A data-driven real-time online lithium battery SoC accurate estimation method

Abstract

The present invention discloses a data-driven real-time online lithium battery SoC accurate estimation method. The data-driven method is used to calculate SOC in real time, which can avoid the complexity of physical modeling with the same estimation accuracy. When the model is applied to similar projects, it only needs to adjust the model parameters through data self-learning without the need for re-modeling. The model parameters are updated in time through real-time data to adapt to changes in the battery working state, battery performance degradation state, and environmental conditions to obtain the best estimation accuracy. By making full use of test data and real-time data, a coarse model is constructed based on test data, which can avoid the insufficiency of being able to construct a data-driven model when there is insufficient real-time data. At the same time, the test data can provide empirical guidance for the construction of the SOC model in terms of parameter selection and effect evaluation, thereby improving modeling efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of electrochemical energy storage, and in particular to a data-driven real-time online lithium battery SoC accurate estimation method. Background Art

[0002] The lithium-ion battery energy storage industry is currently experiencing rapid growth. According to the "China Energy Storage Industry Market Outlook and Investment Forecast Analysis Report" by the Forward-looking Industry Research Institute, the installed capacity of electrochemical energy storage is expected to reach 55.9GW by 2025. Lithium-ion battery energy storage has a wide range of applications, including base station energy storage, IDC energy storage, industrial and commercial energy storage, user-side energy storage, grid-side energy storage, and power generation-side energy storage, and is developing towards diversification.

[0003] The state of charge (SOC) represents the available capacity of a battery and is a crucial basis for energy storage battery charging and discharging to achieve peak shaving and valley filling, as well as backup power. Therefore, accurate real-time online SOC estimation is crucial. Current SOC estimation methods primarily include discharge experimentation, ampere-hour integration, open-circuit voltage, internal resistance, Kalman filtering, and machine learning methods, particularly neural networks. The discharge experimentation method cannot achieve online estimation; the ampere-hour integration method simply records the amount of charge flowing into and out of the battery externally, ignoring changes in the battery's internal state. Furthermore, inaccurate current measurement leads to cumulative SOC calculation errors, necessitating regular calibration. The open-circuit voltage method requires the battery to remain stationary for extended periods, and current fluctuations caused by the battery's charge-discharge ratio can cause changes in the battery's open-circuit voltage, leading to inconsistencies in the battery pack's open-circuit voltage and significant deviations between the predicted remaining capacity and the actual remaining capacity. The accuracy of the Kalman filtering method relies heavily on the development of a battery equivalent model, which is complex and computationally intensive.

[0004] Data indicates that in recent years, data-driven approaches based on machine learning have begun to be applied to SOC estimation, often achieving high estimation accuracy. Data-driven approaches do not require the construction of detailed battery physical models and can fully utilize the large amount of monitoring data generated during battery operation. Compared to other estimation methods that rely on battery models, data-driven approaches offer greater cost-effectiveness and feasibility for achieving the same estimation accuracy.

[0005] Chinese patent document CN107064811A discloses a "lithium battery SOC online estimation method." The method includes the following steps: 1) measuring the battery's open circuit voltage and obtaining the initial state of charge (SOC) value based on the OCV-SOC curve; 2) establishing a second-order RC equivalent model of the battery and estimating the initial parameters of the battery equivalent model; 3) starting the estimation program and setting the initial matching coefficient of the state equation based on the initial state of charge and the initial parameters of the battery equivalent model; 4) using an adaptive unscented Kalman filter algorithm to obtain the current battery SOC value and the current open circuit voltage based on the OCV-SOC curve; 5) starting the least squares method with a forgetting factor to identify the parameters of the current battery equivalent model, updating the matching coefficient of the state equation with the identified parameters, and calculating the battery SOC value at the next moment; 6) repeating steps 4) and 5) to obtain the battery SOC value at each moment. The above technical solution cannot adaptively adjust and switch the model used for real-time online SOC estimation before the system is operational, resulting in inaccurate calculation results. Summary of the Invention

[0006] The present invention mainly solves the technical problem that the original technical solution cannot adaptively adjust and switch the real-time online SOC estimation model before the system is running, and provides a data-driven real-time online lithium battery SoC accurate estimation method. The data-driven method is used to calculate SOC in real time, which can avoid the complexity of physical modeling with the same estimation accuracy. When the model is applied to similar projects, it only needs to adjust the model parameters through self-learning through data without the need for re-modeling. The model parameters are updated in time through real-time data to adapt to changes in battery working status, battery performance degradation status, and environmental conditions to obtain the best estimation accuracy.

