Energy storage battery capacity prediction method based on data driving
Through the data-driven method, electrochemical big data is used to build a capacity prediction model, which solves the problem of poor capacity calibration of traditional energy storage batteries, and achieves rapid and accurate prediction and calibration of energy storage batteries, saving costs and reducing costs Equipment loss.
Patent Information
- Application Number
- CN202310627376.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-30
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The traditional energy storage battery capacity calibration method consumes time, is high in cost and has poor accuracy, and cannot effectively shorten production process time and reduce equipment losses. The existing prediction methods based on physical models are computationally large and difficult to achieve widespread robustness.
Using a data-driven method, by acquiring electrochemical big data, using independent and fusion algorithms to learn the mapping relationship between battery characteristics and capacity, build a capacity prediction model, and realize fast and accurate capacity prediction and intelligent calibration, with only a small amount of charging. Obtain accurate results.
It realizes fast and accurate prediction and calibration of energy storage battery capacity, saves time and power costs, reduces equipment losses, and improves production efficiency and prediction robustness.
Smart Images

Figure CN119064783A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to a data-driven energy storage battery capacity prediction method, which belongs to the battery field. Background Art
[0002] Energy storage is becoming one of the key technologies for many countries today to advance their carbon neutrality goals. Electrochemical energy storage, represented by rechargeable lithium battery systems, has the characteristics of high technical maturity, short construction period, and flexible configuration. It is recognized as the most promising and fastest-growing direction in the field of energy storage. During the production and delivery of energy storage batteries, their capacity needs to be calibrated. The current traditional calibration method is to perform a complete charge and discharge, and then charge the battery to the required state of charge (SOC). However, the above steps not only require a long time to occupy the test equipment, but also generate a lot of time costs, affect production efficiency, and cause large equipment losses; at the same time, the long charging steps will also cause extremely high electricity costs, reducing the final benefits. Using appropriate methods to shorten the factory capacity division process can reduce time and power consumption costs, as well as equipment losses.
[0003] Traditional methods such as the ampere-hour method and the open-circuit voltage method generally have simple principles and small calculation amounts, but they are usually inaccurate, errors accumulate over time, and sampling errors cannot be corrected.
[0004] Prediction methods based on empirical models, such as the Thevenin model, Rint model, and RC model, can predict battery capacity in real time with higher prediction accuracy than traditional methods and can achieve closed-loop control. However, the prediction accuracy is heavily dependent on the accuracy of the empirical model.
[0005] Prediction methods based on physical models provide a better understanding of battery mechanisms. However, due to the complexity of the internal mechanisms, it is difficult to achieve broad robustness for accurate predictions of different batteries using such models, and the computational effort is extremely high.
[0006] In order to solve the above problems, the present application provides a data-driven prediction method, which does not need to consider the internal physical and chemical mechanisms of the battery. The prediction accuracy does not depend on the battery model. Only sufficient battery experimental data is needed to achieve good prediction results, and the prediction has better convenience and robustness.
[0007] DD230175I-DL Summary of the invention
[0008] According to one aspect of the present application, a data-driven energy storage battery capacity prediction method is provided. The method is based on the electrochemical big data information of the energy storage battery, and learns the mapping relationship between the battery electrochemical characteristics and the output capacity through different independent algorithms and fusion algorithms to achieve rapid prediction and intelligent calibration of the energy storage battery capacity.
[0009] To achieve the above object, the present invention provides a data-driven energy storage battery capacity prediction method, comprising the following steps:
[0010] S1: Acquire the data to be tested and perform preprocessing to obtain the derived data features of the data to be tested;
[0011] S2: Feature extraction: evaluate the importance of each of the derived data features, and select the derived data features with high scores as the input of the capacity prediction model;
[0012] S3: Model construction: by comparing the prediction effects of different test algorithm models, the one with the best effect is selected to construct the capacity prediction model;
[0013] Taking different derivative data feature combination ranges or value intervals as input, the output is the battery capacity prediction result of the required predicted capacity object;
[0014] S4: Model training: extracting 90% of the data to be tested to construct a training set to train the capacity prediction model;
[0015] Use the remaining 10% to construct a test set to test the capacity prediction model;
[0016] The capacity prediction model is trained without replacement using cross validation;
[0017] S5: Model evaluation: Compare the errors generated by different models to obtain the best model algorithm framework, select the capacity prediction model parameters with the highest accuracy as the final capacity prediction model, and achieve the most accurate capacity prediction model.
