A battery health status prediction method based on DNN-UKF model

Abstract

The present invention discloses a method for predicting the state of health of a battery based on a DNN-UKF model, the method comprising: constructing an equivalent circuit model according to battery characteristics, taking the battery internal resistance state quantity as a state variable and the battery output terminal voltage as an output variable, and constructing a state equation and an observation equation; based on the battery state equation and the observation equation, using an unscented Kalman filter algorithm to calculate a first estimated value of the battery internal resistance state quantity; correcting the first estimated value using an estimated error to obtain a second estimated value; wherein the estimated error is obtained by estimating through a deep neural network; and calculating the SOH value of the battery using the second estimated value. The present invention solves the problem of decreased prediction accuracy when environmental noise changes or the noise is large, and can effectively extend the battery life and improve system safety.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy storage batteries, and in particular relates to a battery health status prediction method based on a DNN-UKF model. Background Art

[0002] As an energy carrier with a series of advantages such as high energy density, long cycle life, low self-discharge rate, and low environmental pollution, lithium-ion batteries are widely used in electric vehicles and other fields. Due to the degradation of the electrochemical components of lithium-ion batteries, their performance will decrease with the operating time, and electrolyte leakage and micro-short circuits may occur, leading to battery failure. Therefore, accurate assessment of the state of health (SOH) of lithium-ion batteries is the key to improving battery reliability and safety.

[0003] The health status of lithium-ion batteries is a defined quantity that represents the state of lithium-ion batteries, rather than a specific physical quantity, so it cannot be directly measured. At present, its estimation methods mainly include mechanical models, equivalent circuit models, data-driven or a fusion of several methods. The data-driven rule extracts aging characteristics from a large amount of charging and discharging data of lithium-ion batteries, and uses intelligent algorithms to establish a relationship model between battery aging characteristics and SOH. Among them, the Unscented Kalman Filter (UKF) algorithm is widely used, but due to the presence of noise, this method still has certain errors. Summary of the invention

[0004] The purpose of the present invention is to provide a battery health state prediction method based on the DNN-UKF model to further reduce the error of the SOH estimation process and improve the prediction accuracy.

[0005] The present invention adopts the following technical solutions:

[0006] A battery health status prediction method based on a DNN-UKF model includes the following steps:

[0007] Construct an equivalent circuit model based on the battery characteristics, take the battery internal resistance state quantity as the state variable, and the battery output terminal voltage as the output variable, and construct the state equation and observation equation;

[0008] Based on the state equation and observation equation of the battery, an unscented Kalman filter algorithm is used to calculate a first estimated value of the battery internal resistance state quantity;

[0009] Correcting the first estimated value using the estimated error to obtain a second estimated value; wherein the estimated error is obtained by estimating through a deep neural network;

[0010] The SOH value of the battery is calculated using the second estimated value.

[0011] Furthermore, the equivalent circuit model is a second-order Thevenin equivalent circuit model.

[0012] Furthermore, the state equation is expressed as:

[0013] R0(k+1)=R0(k)+w k ,

[0014] Among them, R0(k) is the battery internal resistance state at time k, R0(k+1) is the battery internal resistance state at time k+1, and w k is the system white noise for estimating the battery internal resistance at time k.

[0015] Furthermore, the observation equation is expressed as:

[0016] U(k)=U oc (k)-U c1 (k)-U c2 (k)-R0I(k)+v(k),

[0017] Among them, U c1 (k) and U c2 (k) is the voltage of the two RC loops in the second-order Thevenin equivalent circuit model at time k, R0 is the internal resistance of the battery, I(k) is the charge and discharge current at time k, U(k) is the output voltage at time k, U oc (k) is the battery voltage at time k, and v(k) is the measurement noise of the battery output voltage at time k.

[0018] Further, calculating the first estimated value of the battery internal resistance state quantity using the unscented Kalman filter algorithm includes:

[0019] The battery internal resistance state quantity at time k is taken as input, and the state equation and observation equation are used to determine the first estimated value of the battery internal resistance state quantity at time k+1 and the Kalman gain.

[0020] Further, using the estimation error to correct the first estimation value to obtain the second estimation value includes:

[0021] After each estimation of the battery internal resistance state quantity, the current at that moment, the output voltage, the first estimated value of the battery internal resistance state quantity at time k+1 and the Kalman gain are used as input quantities and input into the trained deep neural network model, with the estimated error as the output. The estimated error is removed from the first estimated value to obtain the second estimated value at time k+1.