[0007] The above technical problems of the present invention are mainly solved by the following technical solutions: The present invention comprises the following steps:

[0008] S1 collects lithium battery laboratory test data and groups them to construct a dataset;

[0009] S2 processes the data set and constructs necessary features;

[0010] S3 performs battery degradation identification on the dataset;

[0011] S4 constructs a model for each dataset in the dataset; all constructed models can fit each dataset with sufficient accuracy. The model construction method is a machine learning method, specifically a feedforward neural network, support vector machine, etc., and is not limited to a specific algorithm model. Each of the above datasets corresponds to a model, and the model input is used as a labeled dataset describing the model, forming a model library.

[0012] S5 performs real-time data collection;

[0013] S6 selects an SOC estimation model from a model library as an initial model;

[0014] S7 uses the initial model for real-time SOC estimation;

[0015] S8 performs drift detection on the collected actual battery data;

[0016] S9 updates the model using the current data in step S7.

[0017] Preferably, the step S1 of constructing the data set is expressed as:

[0018]

[0019] Among them, W i Indicates the charge and discharge working status; T j Indicates the temperature range; Represents the measurement data set under the above working conditions and temperature conditions, including charge and discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number and battery capacity. i Indicates the charge and discharge working state, which can be either the charge state or the discharge state; T j Indicates the temperature range, usually a smaller temperature range or a specific temperature value, such as 25 degrees, 35 degrees, or 24-26 degrees. The temperature range depends on the temperature variation range during the test and the accuracy of the temperature sensor used; Represents the measurement data set under the above operating and temperature conditions, including charge / discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number, and battery capacity.

[0020] Preferably, the step S2 is specifically to process the errors and missing data in each data set, and construct necessary features using a feature construction method, which includes statistical quantity construction, data transformation, time series analysis, signal processing, or directly using physical quantities as features.

[0021] Preferably, step S3 uses a state recognition algorithm to identify whether the battery is degraded and the degradation stage. If a data set corresponds to two battery degradation stages, the data set is divided into two subsets corresponding to the two stages. The specific algorithm may be a clustering algorithm.

[0022] Preferably, the data collected in step S5 includes energy storage battery status data and environmental operating condition data, specifically including battery charge and discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number and battery capacity, and the collection time is T.

[0023] Preferably, step S6 compares the actual collected data of time length T with the labeled data set corresponding to each model for similarity, and selects the model with the highest similarity as the current estimated model. The similarity algorithm used is not limited and can be DTW or similar methods.

[0024] Preferably, at the beginning of operation, a smaller T value is set to quickly start the model, or a model in the first stage of decay or without decay is designated as the SOC estimation model to achieve a cold start of real-time SOC estimation.

[0025] Preferably, the step S7 sets a quantity threshold, and when data is continuously collected and accumulated to the quantity threshold, the initial model is retrained using the actually collected data, and the newly trained model is used as the latest estimation model to perform real-time SOC estimation.

[0026] Preferably, the step S8 performs drift detection on the actual battery data including the battery charge and discharge current, battery voltage, battery temperature, ambient temperature and the current number of charge and discharge cycles to determine whether the data distribution has changed. If the data distribution has changed, data with a length of T starting from the change point is intercepted, and the method of step S6 is used to reselect the initial model from the model library for SOC estimation. The data distribution change detection method includes t-test, KS test and KL divergence.

