Energy storage system battery health degree estimation method, electronic equipment and medium

Abstract

The invention discloses an energy storage system battery health degree estimation method, electronic equipment and a medium, and the method comprises the steps: constructing a pulse neural network, collecting first historical feature data of a first software-defined energy storage system in a battery relaxation process and a corresponding first battery health state true value, performing source task training on the pulse neural network to obtain a first synaptic conductance matrix; taking the first synaptic conductance matrix as a network weight of the pulse neural network for target task training; collecting second historical characteristic data of a second software defined energy storage system in the battery relaxation process and a corresponding second battery health state true value; performing target task training on the pulse neural network to obtain a second synaptic conductance matrix; and taking the second synaptic conductance matrix as the network weight of the trained spiking neural network, collecting to-be-predicted feature data, and inputting the to-be-predicted feature data into the trained spiking neural network, thereby estimating the health state of the battery.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric energy storage, and in particular relates to a method for estimating the health of a battery in an energy storage system, an electronic device, and a medium. Background Art

[0002] Currently, common methods for estimating the health of batteries in energy storage systems include estimation based on battery aging models and data-driven machine learning methods. For example, health estimation is based on charge and discharge characteristics or battery parameters such as voltage, current, and temperature. However, these methods have the following disadvantages:

[0003] 1. Traditional battery aging models fail to fully account for the battery's recovery effect, resulting in inaccurate estimates. This is especially true after the battery has undergone a certain recovery period, failing to reflect its actual condition. Furthermore, traditional energy storage systems struggle to isolate the interactions between individual batteries, leading to excessive noise when collecting raw data.

[0004] 2. Existing data-driven methods mostly rely on large amounts of historical data. They are unable to achieve real-time updates and adaptive adjustments when faced with battery time series data and dynamic changes, and lack the ability to model time-dependent features. Furthermore, the feature construction of these methods often lacks practical significance, and the models are complex and lack generalization, making them unsuitable for widespread application.

[0005] These shortcomings are primarily due to the inability of existing technologies to isolate individual batteries during energy storage operation, making it difficult to collect battery characteristics. Over-reliance on historical data also prevents capturing battery time series data and dynamic changes during actual operation, impacting the accuracy and applicability of battery health estimation in energy storage systems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention provides a method, electronic device, and medium for estimating the health of batteries in energy storage systems. These methods effectively isolate the interactions between internal batteries during energy storage operation, improving the accuracy of battery estimation in energy storage systems. Furthermore, by combining spiking neural networks with transfer learning, the present invention effectively reduces the reliance of energy storage system battery health estimation on historical data from the target energy storage system, expanding the application scenarios of the estimation method, thereby reducing the number of energy storage system tests and improving the method's generalization and accuracy.

[0007] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a method for estimating the health of a battery in an energy storage system, the method comprising:

[0009] Collecting first historical characteristic data of the first software-defined energy storage system during a battery relaxation process and a corresponding first battery health status true value;

[0010] A spiking neural network is constructed, with the first historical feature data as input and the first battery health status true value as a target, thereby performing source task training on the spiking neural network to obtain a first synaptic conductance matrix;

[0011] The first synaptic conductance matrix is used as the network weight for target task training of the spiking neural network; second historical feature data of the second software-defined energy storage system during battery relaxation and the corresponding second battery health state true value are collected; the second historical feature data is used as input and the second battery health state true value is used as a target to train the spiking neural network for the target task, thereby obtaining a second synaptic conductance matrix;

[0012] The second synaptic conductance matrix is used as the network weight of the trained spiking neural network, and the characteristic data to be predicted of the second software-defined energy storage system during the battery relaxation process is collected and input into the trained spiking neural network to estimate the battery health state.

[0013] In a second aspect, an embodiment of the present invention provides an electronic device comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned energy storage system battery health estimation method.

[0014] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the above-mentioned method for estimating the health status of batteries in an energy storage system when the program is executed by a processor.

