Energy storage lithium ion battery thermal runaway temperature estimation method and system based on internal short circuit quantitative diagnosis

Abstract

The invention belongs to the field of lithium ion battery fault diagnosis, and discloses an energy storage lithium ion battery thermal runaway temperature estimation method and system based on internal short circuit quantitative diagnosis, and the method comprises the steps: building a lithium ion battery internal short circuit equivalent circuit model, and carrying out the online identification of model parameters; based on the internal short-circuit equivalent circuit model of the lithium ion battery after model parameter updating, estimating short-circuit current and short-circuit resistance by using a particle filtering algorithm; the short-circuit resistor and the short-circuit current serve as input, the battery temperature serves as output, the thermal runaway temperature is estimated based on an extreme learning machine algorithm, and when the estimated temperature exceeds the thermal runaway characteristic temperature, thermal runaway warning information is sent out. The lithium ion battery thermal runaway detection is realized by quantitatively identifying the internal short-circuit resistance and short-circuit current of the lithium ion battery and estimating the thermal runaway temperature of the lithium ion battery in combination with an extreme learning machine algorithm.

Description

Technical Field

[0001] The present invention relates to thermal runaway temperature estimation of energy storage lithium-ion batteries, belongs to the field of lithium-ion battery fault diagnosis, and particularly relates to a thermal runaway temperature estimation method and system for energy storage lithium-ion batteries based on quantitative internal short circuit diagnosis. Background Art

[0002] Renewable energy will play an increasingly important role in China's energy system. However, the large-scale integration of renewable energy into the grid will severely impact the safe and stable operation of the power grid. Therefore, energy storage technology is one of the key core technologies supporting the widespread application of renewable energy. With the advancement of battery technology and the reduction of costs, electrochemical energy storage technology is gaining increasing attention. Lithium-ion batteries, in particular, have been widely used in electrochemical energy storage systems due to their high safety and excellent cycle performance. However, the frequent fires and explosions in lithium-ion battery energy storage power stations have sounded the alarm. The safety of the material system does not guarantee the safety of the integrated battery cell. As an energy carrier, lithium-ion batteries pose significant safety risks due to thermal runaway.

[0003] According to statistics, internal short circuits account for 52% of the various factors that can cause thermal runaway in lithium-ion battery systems, far exceeding other battery misuse factors. This indicates that internal short circuits are the most common cause of thermal runaway incidents in lithium-ion batteries. In other words, internal short circuits in lithium-ion batteries are closely linked to thermal runaway incidents. Prompt detection and identification of internal short circuits in lithium-ion batteries can effectively prevent most thermal runaway incidents, thereby ensuring the safe and stable operation of lithium-ion battery systems.

[0004] Existing research has conducted extensive work on the thermal runaway propagation mechanism and suppression strategies, and on early internal short-circuit fault diagnosis, achieving certain progress. However, existing research on thermal runaway primarily relies on advanced sensing technologies such as combustible gas detection, smoke / temperature sensing monitoring, and flame detection to achieve thermal runaway detection and early warning when a thermal runaway accident occurs. Limited consideration is given to thermal runaway detection and early warning based on internal short circuits in lithium-ion batteries. Furthermore, existing internal short-circuit diagnostic methods primarily focus on achieving internal short-circuit diagnosis at an early stage, with limited consideration given to combining internal short-circuit faults with thermal runaway to achieve thermal runaway detection and early warning based on internal short-circuit diagnostic results. Summary of the Invention

[0005] To solve the problems existing in the prior art, the present invention provides a method and system for estimating the thermal runaway temperature of an energy storage lithium-ion battery based on quantitative diagnosis of internal short circuits. By quantitatively identifying the internal short-circuit resistance and short-circuit current of the lithium-ion battery, combined with the extreme learning machine algorithm, the thermal runaway temperature of the lithium-ion battery is estimated, thereby realizing thermal runaway detection of the lithium-ion battery.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for estimating thermal runaway temperature of an energy storage lithium-ion battery based on quantitative diagnosis of internal short circuits, the method comprising:

[0008] Establish an equivalent circuit model for the internal short circuit of lithium-ion batteries and identify the model parameters online;

[0009] Based on the internal short-circuit equivalent circuit model of the lithium-ion battery after the model parameters are updated, the particle filter algorithm is used to estimate the short-circuit current and short-circuit resistance;

[0010] With short-circuit resistance and short-circuit current as input and battery temperature as output, the thermal runaway temperature is estimated based on the extreme learning machine algorithm. When the estimated temperature exceeds the thermal runaway characteristic temperature, a thermal runaway alarm message is issued.

[0011] Preferably, the method for establishing an internal short-circuit equivalent circuit model of a lithium-ion battery includes:

[0012] U p,k =exp(-Δt / (R p C p ))U p,k-1 +(1-exp(-Δt / (R p C p )))R p (I sc,k-1 +I L,k-1 );

[0013] U t,k =U oc,k -U p,k -(I sc,k +I L,k )R0;

[0014] Where Δt represents the sampling time, subscript k represents the kth moment, R0 represents the battery internal resistance, R p and C p Represent polarization resistance and polarization capacitance respectively, U p,k , I sc,k , I L,k 、U t,k 、U oc,k They represent the polarization voltage, short-circuit current, external load current, terminal voltage and open-circuit voltage at the kth moment respectively.

[0015] Preferably, the method for online identification of model parameters includes:

[0016] Based on FFRLS identification, the parameter vector θ is obtained k =[θ1,θ2,θ3,θ4], the model parameters are updated as follows:

[0017]

[0018] Where R sc Indicates the internal short-circuit resistance, U oc is the battery open circuit voltage, R0 represents the battery internal resistance, R p represents the polarization resistance, α is the intermediate process parameter of the equivalent circuit model, and θ1, θ2, θ3, and θ4 are the intermediate process parameters of parameter identification.

