pass the information to the Sigmoid function at the same time, and decide whether to discard or retain the information based on the output value. The calculation formula is: f t =σ(W f [x t ,h t-1 ]+b f ); Where σ(·) is the sigmoid function with an output interval of (0,1), W f is the weight matrix of the forget gate structure; b f is the bias matrix of the forget gate structure; x t Represents the input of the network at the current moment; h t-1 Represents the output of the network at the previous moment; Calculate the input gate i through the sigmoid function t The output value is calculated as follows: i t =σ(W i [x t ,h t-1 ]+b i ); Where W i is the weight matrix of the input gate structure; b i is the bias matrix of the input gate structure; Using c t 'Indicates that the unit state input at the current moment is the unit state at the previous moment. The tanh function is used to filter the extracted valid information. The calculation formula is; c' t =tanh(W c [x t ,h t-1 ]+b c ); Where W c is the weight matrix of the unit; b c is the bias matrix of the unit. Using c t Represents the unit state at the current moment, and the calculation formula is; c t =f t ⊙c t-1 +i t ⊙c t '; Where ⊙ is the product of vector elements, c t It consists of two parts: the first part is based on the forget gate f t and the cell state c at the previous moment t-1 The second part is composed of input gate i t With c t 'Multiply to form; in this way, the current memory c t ' and long-term memory c t-1 Combined; The output of the network at the previous moment and the current input are passed to the sigmoid function, and the unit state at the previous moment is passed to the tanh function; then the output of tanh is multiplied by the output of sigmoid to determine the output h of the network at the next moment t The information carried will eventually pass the new unit state and new output to the next time step. The calculation formula is: the t =σ(W o [x] t ,h t-1 ]+b o ); Where W o is the weight matrix of the output gate structure; b o is the bias matrix of the output gate structure; Multiply the output of tanh by the output of sigmoid to determine the output h of the network at the next moment t The information carried will eventually pass the new unit state and new output to the next time step. The calculation formula is: h t =o t ⊙tanh(c t ); Output prediction value after the fully connected layer The calculation formula is; Where W fc The weight matrix of the fully connected layer; b fc is the bias of the fully connected layer.

3. A method for allocating power of an energy storage system based on battery health status prediction according to claim 2, characterized in that: The optimization of the hybrid neural network model by using the Bayesian optimization algorithm specifically includes: The training set is used to optimize BO hyperparameters, and the calculation formula is: With * =arg max f(z),z∈Z; In the formula, Z is the sample space; z is a sample point randomly selected in the sample space; z * is a globally optimal solution including real, integer, or categorical eigenvalues.

4. A method for allocating power of an energy storage system based on battery health status prediction according to claim 3, characterized in that: The root mean square (RMES) and the coefficient of determination (R) 2 The two indicators are used to verify the prediction effect, and the calculation formula is: Among them, the predicted value y'={y'1,y'2,...,y' n }, the true value y={y1,y2,...,y n }, n represents the number of predicted data.

5. The method for allocating power of an energy storage system based on battery health status prediction according to claim 4, characterized in that: Based on the established lithium battery health status prediction model, the health status of each battery pack is predicted in real time. Based on the entropy weight method, the charge state and health state of the battery pack are used as evaluation indicators to sort the charge / discharge priority of the battery pack, select the appropriate battery pack, and determine the optimal allocation power of each battery pack with the charge state consistency of the battery pack as the objective function, including: The energy storage system receives a power dispatch instruction; The corresponding number of battery pack actions is calculated using the following formula: Where: N t is the number of BESS battery packs charged / discharged at time t; P t is the power that BESS needs to dispatch at time t; is the rated power of the battery pack; λ is the redundancy factor; [] means rounding up; Evaluation indicators are established for multiple samples and multiple indicators, and the calculation formula is: Where: Q xy is the value of the xth evaluated object under the yth indicator, where y∈[1,b] represents the evaluation method has b evaluation indicators; x∈[1,a] represents the number of samples; The data were normalized and standardized, and the calculation formula was: Where: Q' xy1 is a positive indicator, Q' xy2 It is a negative indicator; max(Q y )、min(Q y ) are the maximum and minimum values ​​of all objects under index y; Q' xy is the index value after standardization, and its value is 0≤Q' xy ≤1; Calculate the entropy value of each evaluation index, the calculation formula is; Where: S y represents information entropy, S y The larger the value of, the greater the amount of information contained in the data; Calculate the weight of each evaluation indicator, the calculation formula is; A power optimization allocation model with SOC consistency as the objective function is established, and the calculation formula is: Where: SOC i,t is the SOC of battery energy storage unit i at the end of period t; SOC R is the reference value of SOC; In the energy storage system, during time period t, the total charge / discharge power of each battery group must be equal to the dispatch instruction. The constraint is as follows: Where: is the discharge power of battery pack i in period t; is the charging power of battery pack i in period t; Charge / discharge power constraint, the constraint formula is as follows; Where: and The maximum and minimum values ​​of the battery pack discharge power; and The maximum and minimum charging power for the battery pack; The change constraint of battery pack SOC is as follows: Where: SOC i,t is the SOC of battery pack i at the end of period t; S N Battery rated capacity; They are the discharge and charge amounts of battery pack i in period t; n is the number of samples per hour; are the maximum and minimum allowed SOC of battery pack i respectively.

Images