Capacity joint estimation method based on energy storage battery health state

Abstract

This paper discloses a joint capacity estimation method based on the health status of energy storage batteries. The method first collects and preprocesses the operating and aging data of the energy storage batteries, and simultaneously divides the aging characteristic intervals of the battery health status. Then, a second-order Thevenin equivalent circuit model is constructed, and the model parameters are identified online using the recursive least squares method. Next, two-way capacity pre-estimation is completed by respectively using an improved adaptive extended Kalman filter algorithm and a trained and optimized gated recurrent unit (GRU) network. Finally, based on the aging characteristic intervals, the fusion weights of the two pre-estimation values ​​are dynamically adjusted using a fuzzy control algorithm, and the final usable capacity of the battery is output after weighted calculation. This invention combines the advantages of model-based and data-driven methods, solving the problems of large cumulative errors, unstable accuracy throughout the entire life cycle, and weak anti-interference ability of existing single methods. It can achieve online high-precision capacity estimation of energy storage batteries throughout their entire life cycle and under multiple operating conditions.

Description

Technical Field

[0001] This invention relates to a joint capacity estimation method based on the health status of energy storage batteries. Background Technology

[0002] With the rapid development of the new energy industry, electrochemical energy storage power stations have become core facilities for grid peak shaving and new energy consumption. As the core unit of energy storage power stations, lithium-ion energy storage batteries are crucial indicators for assessing battery health, formulating charging and discharging strategies, and achieving tiered utilization. The accuracy of capacity estimation directly determines the safety, economy, and service life of the energy storage system.

[0003] Currently, existing battery capacity estimation methods are mainly divided into three categories: First, the traditional ampere-hour integration method, which is simple in principle and easy to implement, but suffers from cumulative current measurement errors. After long-term use, the estimation deviation will continue to amplify, making it unsuitable for long-term cycle energy storage battery capacity estimation alone. Second, model-based estimation methods, such as Kalman filter algorithms, which have strong dynamic tracking capabilities and can achieve online estimation, but require high accuracy of the battery equivalent circuit model. After deep battery aging, the model parameter matching degree decreases, and the estimation accuracy will be significantly reduced. Third, data-driven estimation methods, such as neural network algorithms, which have strong fitting ability for the nonlinear aging characteristics of batteries and do not require a precise physical model, but require a large amount of full life cycle aging data to support training. They have poor generalization under small sample conditions and are easily affected by sensor noise.

[0004] In summary, existing single capacity estimation methods all have obvious application limitations and cannot meet the online high-precision estimation requirements of energy storage batteries throughout their entire life cycle and under multiple operating conditions. Therefore, there is an urgent need to develop a capacity estimation scheme that combines the advantages of multiple methods and can be adaptively adjusted. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned deficiencies in the prior art by providing a joint capacity estimation method based on the health status of energy storage batteries. This method solves the problems of large cumulative errors, unstable accuracy throughout the entire life cycle, and weak anti-interference capabilities of existing single estimation methods, thereby achieving online high-precision capacity estimation of energy storage batteries.

[0006] A joint capacity estimation method based on the health status of energy storage batteries includes:

[0007] S1. Data Acquisition and Preprocessing: Collect real-time operating data and historical cycle aging data of energy storage batteries, and perform noise reduction and normalization preprocessing on the collected data; at the same time, collect historical full life cycle information of this type of energy storage battery, and divide the aging characteristic range of the health state S of the energy storage battery.

[0008] S2. Constructing an equivalent circuit model: Based on the preprocessed battery data, construct a second-order Thevenin equivalent circuit model of the battery, and identify the ohmic internal resistance, polarization internal resistance and polarization capacitance parameters of the model online by recursive least squares method.

[0009] S3. Equivalent circuit capacity prediction: Based on the identified equivalent circuit model, an improved adaptive extended Kalman filter algorithm is used to predict the available capacity of the battery using the equivalent circuit capacity, and the equivalent circuit capacity prediction value is obtained.

