CN117074955A

CN117074955A - Cloud-end correction OCV-based lithium battery state joint estimation method

Info

Publication number: CN117074955A
Application number: CN202311042952.5A
Authority: CN
Inventors: 王亚雄; 赵尚宇; 王世权; 欧凯
Original assignee: Fuzhou University
Current assignee: Fuzhou University
Priority date: 2023-08-18
Filing date: 2023-08-18
Publication date: 2023-11-17

Abstract

The invention provides a lithium battery state joint estimation method based on cloud-end correction OCV, which is based on a cloud-end cooperative estimation architecture, wherein the cloud-end cooperative estimation architecture comprises a local battery end BMS, a cloud model training and data management platform; the cloud model training process includes the steps that firstly, parameter identification and model error calculation are conducted on a local battery terminal BMS, a time segment with a model error conforming to a threshold value and a time segment exceeding the threshold value are segmented according to a set error threshold value and a time sliding window, and the time segment is uploaded to the cloud; then OCV correction and SOH estimation are carried out on the cloud end, and the final estimation result is transmitted back to a BMS (battery management system) at a local battery end for model parameter updating and capacity updating; finally estimating the SOC at the BMS of the local battery end by using an unscented Kalman filtering algorithm; according to the invention, cloud platform computing resources can be fully utilized, and the OCV is corrected in a data driving mode, so that the purpose of correcting battery model errors is achieved, the maximum available capacity of the battery is updated through SOH estimation, and the estimation accuracy of the SOC is ensured.

Description

Cloud-end correction OCV-based lithium battery state joint estimation method

Technical Field

The invention relates to the technical field of battery state estimation, in particular to a lithium battery state joint estimation method based on cloud-end correction OCV, namely a lithium battery state joint estimation method based on cloud-end correction OCV (Open Circuit Voltage, open-circuit voltage).

Background

In recent decades, lithium batteries have been widely used in many engineering fields due to their excellent characteristics of high output voltage, high energy density, high safety, and long life.

In practical application, the running State and the internal State of the lithium battery directly affect the working performance of the whole main system, and accurate estimation of State of Charge (SOC) is beneficial to more reasonable and effective management of the battery. Therefore, in order to ensure efficient, stable and safe operation of the lithium ion battery in the energy storage system, it is essential to estimate the SOC with high accuracy through the BMS (Battery Management System ).

Currently, methods commonly applied to state of charge estimation are mainly classified into a traditional ampere-hour integration method, a battery model and mechanism-based estimation method and a data-driven estimation method. However, the ampere-hour integration method is influenced by current noise and an initial SOC value, and has larger limitation; the estimation method based on the battery model depends on the battery model with high precision and the parameter identification method; the data-driven estimation method relies on a large amount of data sets and high quality data, however, the storage and calculation resources of the on-board BMS are extremely limited, and the data-driven estimation accuracy is closely related to the adjustment of the super parameters, which is a very challenging task, time-consuming and very experience-dependent. The single lithium battery state of charge estimation method is difficult to meet the high-precision requirement in practical application.

In practical engineering, the Kalman filtering estimation method based on the circuit model is widely applied due to stability and accuracy, but the battery model error directly influences the estimation accuracy of the method, so that the correction of the battery model error becomes a necessary means for ensuring the SOC estimation accuracy. And as the battery ages, the maximum available capacity of the battery also changes, and the State of Health (SOH) State of Health (SOC) State of Health) State estimation method can effectively solve the problem.

Disclosure of Invention

According to the cloud-end-correction-based lithium battery state joint estimation method, cloud platform computing resources can be fully utilized, OCV is corrected in a data driving mode, the purpose of correcting battery model errors is achieved, the maximum available capacity of a battery is updated through SOH estimation, and the estimation accuracy of SOC is guaranteed.

The invention adopts the following technical scheme.