[0027] Preferably, the step S9 adopts the method of step S7 to perform real-time SOC estimation, and uses the actual state of the battery and environmental operating condition data to update the model so that the model reflects the current state of the battery.

[0028] The beneficial effects of the present invention are:

[0029] 1. Using a data-driven approach to calculate SOC in real time can avoid the complexity of physical modeling with the same estimation accuracy. When the model is applied to similar projects, it only needs to adjust the model parameters through self-learning through data without the need for re-modeling.

[0030] 2. Model parameters can be updated in a timely manner through real-time data to adapt to changes in battery operating status, battery performance degradation, and environmental conditions to achieve optimal estimation accuracy.

[0031] 3. Compared to other data-driven approaches, the present method fully utilizes both experimental and real-time data. Building a coarse model based on experimental data can avoid the inability to build a data-driven model when insufficient real-time data is available. Furthermore, experimental data can provide empirical guidance for parameter selection and effect evaluation in SOC model construction, improving modeling efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1It is a flow chart of the present invention.

[0033] Figure 2 It is a system structure diagram of the present invention.

[0034] Figure 3 It is a schematic diagram of the structure of a cloud computing device of the present invention.

[0035] Figure 4 It is a structural schematic diagram of a local edge computing device of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention will be further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0037] Example: This embodiment is a data-driven real-time online lithium battery SoC accurate estimation method, such as Figure 1 As shown, the following steps are included:

[0038] S1 collects lithium battery laboratory test data, groups the data according to charge and discharge conditions and temperature, and obtains a data set for building an SOC estimation model, which can be expressed as:

[0039]

[0040] Among them, W i Indicates the charge and discharge working state, which can be either the charge state or the discharge state; T j Indicates the temperature range, usually a smaller temperature range or a specific temperature value, such as 25 degrees, 35 degrees, or 24-26 degrees. The temperature range depends on the temperature variation range during the test and the accuracy of the temperature sensor used; Represents the measurement data set under the above operating and temperature conditions, including charge / discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number, and battery capacity.

[0041] S2 processes the datasets obtained in the previous steps and constructs the necessary features. Specifically, it processes errors and missing data in each dataset and constructs the necessary features using feature construction methods. Feature construction methods include statistical construction, data transformation, time series analysis, signal processing, and other methods. Physical quantities such as current can also be directly used as features.

[0042] S3 further processes the data set obtained in the above steps and uses a state recognition algorithm to identify whether the battery has decayed and the decay stage. The specific algorithm can be a clustering algorithm. If a data set corresponds to two battery decay stages, the data set should be divided into two subsets corresponding to the two stages.

[0043] S4 constructs a model for each of the above datasets, and all constructed models can fit each dataset with sufficient accuracy. The model construction method is a machine learning method, specifically, a feedforward neural network, a support vector machine, etc., and is not limited to a specific algorithm model. Each of the above datasets corresponds to a model, and the model input corresponds to the corresponding dataset as a labeled dataset describing the model, forming a model library.

[0044] S5 uses a data acquisition device to collect real-time data on the energy storage battery status and environmental conditions, such as battery charge and discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle count, battery capacity, etc. The data is collected continuously for a period of time T.

[0045] S6 selects an SOC estimation model from the model library as the initial model. Specifically, the data actually collected for a period of time T is compared with the labeled data set corresponding to each model for similarity, and the model with the highest similarity is selected as the current estimation model. The similarity algorithm used is not restricted and can be DTW or similar methods. In the early stages of system operation, a smaller T value can be set to quickly start the model, or a model in the first stage of decay (or no decay) can be designated as the SOC estimation model to achieve a cold start of real-time SOC estimation.

[0046] S7 uses the initial model for real-time SOC estimation. When data is continuously collected and accumulated to a certain amount, the initial model is retrained using the actual collected data, and the newly trained model is used as the latest estimation model for real-time SOC estimation.