[0022] Furthermore, the loss function of the deep neural network is the mean square error function, the neuron activation function of the hidden layer is the Sigmoid function and the Relu function, and the neuron activation function of the output layer is the Linear function.

[0023] Furthermore, the SOH value of the battery is calculated using the second estimated value and is expressed as:

[0024]

[0025] Among them, SOH is the battery health state, C t is the current maximum available capacity of the battery, C new is the nominal capacity, R t is the second estimated value, R end is the internal resistance of the battery at the end of its service life, R new is the nominal resistance.

[0026] An electronic device, comprising:

[0027] at least one processor;

[0028] and, a memory communicatively coupled to the at least one processor;

[0029] Among them, the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor so that at least one processor can execute any of the above battery health status prediction methods based on the DNN-UKF model.

[0030] The beneficial effects of the present invention are as follows: the present invention first obtains a state equation and an observation equation based on an equivalent circuit model, then uses a UKF algorithm to make a preliminary estimate of the battery internal resistance state quantity, and uses a deep neural network to correct a first estimate of the battery internal resistance state quantity, and excludes the estimation error from the first estimate to obtain a more accurate second estimate, i.e., a battery internal resistance prediction result, and finally uses the battery internal resistance prediction result to calculate the SOH value of the battery; in view of the problem that the estimation accuracy of the UKF algorithm decreases due to the preset noise being a constant value, the present invention uses a deep neural network to spontaneously adjust the weights of error transmission for different environmental noises, thereby reducing the impact of noise and improving the accuracy of the SOH prediction results. The present invention has broad application prospects in battery management systems and can effectively extend the battery life and improve system safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The present invention is a flow chart of the method.

[0032] Figure 2 It is a schematic diagram of the second-order Thevenin equivalent circuit model in the present invention;

[0033] Figure 3 The figure is a flow chart of battery internal resistance estimation based on UKF algorithm in the present invention;

[0034] Figure 4 This is a structural diagram of a deep neural network in the present invention;

[0035] Figure 5 This is a flowchart of battery internal resistance compensation optimization in the present invention. DETAILED DESCRIPTION

[0036] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Although the Unscented Kalman Filter (UKF) algorithm is widely used in SOH prediction, the algorithm presets the noise as a constant value and fails to improve the SOH prediction results for different noise environments. When the environmental noise changes or the noise is large, the prediction accuracy will decrease.

[0038] The battery charging and discharging process is accompanied by the insertion and deinsertion of lithium ions. The chemical reactions in the battery cause the battery capacity to decay and performance to decline. SOH characterizes the degree of battery aging and is defined as the ratio of the maximum available capacity to the rated capacity. It is usually expressed by the trend of internal resistance and capacity change, expressed as:

[0039]

[0040] Among them, SOH is the battery health state, C t is the current maximum available capacity of the battery, C new is the nominal capacity, R t is the actual internal resistance of the battery in the current cycle, i.e., the second estimated value in the present invention, R end is the internal resistance of the battery at the end of its service life, R new is the nominal resistance.

[0041] Since the battery internal resistance is closely related to the battery state, the battery SOH can be characterized by the battery internal resistance. Therefore, real-time estimation of the battery internal resistance can achieve real-time estimation of the battery SOH.

[0042] The present invention is specifically:

[0043] A battery health status prediction method based on DNN-UKF model, such as Figure 1 As shown, the following steps are included:

[0044] S110, construct an equivalent circuit model according to the battery characteristics, take the battery internal resistance state quantity as the state variable, and the battery output terminal voltage as the output variable, and construct the state equation and observation equation, specifically:

[0045] like Figure 2 As shown, the equivalent circuit model is a second-order Thevenin equivalent circuit model. The state equation and observation equation of the lithium iron phosphate single battery SOH characterized by the battery internal resistance are as follows:

[0046] The state equation and observation equation are expressed as:

[0047]

[0048] Among them, R0(k) is the battery internal resistance state at time k, R0(k+1) is the battery internal resistance state at time k+1, and w k To estimate the system white noise of the battery internal resistance at time k, U c1 (k) and U c2 (k) is the voltage of the two RC loops in the second-order Thevenin equivalent circuit model at time k, R0 is the internal resistance of the battery, I(k) is the charge and discharge current at time k, U(k) is the output voltage at time k, U oc (k) is the battery voltage at time k, v(k) is the measurement noise of the battery output voltage at time k, Q0 is the rated capacity of the battery, R0, R1, R2, C1, and C2 represent the battery internal resistance and the resistance and capacitance of the two circuits, respectively.