A lithium battery state joint estimation method based on cloud-end correction OCV is based on a cloud-end cooperative estimation architecture, wherein the cloud-end cooperative estimation architecture comprises a local battery end BMS, a cloud model training and data management platform; the cloud model training process includes the steps that firstly, parameter identification and model error calculation are conducted on a local battery terminal BMS, a time segment with a model error conforming to a threshold value and a time segment exceeding the threshold value are segmented according to a set error threshold value and a time sliding window, and the time segment is uploaded to the cloud; then OCV correction and SOH estimation are carried out on the cloud end, and the final estimation result is transmitted back to a BMS (battery management system) at a local battery end for model parameter updating and capacity updating; and finally estimating the SOC at the BMS of the local battery end by using an unscented Kalman filtering algorithm.

The local battery terminal BMS comprises a main control unit BCU and a slave control unit BMU; the application layer of the main control unit BCU comprises an SOC estimation module, an equalization control module, an SOH estimation module and a charging management module, and the hardware layer of the main control unit BCU comprises a wireless communication module and a communication interface; the slave control unit BMU is used for monitoring, collecting and issuing management instructions of the master control unit BCU, wherein the current, the voltage and the temperature of the battery module are monitored and collected;

the master control unit BCU and the slave control unit BMU are communicated through a CAN bus; receiving battery voltage, current and temperature acquired by a slave control unit BMU; the SOC estimation module comprises three sub-modules of parameter identification, model error calculation, error time slice data extraction and unscented Kalman filter, the wireless communication module of the main control unit BCU uploads the acquired voltage, temperature, charge and discharge time and cycle times of the battery measurement end to the data management platform;

the cloud model training and data management platform comprises a lithium battery database, a cloud estimation module and a downlink control platform; the lithium battery database is used for receiving the data packet uploaded by the wireless communication module of the main control unit BCU in the BMS of the local battery terminal and preprocessing the data; the cloud estimation module comprises an OCV correction estimation module and an SOH estimation module, and is used for deploying a deep learning model, uploading the deep learning model to a training set and a testing set by using a battery database in the two modules respectively to train the model, and carrying out OCV correction and SOH estimation respectively; the downlink control platform is used for transmitting the OCV correction estimation and SOH estimation results of the cloud estimation module back to the local battery terminal;

The data management platform is connected with the local battery BMS through the wireless communication module of the local battery BMS and performs information interaction.

The main control unit BCU of the local battery terminal BMS comprises a wireless communication module, wherein the wireless communication module is one of a Wi-Fi module, a Bluetooth module, a 4G/5G module and an NB-IoT module, the wireless communication module performs information interaction with a cloud model training and data management platform through a network, and the uplink frequency and the downlink frequency of data are determined by different frequency bands of the selected wireless communication module;

the SOH estimation module in the main control unit BCU of the local battery terminal BMS is used for receiving the SOH estimation value transmitted back by the cloud end in a downlink mode, updating the maximum available capacity of the battery, and inputting the updated maximum available capacity into the SOC estimation module for estimating the SOC;

the downlink control platform returns the OCV correction estimation result and the SOH estimation result of the cloud estimation module to the local battery end at a frequency of one cycle of battery cyclic charge and discharge, namely, performs downlink return of OCV correction estimation and SOH estimation data when the battery is charged and discharged for one time and is kept stand;

the method comprises the steps of constructing an equivalent circuit model in an SOC estimation module, carrying out parameter identification, battery model error calculation and error time segment corresponding data extraction, wherein the parameter identification adopts a least square method with forgetting factors, the battery model error is calculated through a model output terminal voltage and an actual measurement terminal voltage, and the SOC estimation method is a Kalman filtering algorithm deployed based on the equivalent circuit model;

The error time slice corresponding data comprises an SOC, a measuring terminal voltage, a measuring terminal current and a parameter identification result corresponding to a short time slice exceeding a model error threshold value, and also comprises an SOC, a model output terminal voltage, a model output terminal current, a parameter identification result and an OCV value corresponding to a long time slice conforming to the model error threshold value; the division of the error time segment is determined by defining a model error threshold, a time sliding window and a window sliding step size;

the error time slice corresponding data comprise an SOC, a measuring terminal voltage, a measuring terminal current and a parameter identification result corresponding to a short time slice exceeding a model error threshold, and an SOC, a model output terminal voltage, a model output terminal current, a parameter identification result and an OCV value corresponding to a long time slice meeting the model error threshold, the cloud is uploaded through a wireless communication module of a main control unit BCU in a local battery terminal BMS, the time slice corresponding data meeting the error threshold is used for training a cloud OCV correction estimation model, and the time slice exceeding the error threshold corresponds to OCV to be corrected;

the battery measurement terminal voltage, temperature, charge and discharge time and cycle times are characteristic data required by cloud SOH estimation, and the cloud is uploaded through a wireless communication module of a main control unit BCU in a local battery terminal BMS;