[0019] Preferably, the method for estimating thermal runaway temperature based on the extreme learning machine algorithm includes:

[0020] With short-circuit current and short-circuit resistance as input and lithium-ion battery temperature as output, based on the extreme learning machine algorithm, by establishing the relationship between short-circuit current and short-circuit resistance, which characterize the severity of internal short circuit, and temperature, which characterizes thermal runaway, the relationship between internal short circuit and thermal runaway is established, and the thermal runaway temperature is estimated.

[0021] The present invention also provides a thermal runaway temperature estimation system for energy storage lithium-ion batteries based on internal short circuit quantitative diagnosis, the system is used to implement the above method, the system includes: an identification module, a first estimation module, and a second estimation module;

[0022] The identification module is used to establish an internal short-circuit equivalent circuit model of the lithium-ion battery and identify model parameters online;

[0023] The first estimation module is used to estimate the short-circuit current and short-circuit resistance using a particle filter algorithm based on the lithium-ion battery internal short-circuit equivalent circuit model after the model parameters are updated;

[0024] The second estimation module is used to estimate the thermal runaway temperature based on the extreme learning machine algorithm using the short-circuit resistance and short-circuit current as input and the battery temperature as output, and to issue a thermal runaway alarm message when the estimated temperature exceeds the thermal runaway characteristic temperature.

[0025] Preferably, the process of establishing an internal short-circuit equivalent circuit model of a lithium-ion battery includes:

[0026] U p,k =exp(-Δt / (R p C p ))U p,k-1 +(1-exp(-Δt / (R p C p )))R p (I sc,k-1 +I L,k-1 );

[0027] U t,k =U oc,k -Up,k -(I sc,k +I L,k )R0;

[0028] Where Δt represents the sampling time, subscript k represents the kth moment, R0 represents the battery internal resistance, R p and C p Represent polarization resistance and polarization capacitance respectively, U p,k , I sc,k , I L,k 、U t,k 、U oc,k They represent the polarization voltage, short-circuit current, external load current, terminal voltage and open-circuit voltage at the kth moment respectively.

[0029] Preferably, the process of online identification of model parameters includes:

[0030] Based on FFRLS identification, the parameter vector θ is obtained k =[θ1,θ2,θ3,θ4], the model parameters are updated as follows:

[0031]

[0032] Where R sc Indicates the internal short-circuit resistance, U oc is the battery open circuit voltage, R0 represents the battery internal resistance, R p represents the polarization resistance, α is the intermediate process parameter of the equivalent circuit model, and θ1, θ2, θ3, and θ4 are the intermediate process parameters of parameter identification.

[0033] Preferably, the process of thermal runaway temperature estimation based on the extreme learning machine algorithm includes:

[0034] With short-circuit current and short-circuit resistance as input and lithium-ion battery temperature as output, based on the extreme learning machine algorithm, by establishing the relationship between short-circuit current and short-circuit resistance, which characterize the severity of internal short circuit, and temperature, which characterizes thermal runaway, the relationship between internal short circuit and thermal runaway is established, and the thermal runaway temperature is estimated.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention proposes a method for estimating the thermal runaway temperature of energy storage lithium-ion batteries based on quantitative diagnosis of internal short circuits. The proposed method for estimating the thermal runaway temperature of lithium-ion batteries includes the following three main steps: (1) Considering the balance between model accuracy and computational complexity, an equivalent circuit model of the internal short circuit of the lithium-ion battery is established, and the model parameters are identified online based on the forgetting factor recursive least squares (FFRLS) algorithm. (2) With the short-circuit current as one of the state variables, a lithium-ion battery state space model is constructed, and the short-circuit current is estimated based on the particle filter algorithm. Subsequently, the short-circuit resistance is calculated based on Ohm's law to achieve quantitative estimation of the internal short-circuit fault. (3) With the short-circuit resistance and short-circuit current as inputs and the battery temperature as output, the thermal runaway temperature is estimated based on the extreme learning machine algorithm. When the estimated temperature exceeds the thermal runaway characteristic temperature, a thermal runaway alarm message is issued, thereby ensuring the safe and stable operation of the battery system. The present invention realizes thermal runaway detection of lithium-ion batteries by quantitatively identifying the internal short-circuit resistance and short-circuit current of lithium-ion batteries and estimating the thermal runaway temperature of lithium-ion batteries in combination with an extreme learning machine algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in 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 paying any creative work.

[0038] Figure 1 Schematic diagram of the Thevenin model of a battery under internal short circuit according to an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the three-layer ELM neural network structure according to an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the self-heating temperature estimation result according to an embodiment of the present invention;

[0041] Figure 4 Schematic diagram of the pressure relief starting temperature estimation result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example 1

[0045] This paper proposes a method for estimating the thermal runaway temperature of energy storage lithium-ion batteries based on quantitative diagnosis of internal short circuits. The proposed method includes the following three main steps:

[0046] (1) Considering the balance between model accuracy and computational complexity, an equivalent circuit model of the internal short circuit of lithium-ion batteries is established, and the model parameters are identified online based on the forgetting factor recursive least squares (FFRLS) algorithm. (2) With the short-circuit current as one of the state variables, a state space model of the lithium-ion battery is constructed. The short-circuit current is estimated based on the particle filter algorithm, and then the short-circuit resistance is calculated based on Ohm's law to achieve quantitative estimation of internal short-circuit faults. (3) With the short-circuit resistance and short-circuit current as inputs and the battery temperature as output, the thermal runaway temperature is estimated based on the extreme learning machine algorithm. When the estimated temperature exceeds the thermal runaway characteristic temperature, a thermal runaway alarm is issued to ensure the safe and stable operation of the battery system.