[0049] Based on the second-order Thevenin equivalent circuit model, the parameter values ​​of each component in the circuit are obtained through mixed power pulse characteristic test.

[0050] The second-order Thevenin equivalent circuit model can simulate the dynamic and static characteristics of the lithium battery system more realistically and quickly, and describe the response characteristics of the battery more accurately.

[0051] The state equation and observation equation together constitute the system model, which can predict and correct the system state, thereby achieving effective control and filtering.

[0052] S120, based on the state equation and observation equation of the battery, using an unscented Kalman filter algorithm to calculate a first estimated value of the battery internal resistance state quantity.

[0053] The battery internal resistance state quantity at time k is taken as input, and the state equation and observation equation are used to determine the first estimated value of the battery internal resistance state quantity at time k+1 and the Kalman gain.

[0054] The unscented Kalman filter algorithm uses unscented transformation to deal with the error caused by nonlinear transmission and the filter divergence problem caused by the failure of linearization assumption, and then uses the current observation value to make continuous corrections to gradually approximate it, making it closer and closer to the true value of the system.

[0055] The estimated value of the battery internal resistance state is calculated using the unscented Kalman filter algorithm. Figure 3 As shown, specifically:

[0056] S1201, initializing the state equation and covariance matrix;

[0057] S1202, input the battery internal resistance state at time k;

[0058] S1203, perform an untraceable transformation on the battery internal resistance state quantity at time k, determine a Sigma point near it, obtain 2n+1 Sigma points and weights of the battery internal resistance state quantity at time k, and then express these Sigma points as a probability density function similar to the Gaussian density state.

[0059] Sigma point is expressed as:

[0060]

[0061] The weight of the Sigma point is expressed as:

[0062]

[0063] Among them, n represents the dimension of the state, X represents the sampling point, i represents the i-th sampling point, P represents the covariance of the state variable, ω m represents the mean weight, ω c Represents the covariance weight. Parameter λ=α 2 (n+κ)-n represents a scaling parameter used to reduce the estimation error of the nonlinear system; parameter β≥0, which is determined by the distribution method of the Sigma point set, and the present invention takes β=2; κ is an adjustment factor, and the present invention takes a value of 1, which can ensure that the (n+λ)P matrix is ​​semi-positive definite; α is a scaling factor, and the present invention takes a value of 1.

[0064] S1204, substituting the 2n+1 Sigma points of the battery internal resistance state quantity at time k into the state equation in formula (2), and obtaining the prediction of the 2n+1 Sigma points of the battery internal resistance state quantity at time k at time k+1;

[0065] S1205, according to S1204, 2n+1 prediction results are obtained, and the prediction mean and covariance matrix of the system state quantity at time k+1 are calculated;

[0066] S1206, according to the predicted mean and covariance matrix of S1205, untraceable transformation is used again to generate new 2n+1 Sigma point sets, that is, 2n+1 Sigma point sets of the battery internal resistance state quantity at time k+1;

[0067] S1207, substitute the Sigma point set of S126 into the observation equation of formula (2) to obtain the predicted observation value at time k+1;

[0068] S1208, calculating the predicted mean and covariance matrix of the system observation at time k+1 according to the 2n+1 prediction results obtained in S1207;

[0069] S1209, calculating the Kalman gain matrix at time k+1;

[0070] S1210, update the state quantity and covariance matrix of the system.

[0071] S130, using the estimation error to correct the first estimation value to obtain a second estimation value; wherein the estimation error is obtained by estimating through a deep neural network.

[0072] In the process of estimating the internal resistance of the battery using the UKF algorithm, errors may occur in the result of the internal resistance estimation due to the influence of various external factors. To compensate for this error, the present invention introduces a deep neural network to compensate and optimize the UKF estimation value.