[0015] In a fourth aspect, an embodiment of the present invention provides a computer program product, including a computer program / instruction, which implements the above-mentioned method for estimating the health status of batteries in an energy storage system when executed by a processor.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] This application proposes a software-defined method for estimating the battery health of energy storage systems based on spiking neural networks and transfer learning. By isolating the interactions between batteries in the energy storage system, it can significantly improve the accuracy of battery health estimation. The spiking neural network improves sensitivity to timing information and captures dynamic battery information. Transfer learning reduces the testing cost of the energy storage system and prevents the energy storage system from shutting down due to battery health estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 This is a flow chart of a method for estimating the health of a battery in an energy storage system provided by an embodiment of the present invention;

[0020] Figure 2 This is a framework diagram of a software-defined energy storage system provided by an embodiment of the present invention;

[0021] Figure 3 is a schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below with reference to the following examples. The following examples are provided only to facilitate understanding of the present invention. It should be noted that, without departing from the principles of the present invention, a number of improvements and modifications may be made to the present invention by those skilled in the art, and such improvements and modifications fall within the scope of the claims of the present invention.

[0023] In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0024] like Figure 1 As shown, an embodiment of the present invention provides a method for estimating the health of a battery in an energy storage system, the method comprising the following steps:

[0025] Step S1: collecting first historical characteristic data of a first software-defined energy storage system during a battery relaxation process and a corresponding first battery health status real value.

[0026] Further, if Figure 2 As shown, the software-defined energy storage system includes:

[0027] The battery cell management unit includes a battery cell and a high-frequency power electronic switch connected in series with the battery cell; it is used to collect the voltage, current and temperature of the battery cell and control the high-frequency power electronic switch to control the connection and disconnection of the target battery cell;

[0028] A battery cluster management unit, comprising a plurality of battery cell management layers connected in series and / or in parallel;

[0029] The battery array management unit includes several battery cluster management units.

[0030] Furthermore, the historical characteristic data includes: maximum voltage, temperature at the end of charging, voltage within 5 minutes after the end of charging, voltage within 10 minutes after the end of charging, and voltage within 15 minutes after the end of charging.

[0031] It should be noted that the core idea of the software-defined energy storage system is to combine a large number of battery cells with high-frequency power electronic switches in series and parallel. In actual operation, the high-frequency action of the switch is used to select specific battery cells to connect to the system, thereby changing the series and parallel relationship of the existing topology. Figure 2 As shown, the bottom layer is the cell management layer. The cell management unit manages the battery cells and their associated high-frequency power electronic switches. This management unit is primarily responsible for collecting information such as cell voltage, current, and temperature, controlling the high-frequency power electronic equipment to connect and disconnect the target battery cells, generating alarms for abnormal conditions, and communicating with the battery cluster management unit. Next is the battery cluster management layer, controlled by the battery cluster management unit. This management unit manages a battery cluster consisting of multiple battery cells and is responsible for current and total voltage acquisition, leakage detection, and power-off protection when a cell's status is abnormal. The battery array management unit centrally manages the batteries in the entire energy storage battery stack. It connects downward to each battery cluster management unit and upwardly interacts with other devices, providing feedback on the operating status of the battery array. The energy flow in a software-defined energy storage system is mediated by charge transfer between battery cells, high-frequency power electronic switches, and PCs. This information flow is transmitted between the various management units via buses and other means. Unlike traditional energy storage systems, which regulate energy dissipation by adjusting current levels, software-defined energy storage systems regulate battery energy by controlling the charge and discharge times of the batteries, thereby converting a continuous energy flow into a digital signal associated with the signal stream. The specific control method is shown in the following formula:

[0032]

[0033] Where ΔQ represents the charge output by the energy storage system, n B is the number of battery cells in the energy storage system, I i and t i Represent the charge and discharge current and charge and discharge time of the i-th battery cell respectively. When the battery cell is discharged, I i Take positive when charging and negative when charging.

[0034] Through the above structure, various physical information about the energy storage system can be represented in software. Specifically, the charge and discharge capacity of a battery cell can be represented by the product of its charge and discharge time and current. High-frequency power electronic components are represented in the software as Boolean variables. The PCS (energy storage converter) is primarily represented in the software as a filtering and rectification function, and each management unit represents the generation and reception of communication signals.

[0035] Furthermore, because the discharge process of energy storage systems varies during application, the depth of discharge is difficult to standardize, and the current of energy storage systems fluctuates during charging, making it difficult to extract stable features for model training during the charging and discharging process. After the external current is interrupted (such as when charging / discharging stops), the phenomenon of the non-equilibrium state caused by electrochemical reactions or uneven material distribution within the battery gradually returning to equilibrium is a relaxation process that energy storage systems inevitably undergo after being fully charged. The relaxation voltage of an energy storage battery refers to the process by which its terminal voltage gradually stabilizes over time after charging and discharging stops, and the equilibrium value ultimately reached. In a software-defined energy storage system, the state of each battery can be well collected, allowing for accurate estimation.