[0010] S4. Real-time state capacity prediction: Using the historical operation and aging characteristics of this type of energy storage battery as input features and the actual battery capacity as output label, a gated recurrent unit (GRU) deep learning network is trained. The optimized network after training is used to perform real-time state capacity prediction of the battery's available capacity, and the real-time state capacity prediction value is obtained.

[0011] S5. Weighted Fusion and Result Output: Based on the battery aging characteristic intervals divided in step S1, the fusion weights of the equivalent circuit capacity estimate and the real-time state capacity estimate are dynamically adjusted through a fuzzy control algorithm. After weighted calculation, the final available capacity estimate of the battery is output.

[0012] Furthermore, the aging characteristic intervals are divided as follows: fresh interval S≥90%, mildly aged interval 80%≤S<90%, moderately aged interval 70%≤S<80%, and deeply aged interval S<70%.

[0013] Furthermore, in step S2, the state equations and observation equations of the second-order Thevenin equivalent circuit model are as follows:

[0014] Equations of state:

[0015]

[0016] , ,

[0017] Observation equation:

[0018]

[0019] in, Let k be the state of charge at time k. , Let be the polarization voltage at time k. The sampling interval is... , The polarization time constant is For charging and discharging efficiency, For the rated capacity of the battery, Let K be the charging and discharging current at time k. Let be the battery terminal voltage at time k. This is the battery open-circuit voltage. This refers to the battery's internal resistance. , For polarization internal resistance, , It is a polarizing capacitor.

[0020] Furthermore, in step S3, the estimated equivalent circuit capacity... The calculation formula is:

[0021]

[0022] in, This represents the initial SOC value during charging and discharging. The SOC value at the end of the charge / discharge cycle is denoted as n, where n is the number of sampling points.

[0023] Furthermore, in step S4, the construction and training process of the gated recurrent unit (GRU) deep learning network is as follows:

[0024] S41. Dataset Construction: Select cyclic test data of the entire battery life cycle, with the battery terminal voltage, charge and discharge current, surface temperature, ohmic internal resistance and number of cycles as input features, and the actual battery capacity obtained from standard charge and discharge tests as output labels, and divide the training set and test set in an 8:2 ratio.

[0025] S42. Network structure construction: The network consists of an input layer, two GRU hidden layers, a Dropout layer, a fully connected layer, and an output layer. The two GRU hidden layers have 64 and 32 neurons respectively, and the inactivation rate of the Dropout layer is 0.2.

[0026] S43. Network Training: The Adam optimizer is used, the learning rate is set to 0.001, the loss function is the mean squared error (MSE), and the network is trained iteratively for 100 rounds to complete the training and optimization.

[0027] S44. Input the preprocessed real-time battery operating data into the trained GRU network, and output the real-time state capacity estimate of the battery. .

[0028] Furthermore, in step S5, the formula for calculating the final available capacity estimation result is as follows:

[0029]

[0030] in , As weight, This is the final predicted capacity value.

[0031] Furthermore, the rules for determining the fusion weights are as follows:

[0032] When the battery's S value is in the freshness range and the mild aging range... The value range is 0.6 to 0.8. The value range is 0.2 to 0.4;

[0033] When the battery's S value is in the moderate aging range or the deep aging range The value range is 0.2 to 0.4. The value range is 0.6 to 0.8.

[0034] Furthermore, step S5 also includes adaptive weight adjustment based on operating conditions: when the battery current fluctuation exceeds a preset threshold of 0.5C, the weight adjustment will be adjusted accordingly. Increase by 0.1~0.2, corresponding to a decrease. And after adjustment, it meets the requirements. + =1.

[0035] Beneficial effects:

[0036] This invention integrates the advantages of both capacity estimation methods based on equivalent circuit models and data-driven capacity estimation methods based on GRU networks, complementing the inherent defects of single methods. Furthermore, it dynamically matches the fusion weights of the two estimation results based on battery aging characteristic intervals. In the fresh and mildly aged stages, it relies on a model method with high matching degree to ensure estimation accuracy, while in the moderate and deep aged stages, it relies on a data-driven method with stronger nonlinear fitting capabilities to maintain estimation stability. This fundamentally solves the problems of large fluctuations in accuracy throughout the entire lifecycle and significant amplification of estimation deviations after deep battery aging in existing single methods, achieving stable and high-precision capacity estimation throughout the entire lifecycle of energy storage batteries. Attached Figure Description