[0073] Compared with ordinary shallow neural networks, DNN has a more complex structure and has a strong ability to extract local abstract high-level features. It also prevents model overfitting through Dropout and L1 and L2 regularization, speeds up model training, and improves the robustness of the description of the nonlinear relationship between input variables and output variables. Figure 4 As shown in Figure 1, DNN includes input layer, hidden layer, and output layer. Each layer is fully connected, and neurons in the same layer are not connected and are independent of each other. The battery internal resistance compensation optimization model based on DNN is shown in Figure 1. Figure 4 shown.

[0074] In the present invention, after each estimation of the battery internal resistance state quantity, the current at that moment, the output voltage, the first estimated value of the battery internal resistance state quantity at time k+1 and the Kalman gain are used as input quantities and input into the trained deep neural network model, with the estimated error as the output, and the estimated error is removed from the first estimated value to obtain the second estimated value at time k+1.

[0075] The steps of DNN training are as follows Figure 5 As shown, including:

[0076] S131, network initialization. The estimated values ​​of current, output voltage, internal resistance state quantity and Kalman gain are used as the input of DNN, and the estimated error of the internal resistance state quantity of the battery is used as the output Y of DNN. Determine the total number of layers L of the neural network and the number of nodes z in each layer l (l=1,2,…,L); Randomly initialize the weight matrix W between each hidden layer and the output layer l and the offset vector b l (l=1,2,…,L), determine the learning rate η, iteration threshold ε and neuron activation function.

[0077] S132, calculate the output H of each hidden layer and output layer l .

[0078] H l =f(zl )=f(W l H l-1 +b l ),l=1,2,3,…,L (5)

[0079] In the formula, f is the neuron activation function, the neuron activation function of the hidden layer is the Sigmoid function and the Relu function, and the neuron activation function of the output layer is the Linear function.

[0080] Among them, the Sigmoid function is continuous and its gradient is smooth, which helps to avoid jumps in output values ​​and makes the model training more stable. The derivative calculation of the Sigmoid function is simple, which helps the gradient calculation in the back propagation process. At the same time, the Sigmoid function introduces nonlinear characteristics, which enables the neural network to learn and simulate complex nonlinear relationships.

[0081] As a neuron activation function, the ReLU function has the advantages of simplicity, high efficiency and nonlinear characteristics.

[0082] The linear function has the advantages of simple calculation and constant gradient.

[0083] S133, select the mean square error function shown in formula (3) as the loss function, and then calculate the gradient δ of the output layer according to the loss function J L .

[0084]

[0085] As one of the commonly used loss functions, mean square error is widely used to evaluate the prediction performance of the model. By calculating the mean square error between the model's predicted value and the actual value, we can intuitively understand the model's prediction accuracy and then optimize and improve the model.

[0086] S134, calculate the gradient δ of each hidden layer l , the calculation formula is:

[0087]

[0088] S135, adjust the weight matrix W of each hidden layer and output layer l and the offset vector b l , whose expression is:

[0089] W l =W l -ηδ l (H l-1 ) T (9)

[0090] b l =b l -ηδl ,l=2,3,…,L (10)

[0091] S136, judging whether the algorithm iteration is completed according to the threshold ε, if not completed, returning to S132.

[0092] S140: Calculate the SOH value of the battery using the second estimated value.

[0093] The present invention proposes a battery health status prediction method based on the DNN-UKF model. The UKF algorithm does not need to perform Taylor formula expansion on the system function, and omits the linearization processing step of high-order terms. The unscented transform (UT) is used in the estimation process to solve the nonlinear problem, avoiding the estimation error of the entire system caused by ignoring the high-order terms of the Taylor series when linearizing the nonlinear system. However, since the UKF algorithm presets the noise as a constant value, it fails to improve the SOH estimation results for different noise environments, and the estimation accuracy will decrease when the environmental noise changes or the noise is large. Neural network related algorithms have a good prediction effect on nonlinear systems, and can spontaneously adjust the weights of error transfer according to different environmental noises, but they are prone to fall into local optimal solutions during the solution process. A deep neural network is used to predict the error of UKF filtering under different noises, and the estimation error is excluded from the initial estimation value to compensate for the SOH estimation result. This ensures that the algorithm has strong robustness while effectively avoiding the problem of falling into the local optimal solution, improves the accuracy of the estimation, realizes real-time prediction of the battery health status, provides reliable decision support for the battery management system, improves the safety and reliability of the system, and effectively copes with the complex working environment and nonlinear characteristics of the battery. It has broad application prospects and important practical significance.