[0036] The software-defined energy storage system on which this example is based can determine whether a battery cell is fully charged based on the change in charging current. After it is fully charged, it can be disconnected and left idle for a period of time without being called, thereby collecting its relaxation process. This example extracts five features to estimate the battery health. They are: maximum voltage V r_max , the temperature Te at the end of charging, the voltage V every 5 minutes within 15 minutes after the end of charging r,5min 、V r,10min 、V r,15min The specific collected data is shown in the following matrix:

[0037]

[0038] The dimensions of the five features mentioned above are not consistent, which can easily lead to a certain feature dominating the model training due to a significant difference. To avoid this phenomenon, the data needs to be normalized:

[0039]

[0040] Where, Ch' i is the normalized data, Ch i is any item in the collected data matrix, express The corresponding vectors in , Max[] and Min[] represent the maximum and minimum values in the vector respectively.

[0041] Step S2: construct a pulse neural network, take the first historical feature data as input, take the first battery health status true value as target, and perform source task training on the pulse neural network to obtain the first synaptic conductance matrix

[0042] It's important to note that spiking neural networks (SNs) are a new generation of biologically inspired artificial neural network models. They are a subset of deep learning and have a strong biological basis. In SNs, neurons are modeled to simulate the current state, typically represented by their current activation level. Pulse encoding of input signals allows them to propagate through the SN, thereby training the neural network.

[0043] The membrane potential of a neuron changes over time and receives input pulses from excitatory and inhibitory synapses. The intensity of the impact of the input pulses on the neuron is related to the weights between the synapses. The membrane potential of a neuron will continue to accumulate with the input signal. When the membrane potential exceeds the threshold, the neuron will discharge, that is, emit a pulse signal. After the neuron exceeds the threshold and releases a pulse signal, the neuron will perform a reset operation to reset its membrane potential to the resting potential. At the same time, it will be in a fatigue state for a certain period of time and cannot be activated again. The dynamics of the membrane potential are affected by excitatory and inhibitory synapses. Each synaptic type has its own time constant, which controls the rate at which the membrane potential decays after the neuron discharges. The neuron in the present invention adopts the LIF model, and the module voltage V is expressed by the following formula:

[0044]

[0045] Among them, E r is the resting potential of the neuron, τ is the decay time constant of the membrane potential, g e and g in are the weights of excitatory and inhibitory synapses, E e and E in are the equilibrium potentials of excitatory and inhibitory synapses, respectively. The weights in the formula represent the connections between neurons, and the size of the weights indicates the degree of influence of other synapses on the current neuron.

[0046] The present invention constructs a neural network by an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is the same as the number of features, and the output layer is a single neuron. The estimation effect can be improved by gradually adjusting the number of neurons in the hidden layer. The number of neurons in the three hidden layers set by the present invention is 62, 34, and 13, respectively. A full connection is used between the input layer and the first hidden layer, and a probabilistic link is used between the second hidden layer and the third hidden layer. The connection between neurons can be interpreted as the pulse emitted by the previous neuron affecting the pulse of the connected neuron, while the unconnected neurons will not be affected, and the prominence weight between them can be regarded as 0.

[0047] The output layer converts pulse intensity into the energy storage system's battery health. This layer consists of one neuron. The pulse frequency of this neuron is detected and normalized to represent the battery health.

[0048] Step S3: the first synaptic conductance matrix As the network weights for target task training of the pulse neural network; collecting the second historical feature data of the second software-defined energy storage system during the battery relaxation process and the corresponding second battery health state true value; using the second historical feature data as input and the second battery health state true value as the target, the pulse neural network is trained for the target task to obtain the second synaptic conductance matrix

[0049] Furthermore, in spiking neural networks, information is transmitted based on the timing characteristics of pulses. Once the input signal is converted into a pulse, the pulse propagates along the synaptic connections between neurons. The membrane potential of each neuron changes as the input signal is transmitted, and the neuron will only emit a pulse when the membrane potential exceeds a threshold. Therefore, the processed data needs to be encoded to form a pulse signal. Taking frequency encoding as an example, the intensity of the input signal is represented by the frequency of the neuron's pulse emission. The input signal and the frequency of the neuron's pulse emission can be expressed by the following formula:

[0050] f Ch,i =a·Ch i '

[0051] Where a is the gain coefficient, f Ch,i is the frequency at which the neuron fires.