[0047] The implementation process of the method proposed in the present invention mainly involves the following technologies: lithium-ion battery modeling and parameter identification under internal short circuit, quantitative diagnosis of internal short circuit based on particle filtering algorithm, and thermal runaway temperature estimation based on extreme learning machine. The present invention below focuses on the implementation process of the above three steps.

[0048] In this embodiment, the lithium-ion battery modeling and parameter identification under internal short circuit are as follows:

[0049] (1) Battery modeling under internal short circuit:

[0050] Considering that the Thevenin model of lithium-ion batteries can achieve a good balance between accuracy and computational complexity, the present invention captures the internal short-circuit fault characteristics of lithium-ion batteries based on the Thevenin model. During the process of internal short circuit of a single cell, the short-circuit resistance will continue to consume battery energy. This process is equivalent to connecting a short-circuit resistor in parallel at both ends of the battery. Therefore, the battery model under internal short circuit is as follows: Figure 1 As shown. Among them, U oc is the open circuit voltage of the battery. I represents the current flowing through the battery. R0 represents the internal resistance of the battery, R p and C p Represent polarization resistance and polarization capacitance respectively. p is the voltage drop across the polarization resistor.L Represents the external load current, which can be measured in real time by the current sensor. sc Indicates the internal short-circuit resistance, I sc By R sc The short-circuit current caused by this is not measurable during the actual operation of the battery pack. t Indicates the battery terminal voltage.

[0051] According to Kirchhoff's law, Figure 1 The dynamic behavior of the equivalent circuit model under internal short circuit can be expressed as:

[0052]

[0053] U t =U oc -U p -(I sc +I L )R0 (2)

[0054] Discretizing the equivalent circuit model yields:

[0055] U p,k =exp(-Δt / (R p C p ))U p,k-1 +(1-exp(-Δt / (R p C p )))R p (I sc,k-1 +I L,k-1 ) (3)

[0056] U t,k =U oc,k -U p,k -(I sc,k +I L,k )R0 (4)

[0057] Where Δt represents the sampling time, subscript k represents the kth moment; U p,k , I sc,k , I L,k 、U t,k 、U oc,k They represent the polarization voltage, short-circuit current, external load current, terminal voltage and open-circuit voltage at the kth moment respectively. Let α=exp(-Δt / (C p R p )), then:

[0058] U p,k =αU p,k-1 +(1-α)R p (I sc,k-1 +I L,k-1) (5)

[0059] (2) Online update of model parameters based on FFRLS algorithm:

[0060] The present invention first assigns a larger initial value to the short-circuit resistance. Then, based on the real-time measurement of the battery pack current and the terminal voltage of the internal short-circuit monomer, the short-circuit resistance is assigned a larger initial value. α 、R p , R0 and open circuit voltage U oc Model parameters such as are updated online. Figure 1 As shown in Figure 2, during the internal short circuit of a lithium-ion battery, the voltage across the short-circuit resistor is equal to the battery terminal voltage, so Substituting it into equations (4) and (5), we can obtain:

[0061]

[0062] Based on formula (6) and formula (7), we can further deduce:

[0063]

[0064] Where: α=exp(-Δt / (C p R p )) has been given in front of formula (5) and is the intermediate process parameter of the equivalent circuit model; θ1, θ2, θ3 and θ4 are the intermediate process parameters of parameter identification, and their calculation formulas are as follows:

[0065]

[0066] Reconstruct Equation (8) into the following form:

[0067] y k =Φ k θ k (13)

[0068] in:

[0069]

[0070] Where y k is the output variable of the system at time k; Φ k is the input data vector of the system at time k; θ k is the parameter vector of the system at time k.

[0071] FFRLS can effectively overcome the uncertainty of model parameters, thereby accurately capturing the real-time characteristics of the system and achieving reliable identification of system parameters. Based on formula (13), FFRLS updates the system parameter vector θ online through the following recursive formula k :

[0072]

[0073] Where K Ls,k is the gain of the algorithm; P Ls,k is the error covariance matrix of the state estimate; I represents the identity matrix; represents the estimated value of the system parameter vector at the kth moment; μ is the forgetting factor. When μ=1, formula (15) degenerates into the traditional recursive least squares algorithm. The present invention sets it to 0.998.

[0074] The parameter vector θ is obtained based on FFRLS identification. k =[θ1,θ2,θ3,θ4], then by combining equations (9)-(12), the model parameters can be updated as follows:

[0075]

[0076] In this embodiment, the short-circuit current and short-circuit resistance are estimated based on the particle filter algorithm:

[0077] The basic idea behind the particle filter algorithm for estimating short-circuit current is that the true short-circuit current at different times is consistent with a certain statistical result of particles that follow a certain probability density distribution. The particle filter algorithm uses a series of steps, such as weight calculation and importance sampling, to screen all particles, thereby obtaining the most likely distribution of these particles under the current observation. The basic principles of the particle filter algorithm are as follows:

[0078] (1) Bayesian theory:

[0079] Bayesian theory regards the unknown quantity x as a random variable. In the absence of an observation, it can only be estimated based on the prior distribution p(x) of x. When the observation y is known, p(x) can be modified according to Bayesian theorem to obtain the posterior distribution of x.

[0080]

[0081] Where: p(x|y) represents the posterior probability of x given the known observation y; p(x,y) represents the joint distribution of x and y; p(y) represents the total probability of y; and p(y|x) represents the likelihood probability of y.

