Method for constructing DCIR model of energy storage lithium ion battery based on neural network
By constructing a DCIR model based on a fully connected neural network, the problem that traditional models cannot describe the nonlinear changes in DCIR of lithium-ion batteries is solved, and higher accuracy DCIR prediction is achieved, especially with an error of less than 1% under high temperature conditions.
Patent Information
- Application Number
- CN202411198511.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-29
AI Technical Summary
Traditional DCIR models are difficult to accurately describe the nonlinear trend of DCIR in lithium-ion batteries during the initial cycling phase, which is characterized by a decrease followed by an increase, resulting in large prediction errors.
A fully connected neural network model was adopted, and the model parameters were optimized by combining test data and the activation function Leaky ReLU. A neural network-based DCIR model was constructed to capture the DCIR changes of the battery under different temperatures and SOCs.
It improves the accuracy of DCIR prediction, accurately captures the DCIR change trend of the battery in the early stage of operation, and reduces prediction error, especially under high temperature conditions, the error is less than 1% of the traditional model.
Smart Images

Figure CN119044779A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of lithium-ion batteries, and in particular to a method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network. Background Art
[0002] For most power lithium-ion batteries, their relative DCIR under working conditions can usually be described by the traditional semi-empirical DCIR model. The relative DCIR that can be predicted by the traditional DCIR model will not be less than 100%. However, in fact, when the battery is in the early stage of the cycle (several hundred cycles), its DCIR will first decrease and then increase. For example, the DCIR of a battery at high temperature will first decrease from 100% to 95%, and then increase to more than 100%. Since there is a turning point in this process, it is difficult to accurately describe this nonlinear change trend using the traditional DCIR model fitting. Summary of the invention
[0003] Based on this, the present application provides a method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network to solve the problem that the traditional DCIR model is difficult to accurately describe the phenomenon that the relative DCIR first decreases and then increases.
[0004] The present application provides a method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network, comprising:
[0005] Collect test data of lithium-ion batteries at different temperatures, different SOCs and different cycle conditions through testing;
[0006] Obtaining a relative DCIR value of the lithium-ion battery under current test conditions based on the test data;
[0007] Fitting the test data using a fully connected neural network model to obtain output parameters;
[0008] By using MSE as the loss function and adjusting the weight parameters and bias parameters of the fully connected neural network model, the difference between the output parameter and the relative DCIR value is minimized, and finally the optimal DCIR model is obtained.
[0009] The test data includes battery storage days, cycle number, battery capacity, and battery real-time temperature.
[0010] The acquisition of the relative DCIR value includes:
[0011] When the battery is at the current temperature and 50% SOC, the battery is pulse-discharged with a current I of 200A for a duration of 10s, and the voltage V of the battery before discharge is recorded. i前 And the voltage after discharge Vi后 , repeat the above discharge operation for multiple times, and take the average value of these multiple times as the current DCIR of the battery;
[0012] The calculation of the battery's current DCIR includes,
[0013] Battery Current
[0014] Current relative
[0015] The fully connected neural network model includes an input layer, a hidden layer, and an output layer.
[0016] include,
[0017] Inputting the test data into the input layer of the fully connected neural network model;
[0018] Each neuron in the hidden layer receives information from all nodes in the input layer and calculates its own activation value based on the weight parameters and bias parameters.
[0019] The activation value is calculated through the activation function Leaky ReLU to obtain the output parameter.
[0020] include,
[0021]
[0022] in, represents the linear combination of the i-th neuron in the l-th layer, t l-1 represents the number of neurons in the l-1th layer, represents the weight parameter between the i-th neuron in the l-th layer and the j-th neuron in the l-1-th layer, represents the activation value of the jth neuron in the l-1 layer, represents the bias parameter of the ith neuron.
[0023] The output parameters The calculation includes,
[0024]
[0025] Among them, f(.) represents the activation function.
[0026] Set the relative DCIR value to y i′ , calculate the mean square error MSE between the output parameter and the relative DCIR value,
[0027]
[0028] Where n represents the total number of test data, and i′ represents the test data index.
[0029] Starting from the output layer, the gradients of the hidden layer weight parameters and bias parameters are reversely calculated, and the Adam algorithm is used to update the weight parameters and bias parameters according to the calculated gradients. The cycle is iterated until the difference between the output parameter and the relative DCIR value is minimized, and finally the optimal DCIR model is obtained.