The wireless communication module packages protocol data information uploaded by the BMU into a data packet conforming to a TCP/IP transmission protocol, the wireless communication module of the BCU in the BMS at the local battery end establishes socket communication with the cloud model training data management platform to upload the data packet to the cloud model training and data management platform, and the downlink control platform in the cloud model training and data management platform returns OCV and SOH estimation results to the local battery end in the same mode.

The training and estimating modes of the deep learning model in the cloud estimating module are as follows:

step one: the cloud estimation module in the cloud model training and data management platform deploys a deep learning model;

step two: preprocessing the characteristic data corresponding to the estimated value received from the local battery terminal in a battery database in the cloud model training and data management platform to obtain a preprocessed normal data set;

step three: carrying out normalization processing on the normal data set to obtain a training set and a testing set of a cloud deep learning model;

step four: performing dimension lifting processing on the training set and the testing set, and processing the two-dimensional data set into a three-dimensional data set;

Step five: defining the network depth, the channel number, the convolution kernel size, the expansion factor size, the loss function, the type of an optimizer, the learning rate, the single training sample size and the training times of a deep learning model deployed in a cloud;

step six: training the deep learning model deployed in the cloud based on the obtained three-dimensional training set and test set of the cloud deep learning model, and storing the training model after the training model reaches a specified number of times;

step seven: the feature data corresponding to the OCV and SOH at the moment to be estimated is used as a feature vector to be input into a deep learning model trained based on historical data to carry out correction estimation of the OCV and estimation of the SOH

The specific implementation steps of the lithium battery state joint estimation method comprise:

step S1: the slave control unit BMU of the local battery terminal BMS acquires the voltage, current and temperature of the battery cell and the battery pack in real time, uploads the voltage, the temperature, the charge and discharge time and the cycle number of the measurement terminal to the cloud model training and data management platform;

step S2: the main control unit BCU of the local battery terminal BMS receives the battery voltage and the battery current acquired by the auxiliary control unit BMU, carries out parameter identification in the SOC estimation module, then calculates a battery model error, carries out SOC estimation by using the identified model parameter and the OCV if the battery model error meets the requirement, and carries out the next step if the battery model error does not meet the requirement;

Step S3: setting a certain model error threshold, time sliding windows and step sizes in an SOC estimation module of a BMS at a local battery end, calculating a battery model error in each time sliding window, and judging whether the battery model error exceeds the threshold or not so as to divide a short time segment exceeding the model error threshold and a long time segment conforming to the model error threshold;

step S4: extracting the SOC, the voltage of the measuring terminal, the current of the measuring terminal, the parameter identification result and the SOC, the voltage of the model output terminal, the current of the model output terminal, the parameter identification result and the OCV corresponding to the short-time segment exceeding the model error threshold and uploading the SOC, the voltage of the measuring terminal, the current of the measuring terminal, the parameter identification result and the OCV to a cloud model training and data management platform;

step S5: the battery database in the cloud model training and data management platform receives and pre-processes and stores the SOC, the measuring terminal voltage, the measuring terminal current, the parameter identification result and the SOC, the model output terminal voltage, the model output terminal current, the parameter identification result and the OCV which correspond to the short-time segments which exceed the model error threshold and are uploaded by the local battery terminal BMS; training a cloud OCV correction estimation model by utilizing the preprocessed SOC, model output end voltage, model output end current, parameter identification result and OCV corresponding to the long-time segment which accords with the model error threshold, and inputting the preprocessed SOC, measurement end voltage and measurement end current corresponding to the short-time segment which exceeds the model error threshold into the cloud OCV correction estimation model to carry out OCV correction estimation;