[0082] The observation quantity from time 1 to k-1 is denoted as y 1:k-1 , the posterior probability density distribution of x before k-1 time can be expressed as p(x k-1 |y 1:k-1 ). The discrete equivalent circuit model under internal short circuit shown in formulas (3) and (4) is used to estimate the prior probability density distribution p(x) of x at time k. k |y1:k-1 ). For the application in the present invention, the system model is the equivalent circuit model under the internal short circuit in discrete form represented by formulas (3) and (4). When the observation value y at time k is obtained k Then, use it to modify the prior probability density distribution to obtain the posterior probability density p(x k |y 1:k ), its mathematical expression is as follows

[0083]

[0084] Where: x k represents the system state variable at time k; y k represents the observation value at the current time k; x k-1 represents the state variable at the previous moment k-1; y 1:k-1 represents the observation sequence from time 1 to k-1; y 1:k represents the observation sequence from time 1 to time k; p(x|y 1:k-1 ) represents the prior distribution; p(x k ,x k-1 |y 1:k-1 ) represents x k and x k-1 The joint distribution of p(x k |x k-1 ) represents the state transition probability; p(x k-1 |y 1:k-1 ) indicates that given the past observation y 1:k-1 Under the condition of , the posterior distribution of the state at the previous moment; p(x|y 1:k ) represents the posterior probability of the current state given the current and past observations; p(x k |y k ,y 1:k-1 ) represents x k and y 1:k The joint distribution of p(y k |x k ,y 1:k-1 ) represents the observation likelihood function; p(x k |y 1:k-1 ) represents the prior distribution of the kth moment; p(y k |y 1:k-1 ) represents the marginal likelihood of the observation.

[0085] For linear Gaussian systems, the probability density function can be completely represented by the mean and variance. For nonlinear, non-Gaussian models, Monte Carlo simulation is needed to approximate the conditional probability distribution of the unknown quantity x.

[0086] (2) Monte Carlo method:

[0087] The basic idea of the Monte Carlo method is to use some simple random samples to approximate the probability density distribution of the unknown quantity x. The conditional distribution p(x k |y 1:k ) can be approximated by the following random discrete probability distribution sequence

[0088]

[0089] Where: Represents state x k Given an observation sequence y 1:k The estimated probability density function under ; N represents the number of particles, that is, the number of samples used to approximate the posterior distribution; represents the normalized weight of the i-th particle at time k; δ(·) represents the Dirac function; x 0:k Represents the state sequence from the initial moment to the current moment; Represents the state trajectory sample of the i-th particle.

[0090] At time k, x k A random sample of is called a “particle”. For any nonlinear function g(x 0:k ) can be used to obtain its approximate least mean square estimate

[0091]

[0092] In the formula Represents the nonlinear function g(x 0:k ) is an estimated value of . According to the law of large numbers, it can be proved that when N approaches infinity, is the expectation k (x 0:k )} is an unbiased estimate of .

[0093] (3) Importance sampling:

[0094] From the Bayesian estimation formula, we know that the normalization constant p(y k |y 1:k-1 ) is unknown, and the posterior probability density p(x k |y 1:k ) cannot be sampled directly. Therefore, it is necessary to sample another probability density function q(x k |y 1:k ) (called importance sampling function) is used for sampling. To ensure that p(x k |y 1:k ) in all samples of q(x k |y 1:k ) are included in the sampling, and q(x k |y1:k ) is approximated as follows

[0095]

[0096] Combining equations (24) and (25), we have

[0097]

[0098] In the formula Represents the likelihood function of the observed data when a particle state trajectory is given; Prior probability density of particle state trajectory; represents the importance sampling function.

[0099] It is assumed that the sampling at time k does not change the sample set of the past state sequence, that is, after time k arrives, there is no need to recalculate the importance weight of the entire state sequence. 0:k |y 1:k ) can be rewritten as follows

[0100]

[0101] Where: q(x k |x 0:k-1 ,y 1:k ) represents the proposed distribution of the current state given the historical state and observation; q(x 0:k-1 |y 1:k-1 ) represents the proposed distribution of state sequences at past time.

[0102] Combining equations (27) and (28), the weight function is transformed into

[0103]

[0104] Where: represents the likelihood function of the observation given the current state; represents the state transition probability; represents the posterior probability of the past state sequence; the parameters in the denominator come from formula (28), corresponding to the situation of the i-th particle; represents the normalized weight of the i-th particle at the k-1th moment.

[0105] When the state estimation process is optimal, further assumptions can be made

[0106]

[0107] Substituting formula (30) into formula (29), we have

[0108]

[0109] By sampling the importance sampling function, we obtain a Monte Carlo filter estimate of the posterior density function of the state, which also leads to the most basic form of the particle filter algorithm. The particle filter algorithm is a globally convergent nonlinear Gaussian filtering algorithm. Its convergence is theoretically guaranteed, but the theoretical proof is not provided here.

[0110] The specific process of the standard particle filter algorithm is as follows:

[0111] Step 1: Initialization. Generate samples based on prior probability p(x) The weight of each sample is set to 1 / N;

[0112] Step 2: Importance weight calculation. At time k, calculate the importance weight of each sample and normalize it.

[0113] Step 3: Obtain an approximate compliance with p(y by importance sampling k |x k ) distribution of samples

[0114] Step 4: Output the result. Use the mean of the particle set Approximate surrogate function expected value.

[0115] It can be seen that in the lithium-ion battery internal short-circuit model shown in equations (3) and (4), the only unknown quantity we need to solve is the short-circuit current I sc From the above process, we can see that the core principle of particle filtering is to approximate the posterior probability distribution of the system state through a group of random samples (particles) with weights. For the application in the present invention, the system state is the short-circuit current I sc Therefore, using the recursive formulas (23) to (31) of the above particle filter algorithm, the short-circuit current I can be estimated sc Since the voltage across the short-circuit resistor is equal to the battery terminal voltage, the short-circuit resistor is updated based on Ohm's law, that is,

[0116]

[0117] Where R sc Indicates short-circuit resistance, U t,k Indicates the battery terminal voltage.