[0030] Beneficial effect: The DCIR model proposed in the present invention can capture the relative DCIR "decline" phenomenon of the energy storage lithium-ion battery in the early stage of operation and can estimate the DCIR "zero" growth time under specific working conditions, while making the DCIR prediction curve better consistent with the test data.
[0031] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present application.
[0033] Figure 1 is a flow chart according to the present invention;
[0034] Figure 2 is a schematic diagram of the hidden layer;
[0035] Figure 3 It is a prediction curve diagram based on the traditional DCIR model and the DCIR model proposed by the present invention;
[0036] Figure 4 It is a line graph of the error between the simulation curve and the measured data based on the traditional DCIR model and the DCIR model proposed by the present invention at room temperature;
[0037] Figure 5 It is a line graph showing the error between the simulation curves based on the traditional DCIR model and the DCIR model proposed in the present invention and the measured data under high temperature. DETAILED DESCRIPTION
[0038] The following is a description of exemplary embodiments of the present application in conjunction with the accompanying drawings, including various details of the embodiments of the present application to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.
[0039] like Figure 1 As shown, the present application provides a method for constructing a DCIR model of a lithium-ion battery for energy storage based on a neural network, comprising:
[0040] S1: Collect test data of lithium-ion batteries at different temperatures, different SOCs and different cycle conditions through testing. It should be noted that:
[0041] The test data includes battery storage days, cycle number, battery capacity, and battery real-time temperature.
[0042] S2: Obtaining the relative DCIR value of the lithium-ion battery under the current test conditions based on the test data. It should be noted that:
[0043] The acquisition of the relative DCIR (Direct Current Internal Resistance) value includes:
[0044] When the battery is at the current temperature and 50% SOC, the battery is pulse-discharged with a current I of 200A for a duration of 10s, and the voltage V of the battery before discharge is recorded. i前 And the voltage after discharge V i后 , repeat the above discharge operation for multiple times, and take the average value of these multiple times as the current DCIR of the battery;
[0045] Since the internal resistance of the battery will be different under different temperature and SOC conditions, an accurate DCIR value can be obtained by pulse discharging the battery with a current I of 200A for 10s under the condition that the battery is at the current temperature and 50% SOC; setting the current size to 200A and the discharge time to 10s can avoid the DCIR value of the battery changing with the current size and discharge time; pulse discharge is used instead of continuous discharge because pulse discharge can more accurately measure the DCIR value of the battery.
[0046] By repeating the discharge operation multiple times and taking the average value, the measurement error can be reduced and the accuracy of DCIR measurement can be improved.
[0047] The calculation of the battery's current DCIR includes,
[0048] Battery Current
[0049] Current relative
[0050] S3: Fit the test data using a fully connected neural network model to obtain output parameters. It should be noted that:
[0051] The present invention adopts a fully connected neural network model to fit the test data, and uses the fully connected neural network to predict future trends, so as to estimate the time of "zero" growth of DCIR.
[0052] The fully connected neural network model includes an input layer, a hidden layer, and an output layer.
[0053] include,
[0054] Inputting the test data into the input layer of the fully connected neural network model;
[0055] Each neuron in the hidden layer receives information from all nodes in the input layer and calculates its own activation value based on the weight parameters and bias parameters.
[0056] The activation value is calculated through the activation function Leaky ReLU to obtain the output parameter.
[0057] Furthermore, using Leaky ReLU as the activation function can solve the dead zone problem of traditional ReLU in the negative half axis, increase the nonlinearity of the model, and enable the DCIR model proposed in the present invention to accurately capture the behavior of the DCIR of the energy storage lithium-ion battery decreasing first and then increasing in the early stage of operation.
[0058] include,
[0059]
[0060] in, represents the linear combination of the i-th neuron in the l-th layer, t l-1 represents the number of neurons in the l-1th layer, represents the weight parameter between the i-th neuron in the l-th layer and the j-th neuron in the l-1-th layer, represents the activation value of the jth neuron in the l-1 layer, represents the bias parameter of the ith neuron.
[0061] The output parameters The calculation includes,
[0062]
[0063] Among them, f(.) represents the activation function.
[0064] The measurement parameters are input into the input layer of the fully connected neural network model. The measurement parameters will be used as the node values of the input layer and will be transmitted and transformed through a series of weight parameters and bias parameters to form the optimal DCIR model.
[0065] The input layer contains multiple nodes, each node corresponds to an input parameter, and these input parameters will be used as input of the neural network to predict the target variable.