[0047] S8 performs drift detection on the collected actual battery data, such as battery charge and discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number, etc., to determine whether the data distribution has changed. If the data distribution has changed, the data with a time length of T from the change point is intercepted, and the method of step S6 is used to reselect the initial model from the model library for SOC estimation. The data distribution change detection method can be a traditional statistical method, such as t-test, KS test, KL divergence, etc.

[0048] S9 adopts the method in step S7 and uses the actual battery status and environmental operating condition data to update the model so that the model can reflect the current latest status of the battery to improve the estimation accuracy.

[0049] The specific embodiments are as follows:

[0050] 1) Collect test data and divide the test data set into four subsets according to the charge and discharge state and test temperature range (T1 / T2), namely:

[0051] {W 充电:(T1,data1)},{W 充电 :(T2,data2)}, {W 放电 :(T1,data1)},{W 放电 :(T2,data2)}

[0052] 2) Clean the data set and construct necessary features

[0053] 3) Identify the battery degradation stage corresponding to the data set. The clustering algorithm identifies that each data set has two stages, marked as p1 and p2. Thus, the data is further divided into 8 data subsets.

[0054] 4) Build a model for each subset, and use the dataset corresponding to the model input as the model tag. For example, if a model is built for the T1 temperature range, discharge state, and battery stage P1 dataset, and its inputs include current, voltage, battery temperature, current cycle number, and temperature, then the dataset corresponding to these variables in the dataset is the model tag. This results in a model library containing eight models.

[0055] 5) Collect real-time data of the battery's working status. The parameters that need to be collected are consistent with the parameters collected in the test. When the length of the collected data is T, the DTW method is used to compare the similarity between the collected data and the model label data. The model corresponding to the label with the highest similarity is used as the current SOC estimation model for SOC estimation. At this time, the battery is in the p1 stage.

[0056] 6) Continuously collect real-time data on the battery's working status. When the data length is 10T, use the accumulated actual working status data to retrain the currently used SOC estimation model and update the model parameters to match the battery's current status, environment, and working conditions, thereby maintaining the best estimation accuracy.

[0057] 7) Use the KS test method to perform drift test on the collected actual battery working data. When the data distribution changes, it indicates that the battery health state has degraded and is no longer in the p1 stage, indicating that the current model is no longer suitable for SOC estimation. Therefore, select the actual working data with a length of T from the drift starting point, and according to the similarity method used in step 5), select the model with the battery health state in the p2 stage from the model library as the current SOC estimation model.

[0058] 8) Continuously collect actual battery operating data and use the method in step 6) to update the model to match the current state, environment, and operating conditions of the battery, thereby maintaining optimal estimation accuracy.

[0059] like Figure 2As shown in FIG, the implementation system of the real-time online lithium battery SoC accurate estimation method includes two parts: the cloud and the local.

[0060] The local system is responsible for real-time data collection, data exchange, and receiving models sent from the cloud to perform SOC calculations. Data collection involves collecting real-time battery and environmental condition data from the BMS and sensors. Data exchange includes forwarding real-time data to the cloud and obtaining the latest SOC calculation model from the cloud. SOC calculations are based on the latest SOC calculation model and utilize real-time data to accurately calculate SOC.

[0061] The cloud is responsible for test data and real-time data management, model development and retraining, data drift detection, model management, and model selection. It also distributes appropriate SOC calculation models to the local system through data interaction for SOC calculation. Model development and retraining involve building an initial model using training data and retraining the model using real-time data. Data drift detection examines changes in the distribution of real-time data. Model management manages various models trained using test and real-time data, maintaining the model library. Model selection involves selecting an appropriate model from the library through similarity comparison for accurate SOC calculation.

[0062] like Figure 3 、 Figure 4 As shown, the device includes an edge computing device deployed locally in the energy storage system and a computing service device deployed in the cloud, wherein:

[0063] Cloud computing device: There are no requirements for the number, model, composition, or connection method of the power supply, CPU, memory, network interface, and input / output devices. For example, multiple CPUs may exist. The operating system can be Linux or Windows. The application program includes the software system required to implement the method of the present invention and the necessary supporting software system.