[0118] The short-circuit resistor R scSubstitute the equivalent circuit model of the short-circuit cell into the model, and perform the next online update of model parameters and short-circuit current calculation based on FFRLS and particle filtering. Repeat the above process until the fluctuation of the short-circuit resistance is less than the predetermined threshold, which means that the short-circuit resistance converges to the true value. At this point, the calculated short-circuit resistance can be output. The formula for calculating the fluctuation amplitude of the short-circuit resistance and the comparison process with the threshold can be expressed as follows:

[0119]

[0120] Where: R sc (iter) represents the short-circuit resistance estimated by the current iteration loop, R sc (iter-1) represents the short-circuit resistance estimated in the last iterative cycle, and Th1 represents the fluctuation threshold. In the subsequent method verification process, the present invention sets it to 0.997.

[0121] In this embodiment, the thermal runaway temperature is estimated based on the extreme learning machine:

[0122] After estimating the short-circuit current and short-circuit resistance of the lithium-ion battery based on the above process, the temperature is estimated based on the extreme learning machine algorithm with the short-circuit current and short-circuit resistance as input and the lithium-ion battery temperature as output, thereby realizing thermal runaway detection and early warning.

[0123] In order to solve the problems existing in traditional neural networks (such as the BP algorithm), existing research has proposed an extreme learning machine algorithm (ELM). This algorithm is a new learning method for a single hidden layer feedforward neural network. The connection weights between the input layer and the hidden layer and the thresholds of the hidden layer neurons are randomly generated and do not need to be adjusted during the training process. It only needs to give the number of hidden layer neurons and, based on the least squares criterion, the optimal solution of the matrix equation output weights can be solved by analytical methods, thereby completing the learning task of the single hidden layer feedforward neural network. The extreme learning machine can overcome the inherent defects of the gradient descent algorithm, such as slow convergence speed and easy to fall into local optimality. Compared with the traditional training algorithm based on gradient descent, the extreme learning machine has the advantages of fast learning speed and strong generalization ability. The following two theories are proposed and proved:

[0124] Theorem 1: Given any N different samples of the data set {x i ,t i}(x i ∈R n ,t i ∈R n ,i=1,2,…,N), where x i represents the data sample, t i Represents x i The corresponding label, R nRepresents n-dimensional real space; infinitely differentiable activation function g in any interval, then for a single hidden layer feedforward neural network with L number of hidden layer neurons, in R n Randomly assign connection weights w in space i and hidden layer threshold b i , so that the output matrix H of the hidden layer of the neural network must be reversible, and ||Hβ-T||=0 must hold, where T represents the label vector composed of the labels of each sample.

[0125] From Theorem 1, we can see that if the number of training samples is the same as the number of neurons in the hidden layer, then for any w i and b i , a single hidden layer feedforward neural network can approximate the training sample with zero error, that is:

[0126]

[0127] Among them, y j =[y 1j ,y 2j ,…,y mj ] T (j=1,2,…,N). Here, y j represents the output of the jth hidden layer neuron; y mj Represents the output of the m×jth hidden layer neuron.

[0128] Theorem 2: Given any N different sample datasets {x i ,t i}(x i ∈R n ,t i ∈R n , i=1,2,…,N), an arbitrarily small error ε>0 and an activation function g that is infinitely differentiable in any interval, then for a single hidden layer feedforward neural network with L number of hidden layer neurons, in R n Randomly assign connection weights w in space i and hidden layer threshold b i , there exists L≤N such that ||H N×L β L×m -T N×m ||<ε must hold. Here, H N×L , β L×m 、T N×m They represent the N×L dimensional neural network hidden layer output matrix, the L×m dimensional weight matrix and the N×m dimensional label matrix respectively, and m represents the number of neurons in the output layer.

[0129] When the number of training samples N is large, in order to reduce the amount of calculation, the number of hidden layers L is usually selected to be smaller than N. According to Theorem 2, the training error of a single hidden layer feedforward neural network approaches an arbitrary ε>0, that is:

[0130]

[0131] For the traditional gradient-based feedforward neural network, the network input weight w i and hidden layer threshold b i It is necessary to continuously iterate and correct the gradient descent method during the training process, which has low computational efficiency. According to the above theorem, it is proved that for the extreme learning machine, as long as the activation function is infinitely differentiable in any interval, the input weight of the network w i and hidden layer threshold b i Random values can be assigned within any interval without iterative adjustment. There is no need for a threshold at the output layer, and ||Hβ-T||=0 is guaranteed to hold.

[0132] The topological structure of the ELM network consists of three parts: input layer, hidden layer and output layer. Figure 2 As shown. Figure 2 In the equation, x is the input value of ELM, β is the output weight of ELM, o is the estimated output value of ELM, and g(w, b, x) is the activation function of ELM. Its main function is to control the activation of input to output, thereby performing functional conversion on input and output. Common activation functions used in ELM networks are binary threshold function, Sigmoid function, Gaussian function, Tanh function and Relu function. In the ELM learning process, different activation functions have different information processing characteristics. Through repeated experiments, the present invention selects Sigmoid function as the activation function of hidden layer neurons, which is defined as follows:

[0133]

[0134] For a set of N training samples A single hidden layer feedforward neural network with L hidden layer neurons and activation function g(x) is expressed as:

[0135]

[0136] Where: g i (·) represents the activation function corresponding to the i-th neuron, x j =[x j1 ,x j2 ,…,x jn ] T , t j =[t j1 ,t j2 ,…,tjm T and w i = [w i1 , w i2 , …, w in T and b i are the input data, the corresponding expected output value, the connection weight between the i-th hidden neuron and the input neuron, and the threshold of the i-th hidden node respectively; β i = [β i1 , β i2 , …, β im T is the connection weight between the i-th hidden neuron and the output neuron, and o j is the actual network output value corresponding to the input x j . n, L, and m are the numbers of neurons in the input layer, hidden layer, and output layer of the network respectively. Since the parameters of the hidden layer {w i , b i} can be randomly generated without tuning during the training process, the goal of ELM is to solve the following compact model to minimize the error between t j and o j :

[0137]

[0138] In the formula, H is the output matrix of the hidden layer, β is the output weight matrix, T is the expected output, is the optimized output weight.