[0066] The hidden layer is the core part of the fully connected neural network model, responsible for extracting features and modeling complex relationships. Figure 2 As shown in the figure, this embodiment sets four hidden layers, and the number of neurons in each layer is 32, 64, 128 and 64 respectively. Each neuron receives input from all neurons in the previous layer, and calculates its own activation value through weight parameters and bias parameters. The activation value will get the final output value through the activation function Leaky ReLU. For example, from the second neuron in the second layer (value is θ) to the fourth neuron in the third layer, the mapping relationship is y 22→34 =w 22→34 ×θ+b 22→34 Similarly, the value of the fourth neuron in the third layer is the sum of all the mappings from the neurons in the second layer, that is, y 34 =∑y 2i→34 .
[0067] To speed up model training, we apply Batch Normalization before the activation value of each layer. Batch Normalization normalizes the input to a standard normal distribution with a mean of 0 and a variance of 1, which helps reduce the problem of gradient vanishing and exploding and speeds up convergence.
[0068] Dropout technique is used between hidden layers to avoid overfitting.
[0069] Leaky ReLu is used as the activation function to increase the nonlinearity of the model. In the output layer, we set the final output result to 1 output parameter, that is, the 32 neurons in the 5th layer of the hidden layer output 1 output parameter after passing through the activation function.
[0070] S4: By using MSE as the loss function, by adjusting the weight parameters and bias parameters of the fully connected neural network model, the difference between the output parameter and the relative DCIR value is minimized, and finally the optimal DCIR model is obtained. It should be noted that:
[0071] Set the relative DCIR value to y i′ , calculate the mean square error MSE between the output parameter and the relative DCIR value,
[0072]
[0073] Where n represents the total number of test data, and i′ represents the test data index.
[0074] Starting from the output layer, the gradients of the hidden layer weight parameters and bias parameters are calculated in reverse, and the Adam (adaptive momentum estimation) algorithm is used to update the weight parameters and bias parameters according to the calculated gradients. The cycle is iterated until the difference between the output parameter and the relative DCIR value is minimized, and finally the optimal DCIR model is obtained.
[0075] Adam is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process. It can iteratively update the weight parameters and bias parameters of the fully connected neural network model based on training data. Adam combines the advantages of AdaGrad and RMSProp algorithms, while taking into account the performance of sparse gradients and non-steady-state and online problems.
[0076] At present, a certain energy storage project of the company requires to predict the future trend of battery DCIR based on the existing N-circuit data. Now, the DCIR test data is fitted and predicted based on the traditional DCIR model and the DCIR model based on neural network proposed by the present invention.
[0077] The specific implementation steps are as follows:
[0078] Step 1: At the test center, collect battery data at different temperatures, different SOCs, and different cycle conditions, including battery cell storage days, number of cycles, battery capacity, real-time battery cell temperature, and other data.
[0079] Step 2: Based on the test data obtained in step 1, obtain the relative DCIR value of the battery under the current test conditions (this is the actual value calculated through the test). The test method is: let the battery be subjected to a pulse discharge of 10s with a current of 200A at 45°C and 50% SOC, and then record the battery voltage V before discharge. 前 And the voltage after discharge V 后 . Using the current DCIR=(V 前 -V 后 ) / I to calculate the DCIR value. Repeat the above discharge operation multiple times, and take the average value of these multiple times as the current DCIR of the battery. Then use current DCIR / battery initial DCIR = current relative DCIR value.
[0080] Step 3: Based on the data obtained in step 2, the test data is fitted using the traditional DCIR model and the DCIR model proposed in the present invention to obtain the corresponding fitting parameters and fitting results, and the traditional DCIR prediction curves are drawn respectively, and the fitting results based on the traditional DCIR model and the DCIR model proposed in the present invention are compared, as shown in FIG. Figure 3-5 shown.
[0081] Among them, the calculation formula of the traditional DCIR model is:
[0082]
[0083] Among them, DCIR cell Indicates the relative DCIR (%) of the battery under working conditions. A and B are both pre-exponential factors, E a and E b represents activation energy, T represents battery temperature, Time represents cycle time, SOC represents the state of charge value during the cycle, Cycle represents the number of equivalent 100% DOD cycles, and x, y, and z represent fitting parameters.