[0018] Optionally, the step S1 specifically includes the following steps:
[0019] S1-1: Data collection, collecting the target energy storage cell capacity charging and discharging data as the data to be tested;
[0020] S1-2: Data cleaning: checking the rationality of the data to be tested and deleting data that is obviously abnormal or missing;
[0021] S1-3: Data processing, using filtering and smoothing to process the data to be measured;
[0022] Discrete data with equal electrical quantity intervals are obtained by interpolation, and the electrical quantity interval can be changed according to the total number of discrete data points required.
[0023] Optionally, in order to increase the accuracy of the prediction model, different ranges of derived data feature types can be introduced, and the derived data features include at least one of dV / dQ, dQ / dV, charge and discharge curve, curve characteristic peak position, peak value, peak area, half-peak width, voltage drop in a certain charge range, etc.
[0024] Optionally, the model is constructed by arranging and combining different features of the derived data to form different feature groups.
[0025] Optionally, the step S1-1 specifically includes: collecting different batches of energy storage battery cell capacity charging and discharging data in different environments as test data, that is, the training data is oriented to multiple batches of energy storage battery cells, and includes the influence of different test environments as much as possible.
[0026] Optionally, in order to eliminate numerical problems caused by large dimensional differences, S1-3 further includes normalizing and standardizing the data to be measured;
[0027] The normalization method is (Xi–Xmin) / (Xmax–Xmin), where Xi is the current data point, Xmin is the minimum value of the data, and Xmax is the maximum value of the data;
[0028] The standardization method is (Xi–μ) / σ, where μ is the mean and σ is the standard deviation.
[0029] Optionally, the machine learning algorithm includes any one of random forest, linear regression, LASSO, and XGBoost.
[0030] Optionally, the test algorithm model may be an independent algorithm model or a fusion algorithm of different algorithm models.
[0031] Optionally, the independent algorithm model in the algorithm model adopts any one of linear regression, logistic regression, Gaussian process, Bayes, support vector machine, decision tree, random forest, CatBoost, XGBoost, AdaBoost, LightBoost, and neural network.
[0032] Optionally, the fusion algorithm adopts any one of voting method, averaging method, sorting method, Stacking, and Blending.
[0033] Optionally, the model evaluation uses any one or more indicators of average error, maximum error, mean square error, root mean square error, and determination coefficient for analysis and comparison.
[0034] Optionally, the use of different characteristic value ranges involves the time and power consumption required for the final production line capacity division process. While ensuring the most accurate prediction accuracy possible, consideration should be given to narrowing the required characteristic value range of the derivative data.
[0035] Optionally, in step S2, a machine learning algorithm is used to evaluate the importance of the derived data features.
[0036] The beneficial effects of this application include:
[0037] 1) The data-driven energy storage battery capacity prediction method provided in this application is based on the electrochemical big data information of the energy storage battery, and learns the mapping relationship between the battery electrochemical characteristics and the output capacity through different independent algorithms and fusion algorithms to achieve rapid prediction and intelligent calibration of the energy storage battery capacity;
[0038] 2) The data-driven energy storage battery capacity prediction method provided in this application makes the capacity prediction result more accurate through feature type and interval optimization, feature importance screening, algorithm selection and combination optimization;
[0039] 3) The data-driven energy storage battery capacity prediction method provided in this application applies the above model to the factory calibration link of the energy storage battery production line. Only a small amount of charging time (such as 0-20% SOC charging process) is required in the capacity division process to obtain accurate battery capacity calibration, thereby speeding up the battery factory process. This method can greatly save time and electricity costs and reduce equipment losses during the production of energy storage batteries. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 A data-driven energy storage battery capacity prediction flow chart in one embodiment of the present application;
[0041] Figure 2 In the embodiment of the present application, the original collected data is smoothed and interpolated, the thin solid line is the original data, and the dotted line is the processed data;
[0042] Figure 3 The percentage of power consumption when charging to different voltages in the embodiment of the present application;
[0043] Figure 4 For the data of 0-100Ah in the embodiment of the present application, the differential value of voltage to electricity (dV / dQ) is obtained;
[0044] Figure 5 In the embodiments of the present application, the charge and discharge and dV / dQ data features of 0-50Ah are used, and the XGBoost algorithm is used for capacity prediction, and the obtained predicted value is compared with the actual value. DETAILED DESCRIPTION
[0045] The present application is described in detail below with reference to embodiments, but the present application is not limited to these embodiments.