[0139] During the ELM training process, the connection weight w i and the hidden layer threshold b i can be randomly assigned, and the training of its network can actually be transformed into solving the least squares norm solution of the linear equation Hβ = T Then the above formula can be transformed into:

[0140]

[0141] In the formula, w L and b L represent the weight and bias corresponding to the L-th neuron respectively.

[0142] When the number of training samples is the same as the number of neurons in the hidden layer of the network, that is, N = L, the hidden layer output matrix H is a square matrix. By performing an inverse calculation on it, the hidden layer output weight can be obtained as β = H -1 T. However, in practical applications, the number of training samples is often much larger than the number of neurons in the hidden layer of the network, that is, L << N, then the hidden layer output matrix H is not a square matrix.

[0143] ​​​Theorem 3 In a linear matrix Ax=y, the necessary and sufficient condition for Gy to be its least squares norm solution is G=A + , where A represents the coefficient matrix, x represents the parameter matrix, y represents the output matrix, and G represents the solution vector matrix. + is the Moore-Penrose generalized inverse of A. From Theorem 3, we can obtain the least squares norm solution of the linear equation Hβ=T for:

[0144]

[0145] Where H+ is the Moore-Penrose generalized inverse of the hidden layer output matrix H.

[0146] From the above analysis, we can see that before training, the hidden layer node parameters (weights and thresholds) of the ELM algorithm are randomly or artificially given and do not need to be updated. It is only necessary to determine the number of hidden layer neurons and the activation function (infinitely differentiable), and the hidden layer output weights can be calculated during the training process. Therefore, during the ELM training process, only the hidden layer output weights are parameters that need to be adjusted, that is, the least squares norm solution of the linear equation Hβ=T in,

[0147]

[0148] In fact, training the ELM model is equivalent to solving the hidden layer output weights And solve H + There are many methods to solve H, such as iteration method, orthogonal method and orthogonal projection method. The first two methods require continuous iteration and search, which increases the difficulty of solving H. + The complexity of H reduces its computational efficiency. + Common method. When H T When H is a non-singular matrix, H + =(H T H) -1 H T When H T When H is a singular matrix, H + =H T (H T H) -1 , that is, the least squares solution of the linear equation Hβ=T It can be expressed as:

[0149]

[0150] In addition, ELM has the following characteristics:

[0151] ① Minimum training error. The purpose of network training is to achieve the minimum training error as much as possible, but most training algorithms tend to fall into local optimality and it is difficult to obtain the minimum training error. It is the least squares norm solution of the linear equation Hβ=T, so that it can achieve the minimum training error during training:

[0152]

[0153] ② Minimum weight norm. It is the least squares norm solution of the linear equation Hβ=T, where the smaller the weight norm, the better the generalization performance of the training model:

[0154]

[0155] ③The uniqueness of the solution. It is the least squares norm solution of the linear equation Hβ=T, and it is the only solution.

[0156] Internal short circuit is the main cause of thermal runaway of lithium-ion batteries, and it is also the common link of thermal runaway of lithium-ion batteries caused by external abuse conditions such as mechanical abuse, thermal abuse and electrical abuse. Short-circuit current and short-circuit resistance are the core basis and scientific indicators for measuring the severity and evolution stage of internal short circuits, that is, the larger the short-circuit current and the smaller the short-circuit resistance, the more serious the internal short circuit. For thermal runaway of lithium-ion batteries, temperature is one of the most core parameters to characterize its characteristics. According to the above analysis, the present invention uses short-circuit resistance and short-circuit current as inputs and battery temperature as output, and estimates the thermal runaway temperature based on the extreme learning machine algorithm. When the estimated temperature exceeds the thermal runaway characteristic temperature, a thermal runaway alarm message is issued, thereby ensuring the safe and stable operation of the battery system. In other words, this article establishes the relationship between internal short circuit and thermal runaway by establishing the relationship between short-circuit current and short-circuit resistance (characterizing the severity of internal short circuit) and temperature (characterizing thermal runaway).

[0157] From the above process, we can see that the method proposed by the present invention estimates the short-circuit current and short-circuit resistance corresponding to the internal short-circuit fault based on the voltage and current data of the energy storage lithium-ion battery. Subsequently, based on the estimated short-circuit current and short-circuit resistance, the battery thermal runaway temperature is further predicted and compared with the characteristic temperature to achieve thermal runaway detection. Compared with existing energy storage lithium-ion battery thermal runaway detection methods, the method proposed by the present invention has the following two outstanding advantages:

[0158] (1) After a serious thermal runaway accident occurs in a lithium-ion battery, the safety valve opens and combustible gas is ejected, accompanied by smoke and fire. Therefore, the existing thermal runaway detection method for energy storage lithium-ion battery systems usually determines whether a thermal runaway accident has occurred by coupling multiple sensors such as gas sensors, smoke sensors and flame sensors after the thermal runaway accident occurs. This means that the existing method requires the installation of additional sensor equipment (i.e., the gas sensors, smoke sensors and flame sensors mentioned above). These sensor equipment usually need to be specially customized and are relatively expensive, which means that the existing thermal runaway detection method will undoubtedly significantly increase the hardware cost and equipment complexity. In contrast, the method proposed in the present invention only requires the input of the voltage and current parameters of the battery, which means that only voltage and current sensors are required (already equipped in the energy storage lithium-ion battery system), and there is no need to install additional advanced sensing technology. This means that the method proposed in the present invention requires lower costs, which is conducive to reducing the installation and operation and maintenance costs of the energy storage system.