[0084] like Figure 3 As shown in the figure, at both normal and high temperatures, the traditional DCIR model is completely unable to fit the relative DCIR decrease part, and the relative DCIR increases monotonically starting from 1, so it will cause a lot of deviations. The DCIR formula proposed in the present invention can fit the original data very well, and can be used to make a trend of first decreasing and then increasing. This greatly improves the prediction accuracy of DCIR. At the same time, we can predict the time of DCIR "zero" growth, such as the number of DCIR "zero" growth cycles in high temperature cycles is about 2N cycles.
[0085] like Figure 4 As shown in the figure, at room temperature, the fitting error based on the traditional DCIR model (dashed line part) is very large, and as the number of turns increases, the error continues to increase. When the number of turns is N, the average error is already greater than 20%. However, the fitting error of the DCIR model (solid line part) proposed in the present invention is smaller, and the error value does not fluctuate dramatically with the number of turns.
[0086] like Figure 5 As shown, at high temperatures, the fitting error based on the traditional DCIR model (dashed line portion) is large, and as the number of turns increases, the error continues to increase. However, unlike at room temperature, when the number of cycles at high temperature is N, the average error is about 7%. This shows that the magnitude of the DCIR decrease is strongly correlated with temperature. The higher the temperature, the smaller the magnitude of the DCIR decrease. The DCIR model (solid line portion) proposed in the present invention has a smaller fitting error, with an average error of about 1%, and the error value does not fluctuate dramatically with the number of turns.
[0087] In summary, the fitting effect of the DCIR model proposed in the present invention is significantly better than that of the traditional DCIR model. The traditional DCIR model cannot capture its "decline" area, while the DCIR model proposed in the present invention can capture this area well, and the DCIR model proposed in the present invention can estimate the time of DCIR "zero" growth.
[0088] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A method for constructing a DCIR model of energy storage lithium-ion battery based on a neural network, characterized in that: include: Collect test data of lithium-ion batteries at different temperatures, different SOCs and different cycle conditions through testing; Obtaining a relative DCIR value of the lithium-ion battery under current test conditions based on the test data; Fitting the test data using a fully connected neural network model to obtain output parameters; By using MSE as the loss function and adjusting the weight parameters and bias parameters of the fully connected neural network model, the difference between the output parameter and the relative DCIR value is minimized, and finally the optimal DCIR model is obtained.
2. The method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network according to claim 1, characterized in that: The test data includes battery storage days, cycle number, battery capacity, and battery real-time temperature.
3. The method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network according to claim 2, characterized in that: The acquisition of the relative DCIR value includes: When the battery is at the current temperature and 50% SOC, the battery is pulse-discharged with a current I of 200A for a duration of 10s, and the voltage V of the battery before discharge is recorded. i前 And the voltage after discharge V i后 , repeat the above discharge operation for multiple times, and take the average value of these multiple times as the current DCIR of the battery; The calculation of the battery's current DCIR includes, Battery Current Current relative 4. The method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network according to claim 3, characterized in that: The fully connected neural network model includes an input layer, a hidden layer, and an output layer.
5. The method for constructing a DCIR model of a lithium-ion battery based on a neural network according to claim 7, characterized in that: include, Inputting the test data into the input layer of the fully connected neural network model; Each neuron in the hidden layer receives information from all nodes in the input layer and calculates its own activation value based on the weight parameters and bias parameters. The activation value is calculated through the activation function Leaky ReLU to obtain the output parameter.
6. The method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network according to claim 5, characterized in that: include, in, represents the linear combination of the ith neuron in the lth layer, tl -1 represents the number of neurons in the l-1th layer, represents the weight parameter between the i-th neuron in the l-th layer and the j-th neuron in the l-1-th layer, represents the activation value of the jth neuron in the l-1 layer, represents the bias parameter of the ith neuron.
7. The method for constructing a DCIR model of a lithium-ion battery based on a neural network according to claim 6, characterized in that: The output parameters The calculation includes, Among them, f(.) represents the activation function.
8. The method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network according to claim 7, characterized in that: Set the relative DCIR value to yi′, calculate the mean square error MSE between the output parameter and the relative DCIR value, Where n represents the total number of test data, and i′ represents the test data index.
9. The method for constructing a DCIR model of an energy storage lithium-ion battery based on a neural network according to claim 8, characterized in that: Starting from the output layer, the gradients of the hidden layer weight parameters and bias parameters are reversely calculated, and the Adam algorithm is used to update the weight parameters and bias parameters according to the calculated gradients. The cycle is iterated until the difference between the output parameter and the relative DCIR value is minimized, and finally the optimal DCIR model is obtained.