[0046] Example
[0047] In this embodiment, see Figure 1 As shown, the present invention proposes a data-driven energy storage battery capacity prediction method. This method does not need to consider the complex physical and chemical processes inside the battery. It builds the intrinsic connection between the battery cell test data and the battery cell capacity through machine learning to achieve efficient and accurate capacity prediction.
[0048] Prediction, the method comprises the following steps:
[0049] Step 1: Data collection: In the capacity division step, the charging data of the 280Ah energy storage cell SOC from 0-100% is collected, that is, the charge Q and voltage V curve.
[0050] Step 2: Data cleaning: Check the rationality of the test data and delete data with obvious abnormal open circuit voltage, obvious abnormal charging curve, incomplete charging or missing data.
[0051] Step 3: Data processing. For the above charging data, firstly, filter smoothing is adopted, and then the data is interpolated to obtain N groups of data points, such as 1001 groups, 2001 groups, 3001 groups, etc. The obtained processed curve is as follows: Figure 2 As shown (the thin solid line is the original data, and the dotted line is the processed data);
[0052] At the same time, it is possible to consider processing the above data in a certain way to obtain derived data features, such as dV / dQ, dQ / dV, curve characteristic peak position / peak value / peak area / half peak width, etc.
[0053] Step 4: Feature extraction: Divide the feature data into different value ranges for discussion, such as 0-50Ah, 0-100Ah, 0-150Ah, 0-200Ah, etc.
[0054] When using different interval data to build models for capacity forecasting, different degrees of cost savings can be achieved, such as Figure 3 As shown;
[0055] Taking 0-100Ah as an example, the dV / dQ curve obtained is as follows Figure 4 shown.
[0056] On the other hand, the different types of features obtained above are combined at different levels to discuss the impact of different feature combination ranges on prediction accuracy and model robustness.
[0057] In addition, in order to reduce the complexity of the model and avoid overfitting, random forest, linear regression, LASSO, XGBoost and other methods can be used to calculate the importance of features, perform necessary feature selection, and discuss the model prediction performance of different percentages of data features before selection.
[0058] Step 5: Model construction. Using different numerical intervals or feature combination ranges as model input and the required predicted capacity object as output, different independent algorithm models and fusion algorithm models are used to build the capacity prediction model, including linear regression, logistic regression, Gaussian process, Bayesian, support vector machine, decision tree, random forest, CatBoost, XGBoost, AdaBoost, LightBoost, neural network, and fusion algorithms of different algorithm combinations.
[0059] Step 6: Model training. Use 90% of the data as the training set and 10% as the test set, and use k-fold cross validation to train without replacement to obtain higher prediction performance.
[0060] Step 7: Model evaluation: Compare the errors generated by different models, including average error, maximum error, etc., to obtain the best model algorithm framework and realize the most accurate capacity prediction model.
[0061] In this example, the capacity prediction results of the prediction model constructed by the DD230175I-DLXGboost algorithm are as follows: Figure 5 As shown, the average prediction error is 0.5% and the maximum error is 1.8%. It should be noted that more accurate prediction can be achieved by further increasing the amount of battery cell data, expanding the data value range, and optimizing the algorithm combination framework.
[0062] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.