[0159] (2) As mentioned above, the existing thermal runaway detection method for lithium-ion batteries usually implements thermal runaway alarms by detecting combustible gases, smoke, and flames after the accident occurs, which is a typical post-alarm method. However, when obvious thermal runaway characteristics such as combustible gases, smoke, and flames are detected, it means that the thermal runaway accident of the lithium-ion battery has become extremely serious and is in an irreversible stage. In other words, when thermal runaway is detected according to the existing method, the lithium-ion battery fire and explosion accident is very close, and there is no time and space to take disposal measures. This means that the existing thermal runaway detection method is difficult to reserve enough time for accident disposal. On the contrary, the method proposed in the present invention can achieve early warning of thermal runaway by predicting the thermal runaway temperature, which is a typical pre-warning method. An alarm message can be sent to the battery management system or operation and maintenance personnel before the thermal runaway becomes extremely serious. In other words, the method proposed in the present invention can warn of thermal runaway in advance, reserve enough time for accident disposal, and provide a valuable safety margin.

[0160] Example 2

[0161] The present invention also provides a thermal runaway temperature estimation system for energy storage lithium-ion batteries based on internal short circuit quantitative diagnosis, the system is used to implement the method described in Example 1, the system includes: an identification module, a first estimation module, and a second estimation module;

[0162] Identification module, used to establish the internal short-circuit equivalent circuit model of lithium-ion batteries and identify the model parameters online;

[0163] The first estimation module is used to estimate the short-circuit current and short-circuit resistance using a particle filter algorithm based on the internal short-circuit equivalent circuit model of the lithium-ion battery after the model parameters are updated;

[0164] The second estimation module is used to estimate the thermal runaway temperature based on the extreme learning machine algorithm, using the short-circuit resistance and short-circuit current as input and the battery temperature as output. When the estimated temperature exceeds the thermal runaway characteristic temperature, a thermal runaway alarm message is issued.

[0165] In this embodiment, the process of establishing the lithium-ion battery internal short-circuit equivalent circuit model includes:

[0166] U p,k =exp(-Δt / (R p C p ))U p,k-1 +(1-exp(-Δt / (R p C p )))R p (I sc,k-1 +I L,k-1 );

[0167] U t,k =U oc,k -U p,k -(I sc,k +I L,k )R0;

[0168] Where Δt represents the sampling time, subscript k represents the kth moment, R0 represents the battery internal resistance, R p and C p Represent polarization resistance and polarization capacitance respectively, U p,k , I sc,k , I L,k 、U t,k 、U oc,k They represent the polarization voltage, short-circuit current, external load current, terminal voltage and open-circuit voltage at the kth moment respectively.

[0169] In this embodiment, the process of online identification of model parameters includes:

[0170] Based on FFRLS identification, the parameter vector θ is obtained k =[θ1,θ2,θ3,θ4], the model parameters are updated as follows:

[0171]

[0172] Where R sc Indicates the internal short-circuit resistance, U oc is the battery open circuit voltage, R0 represents the battery internal resistance, R p represents the polarization resistance, α is the intermediate process parameter of the equivalent circuit model, and θ1, θ2, θ3, and θ4 are the intermediate process parameters of parameter identification.

[0173] In this embodiment, the process of estimating the thermal runaway temperature based on the extreme learning machine algorithm includes:

[0174] With short-circuit current and short-circuit resistance as input and lithium-ion battery temperature as output, based on the extreme learning machine algorithm, by establishing the relationship between short-circuit current and short-circuit resistance, which characterize the severity of internal short circuit, and temperature, which characterizes thermal runaway, the relationship between internal short circuit and thermal runaway is established, and the thermal runaway temperature is estimated.

[0175] Example 3

[0176] In order to verify the effectiveness and feasibility of the proposed method, the present invention built an experimental platform under adiabatic boundary conditions. The adiabatic experimental environment consists of an accelerating rate calorimeter (ARC) and a battery charge and discharge tester. The ARC model used is the ARC-EV of the British THT company; the battery charge and discharge equipment is the 60V / 100A charge and discharge tester of China Xinwei Company. The battery is placed in the center of the ARC calorimetric chamber, and the thermocouple is fixed to the center of the battery side surface using thermocouple glue (satlon D-3), and the connecting wire is fixed using aluminum foil tape.

[0177] The present invention selected 44 lithium iron phosphate battery cells for testing to verify the effectiveness of the technology. The cell specifications are shown in Table 1. The experiment adopted a constant current charge-dynamic discharge composite working condition to carry out a cycle test on the battery. The present invention simulates an internal short circuit fault by connecting an external resistor in parallel at both ends of the battery. This method can not only better control the triggering time and position of the internal short circuit, but also simulate the evolution process of the internal short circuit, and has good controllability and repeatability. Specifically, during the experiment, by continuously reducing the parallel resistance, the internal short circuit is gradually worsened, and eventually induces thermal runaway of the lithium-ion battery. During this process, the self-heating temperature and the pressure relief starting temperature of the battery are measured and recorded.

[0178] Table 1 Specifications of lithium-ion batteries

[0179]

[0180] During the validation process, 50% of the data from all cells were randomly selected as the training set, and the remaining 50% were used as the test set. The self-heating temperature estimation results of 22 battery cells in the test set based on the proposed method are shown in Figure 2. Figure 3 As shown, the pressure relief starting temperature is Figure 4 Table 2 lists the relative errors of the self-heating temperature and the pressure relief starting temperature.