[0037] Figure 1 This is a flowchart of the capacity estimation method based on the equivalent circuit model;

[0038] Figure 2 This is a comparison table of the estimation accuracy of energy storage batteries throughout their entire life cycle;

[0039] Figure 3 This is a test table for anti-interference capability under different working conditions;

[0040] Figure 4 It is a test table for error stability during long-term cycles. Detailed Implementation

[0041] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments and accompanying drawings. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0042] Example 1:

[0043] This embodiment provides a battery prediction method, as detailed below:

[0044] 1. Experimental Subjects and Testing Environment

[0045] Experimental battery cell: Commercial 280Ah square lithium iron phosphate energy storage cell, rated voltage 3.2V, charging cut-off voltage 3.65V, discharging cut-off voltage 2.5V, nominal cycle life ≥6000 cycles (S≥80%).

[0046] Test environment: The temperature and humidity chamber was maintained at 25℃±1℃ and 45%±5% relative humidity.

[0047] Testing equipment: battery charge and discharge test system, high and low temperature alternating damp heat test chamber, BMS performance test platform.

[0048] 2. Test Benchmark and Comparison Scheme

[0049] Actual capacity benchmark: After every 100 cycles of aging, a standard charge-discharge test is performed (1C constant current and constant voltage charging to full charge, resting for 1 hour, 1C constant current discharging to cutoff voltage), and the discharge capacity is recorded as the actual capacity of the cell, while the corresponding S value is calculated.

[0050] 3. Comparison Method: Three mainstream capacity estimation methods commonly used in the industry were selected as control groups: the traditional ampere-hour integration method, the traditional extended Kalman filter (EKF) method, and the single GRU neural network method. The method of this invention was compared with the method of this invention under the same conditions.

[0051] Test Implementation Process

[0052] The capacity estimation is performed according to the technical solution of this invention, and the specific steps are as follows:

[0053] (1) Data acquisition and preprocessing: Real-time charging and discharging current, terminal voltage, surface temperature data of the battery cell, as well as historical cycle aging data are collected. Noise reduction is performed by moving average filtering, and the data is normalized by min-max. The aging history data of the battery cell of this model throughout its entire life cycle is collected, and S is divided into four intervals: fresh interval (S≥90%), lightly aged interval (80%≤S<90%), moderately aged interval (70%≤S<80%), and deeply aged interval (S<70%).

[0054] (2) Construct a second-order Thevenin equivalent circuit model. Based on the preprocessed data, identify the ohmic internal resistance of the model online using the recursive least squares method. Polarization internal resistance , With polarization capacitor , Parameters are used to construct the state equations and observation equations of the model.

[0055] (3) Based on the identified equivalent circuit model, the battery SOC is estimated by using the improved adaptive extended Kalman filter algorithm, and the estimated equivalent circuit capacity is calculated by combining the ampere-hour integral formula. .

[0056] (4) Construct and train the GRU network: Select the cyclic test data of the entire life cycle of the battery cell of this model, with terminal voltage, charge and discharge current, surface temperature, ohmic internal resistance and number of cycles as input features, and the actual capacity obtained from the standard charge and discharge test as the output label. Divide the training set and the test set in an 8:2 ratio; build the network structure as input layer - 64-neuron GRU layer - 32-neuron GRU layer - Dropout layer with inactivation rate of 0.2 - fully connected layer - output layer; use Adam optimizer with learning rate of 0.001 and loss function as mean square error (MSE), and iterate for 100 rounds to complete the optimization; input the real-time data of the battery cell into the trained network, and output the real-time state capacity prediction. .

[0057] (5) Weight fusion and result output: Based on the current S-interval of the battery cell, the fusion weight is determined by the fuzzy control algorithm, and the fresh and slightly aged intervals are selected. Take 0.7, Take 0.3, for the moderate to deep aging range. Take 0.3, Set the current to 0.7; when the detected current fluctuation exceeds 0.5C, [the value will be adjusted]. Increased by 0.15, corresponding to a decrease ,ensure + =1; the final capacity estimation result is obtained by calculating using a weighted formula. .