[0052] In biological neural systems, the process of neuron firing is usually random and follows certain statistical laws. The Poisson process can effectively describe this phenomenon of random firing of pulses. The characteristics of Poisson pulses simulate the firing pattern of biological neurons when processing information. Poisson pulses encode information by adjusting the firing frequency. Even if the time interval between pulses is random, the firing frequency of the pulses (that is, the number of pulses fired per unit time) can carry rich information. Therefore, after obtaining the pulse firing frequency based on the feature generation, using the Poisson process to generate pulses can better utilize the feature information. The probability function of the Poisson distribution is as follows:

[0053]

[0054] Where λ i =f Ch,i is the average pulse emission frequency of the neuron per unit time, Δt is the time interval, and m is the number of pulses emitted per unit time. That is, in each time window Δt, the probability of the neuron emitting a pulse is Then a random number r is generated in the interval [0,1], if r≤p i , then the neuron generates a pulse in this time window.

[0055] Furthermore, in this example, the spiking neural network is trained by combining the synaptic plasticity mechanism STDP and supervised learning loss. Specifically, the following steps are involved:

[0056] STDP is a biologically inspired learning rule that adjusts synaptic weights based on the timing differences in neuronal spikes. When one neuron's spike (presynaptic) precedes the spike of another neuron (postsynaptic) (LTP), the synaptic weight of the connection increases. Conversely, if the presynaptic neuron spikes after the postsynaptic neuron (LTD), the synaptic weight decreases. Through this mechanism, STDP enables neural networks to automatically adjust their connections based on temporal order, thereby achieving learning. Specifically, if a hidden layer neuron spikes before an output layer neuron spikes, this means that the hidden layer neuron's activation (spike) occurs before the predicted SOH value. This neuron's activation is more likely to influence the SOH prediction, conforming to the "presynaptic neuron precedes the postsynaptic neuron" pattern. In this case, the forward part of the STDP rule increases the connection weight. However, if a hidden layer neuron spikes after an output layer neuron spikes, indicating that the output layer neuron activates after the hidden layer neuron, this "postsynaptic neuron precedes the presynaptic neuron" timing pattern will cause the synaptic weight to decrease (via the reverse part of STDP). As training progresses, the STDP rule continuously adjusts synaptic weights, gradually strengthening those spike timing relationships that help correctly predict SOH and weakening those incorrect timing patterns. This learning mechanism enables the network to gradually learn the mapping relationship between input features and SOH by adjusting the timing patterns over multiple training samples. The specific synaptic adjustment formula is as follows:

[0057] Δw=η(x pre -x tar )(w max -w) μ

[0058] In the formula, η represents the learning rate, x pre is the activation moment of the neuron's presynaptic pulse firing, x tar is the activation moment of the postsynaptic neuron, w max The maximum value of the synaptic weight, w is the current synaptic weight, and μ is the factor that controls the dependency of weight changes on the current synaptic weight.

[0059] In a certain step of simulation training, for the two first neurons S1 and second neurons S2 before and after a specific synapse A, it is assumed that the first neuron S1 is at the time [t 11 , t 12 , t 13 ] generates a pulse, and the second neuron S2 generates a pulse at time [t 21 , t 22 ] generates a pulse. In this simulation training, STDP will match all the pulses of the previous and next neurons across time, that is, Δt=[t 11 -t 21 , t11 -t 22 , t 21 -t 21 , t 21 -t 22 , t 31 -t 21 , t 31 -t 22 ]. The obtained time difference is then substituted into the synaptic adjustment formula to obtain the weight update size in this simulation.

[0060] According to the above formula, the influence of neuron connection paths irrelevant to the energy storage system battery health estimation on the results can be further reduced, thereby improving the accuracy of the energy storage system battery health estimation.

[0061] The supervised learning loss is then introduced as an auxiliary signal into the learning process of the SNN, making the network output closer to the true SOH value. The weights between the third hidden layer and the output layer need to be adjusted based on the gap between the predicted value and the true SOH.

[0062] L=soh i '-soh i

[0063] w new =w old -η ij Lw old

[0064] In the formula, soh i ' is the SOH prediction value output after simulation training, soh i is the real SOH value. old and w new They are the weights after stdp update and the final weights after introducing supervised learning loss correction.