[0094] In summary, the present invention first obtains the state equation and observation equation based on the second-order Thevenin equivalent model, then uses the UKF algorithm to make a preliminary estimate of the battery internal resistance state quantity, and uses a deep neural network to correct the estimated value of the battery internal resistance state quantity, and excludes the estimation error from the estimated value of the battery internal resistance state quantity to obtain a more accurate battery internal resistance prediction result, and finally uses the battery internal resistance prediction result to calculate the battery's SOH value. The present invention improves the accuracy of the SOH prediction result, realizes the real-time prediction of the battery health status, provides reliable decision support for the battery management system, improves the safety and reliability of the system, can effectively cope with the complex working environment and nonlinear characteristics of the battery, and has broad application prospects and important practical significance.

Claims

1. A battery health status prediction method based on DNN-UKF model, characterized in that: The following steps are involved: An equivalent circuit model is constructed based on the battery characteristics, with the battery internal resistance state quantity as the state variable and the battery output terminal voltage as the output variable, to construct the state equation and observation equation; Based on the state equation and observation equation of the battery, a first estimated value of the battery internal resistance state quantity is calculated using an unscented Kalman filter algorithm; Correcting the first estimated value using an estimated error to obtain a second estimated value; wherein the estimated error is obtained by estimating a deep neural network; The SOH value of the battery is calculated using the second estimated value.

2. According to claim 1, a battery health status prediction method based on DNN-UKF model is characterized in that: The equivalent circuit model is a second-order Thevenin equivalent circuit model.

3. According to claim 2, a battery health status prediction method based on DNN-UKF model is characterized in that: The state equation is expressed as: R0(k+1)=R0(k)+w k , Among them, R0(k) is the battery internal resistance state at time k, R0(k+1) is the battery internal resistance state at time k+1, and w k is the system white noise for estimating the battery internal resistance at time k.

4. A battery health status prediction method based on DNN-UKF model according to claim 3, characterized in that the observation The equation is expressed as: U(k)=U oc (k)-U c1 (k)-U c2 (k)-R0I(k)+v(k), Among them, U c1 (k) and U c2 (k) is the voltage of the two RC loops in the second-order Thevenin equivalent circuit model at time k, R0 is the internal resistance of the battery, I(k) is the charge and discharge current at time k, U(k) is the output voltage at time k, U oc (k) is the battery voltage at time k, and v(k) is the measurement noise of the battery output voltage at time k.

5. The method for predicting battery health status based on the DNN-UKF model according to claim 4, characterized in that: Calculating the first estimated value of the battery internal resistance state quantity using the unscented Kalman filter algorithm includes: The battery internal resistance state quantity at time k is taken as input, and the state equation and observation equation are used to determine the first estimated value of the battery internal resistance state quantity and the Kalman gain at time k+1.

6. The method for predicting battery health status based on the DNN-UKF model according to claim 5, characterized in that: Correcting the first estimated value by using the estimated error to obtain a second estimated value includes: After each estimation of the battery internal resistance state quantity, the current at that moment, the output voltage, the first estimated value of the battery internal resistance state quantity at time k+1 and the Kalman gain are used as input quantities and input into the trained deep neural network model, with the estimated error as the output. The estimated error is removed from the first estimated value to obtain the second estimated value at time k+1.

7. The method for predicting battery health status based on the DNN-UKF model according to claim 6, characterized in that: The loss function of the deep neural network is the mean square error function, the neuron activation function of the hidden layer is the Sigmoid function and the Relu function, and the neuron activation function of the output layer is the Linear function.

8. A battery health status prediction method based on DNN-UKF model according to claim 1 or 7, characterized in that: The SOH value of the battery calculated using the second estimated value is expressed as: Among them, SOH is the battery health state, C t is the current maximum available capacity of the battery, C new is the nominal capacity, R t is the second estimated value, R end is the internal resistance of the battery at the end of its service life, R new is the nominal resistance.

9. An electronic device, characterized in that: include: at least one processor; and, a memory communicatively coupled to at least one of the processors; Wherein, the memory stores instructions that can be executed by at least one of the processors, and the instructions are executed by at least one of the processors so that at least one of the processors can execute a battery health status prediction method based on a DNN-UKF model as described in any one of claims 1 to 8.

Images