[0181] Table 2 Relative errors of self-heating temperature and pressure relief starting temperature estimation

[0182]

[0183] from Figure 3 and Figure 4 It can be seen that for each battery cell, the estimated self-heating temperature and pressure relief onset temperature agree well with the actual values. Table 2 shows that the average and maximum relative errors of the self-heating temperature estimation results for lithium iron phosphate batteries are less than 2% and 9%, respectively, while the average and maximum relative errors of the pressure relief onset temperature estimation results are less than 1% and 5%, respectively. This analysis demonstrates that the battery thermal runaway temperatures estimated using the proposed method agree well with the measured values, and that the estimated errors are within a reasonable range, demonstrating the effectiveness and feasibility of the proposed method.

[0184] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for estimating thermal runaway temperature of energy storage lithium-ion batteries based on quantitative diagnosis of internal short circuits, characterized in that: The method comprises: Establish an equivalent circuit model for the internal short circuit of lithium-ion batteries and identify the model parameters online; Based on the internal short-circuit equivalent circuit model of the lithium-ion battery after the model parameters are updated, the particle filter algorithm is used to estimate the short-circuit current and short-circuit resistance; With short-circuit resistance and short-circuit current as input and battery temperature as output, the thermal runaway temperature is estimated based on the extreme learning machine algorithm. When the estimated temperature exceeds the thermal runaway characteristic temperature, a thermal runaway alarm message is issued.

2. The method according to claim 1, characterized in that Methods for establishing an equivalent circuit model of an internal short circuit of a lithium-ion battery include: U p,k =exp(-Δt / (R p C p ))U p,k-1 +(1-exp(-Δt / (R p C p )))R p (I sc,k-1 +I L,k-1 ); IN t,k =U oc,k -IN p,k -(AND sc,k +I L,k )R0; Where Δt represents the sampling time, subscript k represents the kth moment, R0 represents the battery internal resistance, R p and C p Represent polarization resistance and polarization capacitance respectively, U p,k , I sc,k , I L,k 、U t,k 、U oc,k They represent the polarization voltage, short-circuit current, external load current, terminal voltage and open-circuit voltage at the kth moment respectively.

3. The method according to claim 2, characterized in that Methods for online identification of model parameters include: Based on FFRLS identification, the parameter vector θ is obtained k =[θ1,θ2,θ3,θ4], the model parameters are updated as follows: Where R sc Indicates the internal short-circuit resistance, U oc is the battery open circuit voltage, R0 represents the battery internal resistance, R p represents the polarization resistance, α is the intermediate process parameter of the equivalent circuit model, and θ1, θ2, θ3, and θ4 are the intermediate process parameters of parameter identification.

4. The method according to claim 1, wherein The methods for estimating thermal runaway temperature based on the extreme learning machine algorithm include: With short-circuit current and short-circuit resistance as input and lithium-ion battery temperature as output, based on the extreme learning machine algorithm, by establishing the relationship between short-circuit current and short-circuit resistance, which characterize the severity of internal short circuit, and temperature, which characterizes thermal runaway, the relationship between internal short circuit and thermal runaway is established, and the thermal runaway temperature is estimated.

5. A thermal runaway temperature estimation system for energy storage lithium-ion batteries based on internal short circuit quantitative diagnosis, the system being used to implement the method according to any one of claims 1 to 4, characterized in that: The system includes: an identification module, a first estimation module, and a second estimation module; The identification module is used to establish an internal short-circuit equivalent circuit model of the lithium-ion battery and identify model parameters online; The first estimation module is used to estimate the short-circuit current and short-circuit resistance using a particle filter algorithm based on the lithium-ion battery internal short-circuit equivalent circuit model after the model parameters are updated; The second estimation module is used to estimate the thermal runaway temperature based on the extreme learning machine algorithm using the short-circuit resistance and short-circuit current as input and the battery temperature as output, and to issue a thermal runaway alarm message when the estimated temperature exceeds the thermal runaway characteristic temperature.

6. The system according to claim 5, characterized in that The process of establishing an equivalent circuit model for an internal short circuit of a lithium-ion battery includes: U p,k =exp(-Δt / (R p C p ))U p,k-1 +(1-exp(-Δt / (R p C p )))R p (I sc,k-1 +I L,k-1 ); IN t,k =U oc,k -IN p,k -(AND sc,k +I L,k )R0; Where Δt represents the sampling time, subscript k represents the kth moment, R0 represents the battery internal resistance, R p and C p Represent polarization resistance and polarization capacitance respectively, U p,k , I sc,k , I L,k 、U t,k 、U oc,k They represent the polarization voltage, short-circuit current, external load current, terminal voltage and open-circuit voltage at the kth moment respectively.

7. The system according to claim 6, characterized in that The process of online identification of model parameters includes: Based on FFRLS identification, the parameter vector θ is obtained k =[θ1,θ2,θ3,θ4], the model parameters are updated as follows: Where R sc Indicates the internal short-circuit resistance, U oc is the battery open circuit voltage, R0 represents the battery internal resistance, R p represents the polarization resistance, α is the intermediate process parameter of the equivalent circuit model, and θ1, θ2, θ3, and θ4 are the intermediate process parameters of parameter identification.

8. The method according to claim 5, characterized in that The process of thermal runaway temperature estimation based on the extreme learning machine algorithm includes: With short-circuit current and short-circuit resistance as input and lithium-ion battery temperature as output, based on the extreme learning machine algorithm, by establishing the relationship between short-circuit current and short-circuit resistance, which characterize the severity of internal short circuit, and temperature, which characterizes thermal runaway, the relationship between internal short circuit and thermal runaway is established, and the thermal runaway temperature is estimated.

Images