[0064] Local computing device: There are no requirements for the quantity, model, composition, or connection method of the power supply, computing processor, and memory. For example, multiple CPUs may be present. The operating system can be Linux or another operating system adapted to the processing performance of the local computing device. The application program includes the software system required to implement the method of the present invention and the necessary supporting software system. The data acquisition interface should be able to meet the requirements for collecting data from the BMS and other energy storage system-related controllers, power electronic equipment, and sensors. The remote communication module must meet the requirements for data uploading and receiving and distributing data and models, but there are no model restrictions.

[0065] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

[0066] Although this document frequently uses terms such as dataset, recession identification, and initial model, the use of other terms is not excluded. These terms are used solely to more conveniently describe and explain the essence of the present invention; interpreting them as any additional limitations is contrary to the spirit of the present invention.

Claims

1. A data-driven real-time online lithium battery SoC accurate estimation method, characterized in that: The following steps are involved: S1 collects lithium battery laboratory test data and constructs a data set by grouping according to charge and discharge conditions and temperature; S2 processes the data set and constructs necessary features; S3 uses clustering algorithms to identify battery degradation in the data set and divide it into subsets; S4 builds a model for each dataset in the dataset; S5 performs real-time data collection; S6 uses the DTW method to calculate data similarity and selects the SOC estimation model corresponding to the highest similarity mark from the model library as the initial model; S7 uses the initial model for real-time SOC estimation; S8 uses multiple data distribution change detection methods to perform drift detection on the collected actual battery data; S9 updates the model using the current data according to step S7.

2. A data-driven real-time online lithium battery SoC accurate estimation method according to claim 1, characterized in that: The step S1 constructing the data set is expressed as: Among them, W l Indicates the charge and discharge working status; T j Indicates the temperature range; Represents the measurement data set under the above operating and temperature conditions, including charge and discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number, and battery capacity.

3. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 1 is characterized in that: The step S2 specifically processes the errors and missing data in each data set and constructs the necessary features using a feature construction method, which includes statistical quantity construction, data transformation, time series analysis, signal processing, or directly using physical quantities as features.

4. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 1, characterized in that: The step S3 uses a state recognition algorithm to identify whether the battery is decaying and the decay stage. If a data set corresponds to two battery decay stages, the data set is divided into two subsets corresponding to the two stages.

5. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 1, characterized in that: The data collected in step S5 include energy storage battery status data and environmental operating condition data, specifically including battery charge and discharge current, battery voltage, battery temperature, ambient temperature, current charge and discharge cycle number and battery capacity, and the collection time is T.

6. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 1, characterized in that: In step S6, the actually collected data of duration T is compared with the labeled data set corresponding to each model for similarity, and the model with the highest similarity is selected as the current estimated model.

7. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 6, characterized in that: At the beginning of operation, a smaller T value is set to quickly start the model, or a model in the first stage of decay or without decay is designated as the SOC estimation model to achieve a cold start of real-time SOC estimation.

8. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 1, characterized in that: The step S7 sets a quantity threshold. When data is continuously collected and accumulated to the quantity threshold, the initial model is retrained using the actually collected data, and the newly trained model is used as the latest estimation model to perform real-time SOC estimation.

9. The data-driven real-time online lithium battery SoC accurate estimation method according to claim 7, characterized in that: Step S8 performs drift detection on the actual battery data, including the battery charge and discharge current, battery voltage, battery temperature, ambient temperature, and the current number of charge and discharge cycles, to determine whether the data distribution has changed. If the data distribution has changed, data with a duration of T starting from the change point is intercepted, and the method of step S6 is used to reselect the initial model from the model library for SOC estimation. The data distribution change detection method includes t-test, KS test, and KL divergence.

10. A data-driven real-time online lithium battery SoC accurate estimation method according to claim 8 or 9, characterized in that: The step S9 uses the method of step S7 to perform real-time SOC estimation, and uses the actual battery status and environmental operating condition data to update the model so that the model reflects the current status of the battery.

Images