[0058] like Figure 2-4 As shown in the experimental results, the average absolute error of capacity estimation based on the joint method is controlled within 1.3% throughout the entire battery life cycle, which is far superior to the control group method. In the deep aging range, the estimation error of the method of this invention is only about 1 / 4 of that of the traditional EKF method, which verifies the design based on dynamic weight adjustment in the S range and effectively solves the problem of significant decrease in accuracy after aging of the traditional model method.

[0059] Under high current fluctuation conditions, the combined method controls the estimation error to within 1.1% and the maximum error to no more than 1.6% through adaptive weight adjustment based on operating conditions, which is far superior to the traditional EKF method. This verifies that the adaptive operating condition mechanism of the present invention effectively improves the anti-interference capability under complex operating conditions.

[0060] As the number of iterations increases, the error of the traditional ampere-hour integration method continues to amplify, while the estimation error of the combined method remains stable within 1.1%, without the problem of cumulative error amplification. This verifies the stability and reliability of the method of the present invention in long-term operation.

[0061] Example 2:

[0062] This embodiment provides a specific implementation method and verification test of the method of the present invention under low temperature conditions and complex grid frequency regulation conditions of ternary lithium energy storage cells, in order to verify the adaptability and reliability of the present invention to different energy storage cell systems, harsh temperature environments, and complex operating conditions.

[0063] Experimental battery cell: Commercial 125Ah square ternary lithium energy storage cell, rated voltage 3.7V, charging cut-off voltage 4.2V, discharging cut-off voltage 2.8V, nominal cycle life ≥4000 cycles (S≥80%).

[0064] Test environment: Two control environments were set up: low temperature condition (constant temperature and humidity chamber controlled ambient temperature at 0℃±1℃, relative humidity at 45%±5%) and normal temperature condition (25℃±1℃, relative humidity at 45%±5%).

[0065] Testing equipment: battery charge and discharge test system, high and low temperature alternating damp heat test chamber, BMS performance test platform.

[0066] Actual capacity benchmark: After every 100 cycles of aging, a standard charge-discharge test is performed at 25°C (1C constant current and constant voltage charging to full charge, resting for 1 hour, 1C constant current discharging to cutoff voltage). The discharge capacity is recorded as the actual capacity of the cell, and the corresponding S value is calculated.

[0067] Comparison method: Consistent with Example 1, the traditional ampere-hour integration method, the traditional extended Kalman filter (EKF) method, and the single GRU neural network method were selected as control groups and compared with the method of the present invention under the same conditions.

[0068] Test conditions: The standard operating conditions of primary frequency regulation of the power grid are adopted. The current fluctuates randomly in an amplitude of 0.2C to 1.2C, with a maximum fluctuation frequency of 1 time / 10s, simulating the high-frequency fluctuation operation scenario of the energy storage power station participating in the actual grid frequency regulation.

[0069] Test results: Under normal operating conditions of 25℃, the average absolute error of the combined method for estimating the full life cycle capacity of ternary lithium battery cells is controlled within 1.2%. Under low operating conditions of 0℃, the average absolute error of the traditional EKF method rises to 4.2%~6.7% due to severe drift of battery model parameters. However, the method of this invention, through dynamic weight adjustment, can still control the average absolute error within 1.5%, with a maximum error of no more than 1.9%, which is far superior to the control group method. This verifies the stable estimation capability of this invention under harsh low-temperature environments.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A joint capacity estimation method based on the health status of energy storage batteries, characterized in that, include: S1. Data Acquisition and Preprocessing: Collect real-time operating data and historical cycle aging data of energy storage batteries, and perform noise reduction and normalization preprocessing on the collected data; at the same time, collect historical full life cycle information of this type of energy storage battery and divide the aging characteristic range of the energy storage battery. S2. Constructing an equivalent circuit model: Based on the preprocessed battery data, construct a second-order Thevenin equivalent circuit model of the battery, and identify the ohmic internal resistance, polarization internal resistance and polarization capacitance parameters of the model online by recursive least squares method. S3. Equivalent circuit capacity prediction: Based on the identified equivalent circuit model, an improved adaptive extended Kalman filter algorithm is used to predict the available capacity of the battery using the equivalent circuit capacity, and the equivalent circuit capacity prediction value is obtained. S4. Real-time state capacity prediction: Using the historical operation and aging characteristics of this type of energy storage battery as input features and the actual battery capacity as output label, a gated recurrent unit (GRU) deep learning network is trained. The optimized network after training is used to perform real-time state capacity prediction of the battery's available capacity, and the real-time state capacity prediction value is obtained. S5. Weighted Fusion and Result Output: Based on the battery aging characteristic intervals divided in step S1, the fusion weights of the equivalent circuit capacity estimate and the real-time state capacity estimate are dynamically adjusted through a fuzzy control algorithm. After weighted calculation, the final available capacity estimate of the battery is output.