Claims
1. A data-driven energy storage battery capacity prediction method, characterized in that: The steps include: S1: Acquire the data to be tested and perform preprocessing to obtain the derived data features of the data to be tested; S2: Feature extraction: evaluating the importance of the derived data features and selecting the derived data features with high scores as inputs to the capacity prediction model; S3: Model construction: by comparing the prediction effects of different test algorithm models, the one with the best effect is selected to construct the capacity prediction model; Taking different derivative data feature combination ranges or value intervals as input, the output is the battery capacity prediction result of the required predicted capacity object; S4: Model training: extract part of the data to be tested to construct a training set, and use the remaining part to construct a test set; The capacity prediction model is trained without replacement using cross validation; S5: Model evaluation: The capacity prediction model parameters with the highest accuracy are selected as the final capacity prediction model.
2. The data-driven energy storage battery capacity prediction method according to claim 1, characterized in that: The step S1 specifically includes the following steps: S1-1: Data collection, collecting the target energy storage cell capacity charging and discharging data as the data to be tested; S1-2: Data cleaning: checking the rationality of the data to be tested and deleting data that is obviously abnormal or missing; S1-3: Data processing: using filtering and smoothing to process the data to be measured; Obtain discrete data of equal electrical quantity intervals by interpolation, wherein the electrical quantity interval can be changed according to the total number of discrete data points required; Preferably, in step S2, a machine learning algorithm is used to evaluate the importance of the derived data features; Preferably, while ensuring accurate prediction accuracy, the required characteristic value range of the derived data is reduced; Preferably, the model evaluation uses average error, maximum error, mean square error, root mean square DD230175I-DL Any one or more indicators of error and determination coefficient can be analyzed and compared.
3. The data-driven energy storage battery capacity prediction method according to claim 1, characterized in that: The derived data features include at least one of the following: dV / dQ, dQ / dV, charge and discharge curve, curve characteristic peak position, peak value, peak area, and half peak width.
4. The data-driven energy storage battery capacity prediction method according to claim 3 is characterized in that: The model is constructed by arranging and combining different features of the derived data to form different feature groups.
5. The data-driven energy storage battery capacity prediction method according to claim 2, characterized in that: The step S1-1 specifically includes: collecting different batches of energy storage cell capacity charging and discharging data in different environments as the data to be tested.
6. The data-driven energy storage battery capacity prediction method according to claim 2, characterized in that: The S1-3 also includes normalizing and standardizing the data to be tested; The normalization method is (Xi–Xmin) / (Xmax–Xmin), where Xi is the current data point, Xmin is the minimum value of the data, and Xmax is the maximum value of the data; The standardization method is (Xi–μ) / σ, where μ is the mean and σ is the standard deviation.
7. The data-driven energy storage battery capacity prediction method according to claim 1, characterized in that: The machine learning algorithm includes any one of random forest, linear regression, LASSO, and XGBoost.
8. The data-driven energy storage battery capacity prediction method according to claim 1, characterized in that: The test algorithm model is an independent algorithm model or a fusion algorithm of different algorithm models.
9. The data-driven energy storage battery capacity prediction method according to claim 8, characterized in that: The independent algorithm model in the algorithm model adopts linear regression, logistic regression, DD230175I-DL Any of Gaussian Process, Bayesian, Support Vector Machine, Decision Tree, Random Forest, CatBoost, XGBoost, AdaBoost, LightBoost, Neural Network.
10. The data-driven energy storage battery capacity prediction method according to claim 8, characterized in that: The fusion algorithm adopts any one of voting method, averaging method, sorting method, Stacking and Blending.
Citation Information
Patent Citations
Lithium battery energy storage power station residual life prediction method based on multiple linear regression
CN112255559A
Battery cell capacity prediction method and device, electronic equipment and medium
CN113687250A
Battery capacity estimation method, electronic device and storage medium
CN113711458A
Lithium battery capacity prediction method based on chaos theory and nonlinear system
CN115018133A
Battery capacity prediction model training method, battery capacity grading method and device
CN115470995A