[0065] In step S4, the second synaptic conductance matrix is used as the network weight of the trained spiking neural network, and the characteristic data to be predicted of the second software-defined energy storage system during the battery relaxation process is collected and input into the trained spiking neural network to estimate the battery health state.

[0066] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method for estimating the health of batteries in an energy storage system. Figure 3 As shown in FIG, a hardware structure diagram of any device with data processing capability in which the method for estimating the health of a battery in an energy storage system provided by an embodiment of the present invention is used, except Figure 3In addition to the processor, memory, and network interface shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0067] Accordingly, the present application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned method for estimating the battery health of an energy storage system. The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The computer-readable storage medium may also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0068] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A method for estimating the health of a battery in an energy storage system, characterized in that: The method comprises: Collecting first historical characteristic data of the first software-defined energy storage system during a battery relaxation process and a corresponding first battery health status true value; A spiking neural network is constructed, with the first historical feature data as input and the first battery health status true value as a target, thereby performing source task training on the spiking neural network to obtain a first synaptic conductance matrix; The first synaptic conductance matrix is used as the network weight for training the spiking neural network for the target task; second historical feature data of the second software-defined energy storage system during the battery relaxation process and the corresponding second battery health state true value are collected; the spiking neural network is trained for the target task using the second historical feature data as input and the second battery health state true value as a target to obtain a second synaptic conductance matrix; The second synaptic conductance matrix is used as the network weight of the trained spiking neural network, and the characteristic data to be predicted of the second software-defined energy storage system during the battery relaxation process is collected and input into the trained spiking neural network to estimate the battery health state.

2. The method for estimating the health of a battery in an energy storage system according to claim 1, wherein: Software-defined energy storage systems include: The battery cell management unit includes a battery cell and a high-frequency power electronic switch connected in series with the battery cell; it is used to collect the voltage, current and temperature of the battery cell and control the high-frequency power electronic switch to control the connection and disconnection of the target battery cell; A battery cluster management unit, comprising a plurality of battery cell management layers connected in series and / or in parallel; The battery array management unit includes several battery cluster management units.

3. The method for estimating the health of a battery in an energy storage system according to claim 1, wherein: Historical characteristic data includes: maximum voltage, temperature at the end of charging, voltage within 5 minutes after charging, voltage within 10 minutes after charging, and voltage within 15 minutes after charging.

4. The method for estimating the health of a battery in an energy storage system according to claim 1, wherein: The spiking neural network includes an input layer, three hidden layers, and an output layer; wherein the number of neurons in the input layer is the same as the number of features, and the output layer has one neuron; a full connection is used between the input layer and the first hidden layer, and a probabilistic link is used between the second hidden layer and the third hidden layer.

5. The method for estimating the health of a battery in an energy storage system according to claim 4, wherein: The neuron adopts the LIF model, and the model voltage V is expressed by the following formula: Among them, E r is the resting potential of the neuron, τ is the decay time constant of the membrane potential, g e and g in are the weights of excitatory and inhibitory synapses, E e and E in are the equilibrium potentials of excitatory and inhibitory synapses, respectively.

6. The method for estimating the health of a battery in an energy storage system according to claim 1, wherein: During the training of the spiking neural network, the expression of synaptic adjustment is as follows: Δw=η(x pre -x tar )(w max -w) μ In the formula, η represents the learning rate, x pre is the activation moment of the neuron's presynaptic pulse firing, x tar is the activation moment of the postsynaptic neuron, w max The maximum value of the synaptic weight, w is the current synaptic weight, and μ is the factor that controls the dependency of weight changes on the current synaptic weight.

7. The method for estimating the health of a battery in an energy storage system according to claim 1, wherein: The spiking neural network training process also includes: Combining the synaptic plasticity mechanism STDP and supervised learning loss to train the pulse neural network, the expression is as follows: L=soh i '-soh i w new =w old -η ij Lw old In the formula, soh i ' is the SOH estimate output by the pulse neural network, soh i is the true value of SOH, w old and w new They are the weights after stdp update and the final weights after introducing supervised learning loss correction.

8. An electronic device comprising a memory and a processor, characterized in that: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the energy storage system battery health estimation method according to any one of claims 1 to 7 above.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for estimating the health status of a battery in an energy storage system as described in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the method for estimating the health status of a battery in an energy storage system as described in any one of claims 1 to 7 is implemented.

Images