2. The capacity joint estimation method based on the health status of energy storage batteries according to claim 1, characterized in that, The aging characteristic intervals are divided as follows: fresh interval S≥90%, mildly aged interval 80%≤S<90%, moderately aged interval 70%≤S<80%, and deeply aged interval S<70%.

3. The capacity joint estimation method based on the health status of energy storage batteries according to claim 1, characterized in that, In step S2, the state equations and observation equations of the second-order Thevenin equivalent circuit model are as follows: Equations of state: 、 ， Observation equation: in, Let k be the state of charge at time k. , Let be the polarization voltage at time k. The sampling interval is... , The polarization time constant is For charging and discharging efficiency, For the rated capacity of the battery, Let K be the charging and discharging current at time k. Let be the battery terminal voltage at time k. This is the battery open-circuit voltage. This refers to the battery's internal resistance. , For polarization internal resistance, , It is a polarizing capacitor.

4. The capacity joint estimation method based on the health status of energy storage batteries according to claim 1, characterized in that, In step S3, the estimated equivalent circuit capacity is... The calculation formula is: in, This represents the initial SOC value during charging and discharging. The SOC value at the end of the charge / discharge cycle is denoted as n, where n is the number of sampling points.

5. The capacity joint estimation method based on the health status of energy storage batteries according to claim 1, characterized in that, In step S4, the construction and training process of the Gated Recurrent Unit (GRU) deep learning network is as follows: S41. Dataset Construction: Select cyclic test data of the entire battery life cycle, with the battery terminal voltage, charge and discharge current, surface temperature, ohmic internal resistance and number of cycles as input features, and the actual battery capacity obtained from standard charge and discharge tests as output labels, and divide the training set and test set in an 8:2 ratio. S42. Network structure construction: The network consists of an input layer, two GRU hidden layers, a Dropout layer, a fully connected layer, and an output layer. The two GRU hidden layers have 64 and 32 neurons respectively, and the inactivation rate of the Dropout layer is 0.

2. S43. Network Training: The Adam optimizer is used, the learning rate is set to 0.001, the loss function is the mean squared error (MSE), and the network is trained iteratively for 100 rounds to complete the training and optimization. S44. Input the preprocessed real-time battery operating data into the trained GRU network, and output the real-time state capacity estimate of the battery. .

6. The capacity joint estimation method based on the health status of energy storage batteries according to claim 5, characterized in that, In step S5, the formula for calculating the final available capacity estimation result is as follows: in , As weight, This is the final predicted capacity value.

7. The capacity joint estimation method based on the health status of energy storage batteries according to claim 1, characterized in that, The rules for determining the fusion weights are as follows: When the battery's S value is in the freshness range and the mild aging range... The value range is 0.6 to 0.

8. The value range is 0.2 to 0.4; When the battery's S value is in the moderate aging range or the deep aging range The value range is 0.2 to 0.

4. The value range is 0.6 to 0.

8.

8. The capacity joint estimation method based on the health status of energy storage batteries according to claim 7, characterized in that, Step S5 also includes adaptive weight adjustment based on operating conditions: when the battery current fluctuation exceeds a preset threshold of 0.5C, the weight adjustment will be adjusted accordingly. Increase by 0.1~0.2, corresponding to a decrease. And after adjustment, it meets the requirements. + =1.

Images