Step S6: the battery database in the cloud model training and data management platform receives the voltage, the temperature, the charge and discharge time and the cycle times of the measuring end uploaded by the BMS of the local battery end, performs data preprocessing and storage, and inputs the preprocessed data into the cloud SOH estimation model for model training and SOH estimation;

step S7: and the OCV correction estimation result and the SOH estimation value obtained by the cloud model training and data management platform are returned to the local battery terminal through the downlink control platform. The corrected OCV is used for carrying out parameter identification in the SOC estimation module again, the SOH estimation value is used for updating the maximum available capacity in the SOH module and carrying out SOC estimation by using an unscented Kalman filtering algorithm, wherein the OCV correction estimation result is subjected to downlink feedback of the estimation result after each operation period of the battery is finished, namely the standing time, SOH is used as a long-time scale variable, the change in short time is small, the downlink feedback of the estimation result is carried out according to the actual conditions of the battery and the operation working condition, and the capacity can be set to be updated for a plurality of cycle period feedback;

the cloud model training and data management platform comprises a battery database, a cloud estimation module and a downlink control platform, and the platform provides a data visualization and man-machine interaction interface, so that a user can monitor battery state, running condition and energy utilization efficiency in real time, and battery management and scheduling strategy optimization by the user is facilitated;

The battery database is used for receiving data uploaded by the local battery terminal BMS in real time, performing cleaning processing operation and preprocessing operation, and storing the processed data into the database;

the deep learning model is deployed on a cloud estimation module, the cloud estimation module comprises an OCV estimation module and an SOH estimation module, and the two estimation modules use the same deep learning model;

the downlink control platform is used for regularly returning the cloud OCV correction result and the SOH estimated value to the local battery terminal;

the SOC estimation model used by the SOC estimation module of the master control BCU in the local battery terminal BMS is an unscented Kalman filtering model based on an equivalent circuit, and the SOC estimation module comprises a battery maximum available capacity which is updated and transmitted by the SOH estimation module;

the equivalent circuit is a Thevenin equivalent circuit model, and the state equation of the model is as follows:

wherein U is _p Representing the polarization voltage, U _L Represents terminal voltage, I _L Representing end current, U _oc Represents the open circuit voltage, Δt represents the sampling interval, R ₀ Represents ohmic internal resistance, R ₁ Represents internal resistance of polarization, C ₁ Representing a polarization capacitance;

the parameter identification method of the equivalent circuit is a least recursive square method with forgetting factors;

The deep learning model deployed by the cloud estimation module is a TCN (Temporal Convolutional Network, time convolutional network) model. The TCN model comprises an input layer, a hidden layer and an output layer, wherein the hidden layers are connected through a residual error module, the residual error module comprises an expansion causal convolution, weight regularization, a ReLU function, a Dropout layer and one-time residual error operation, the extraction of time sequence features is realized through the convolution module, and the gradient disappearance or gradient explosion problem can be avoided simultaneously by using the residual error module;

the SOC estimation module of the main control unit BCU in the local battery terminal BMS comprises three sub-modules of parameter identification, model error calculation and error time segment data extraction and unscented Kalman filter.

The parameter identification process in the step S2 comprises the steps of collecting battery voltage, current and temperature data at different moments through a slave control unit BMU in a local battery terminal BMS; preprocessing the acquired battery data, and removing abnormal data; carrying out parameter identification on the Thevenin equivalent circuit model by using the recursive least square FFRLS with forgetting factors by using the preprocessed acquired data;

further, the second equation in the state equation of the formula (1) is subjected to laplace transformation, bilinear transformation, and inverse transformation of Z to obtain:

U _L (k)＝U _OC (k)-a ₁ [U _L (k-1)-U _OC (k-1)]+a ₂ I _L (k)+a ₃ I _L (k) (2)

Wherein k represents time k, a ₁ 、a ₂ 、a ₃ Representing expressions concerning ohmic internal resistance, polarized capacitance, the expressions are as follows:

equation (2) can be written as follows:

wherein,and θ (k) are respectively represented by the following formula:

by recursive least square method with forgetting factorThe least recursive part is as follows:

wherein P represents a covariance matrix, and lambda represents a forgetting factor;

according to the predicted quantity in equation (6)Inverse solving R from equation (3) ₀ ，R ₁ And C ₁ And completing the parameter identification.

In step S4, the OCV corresponding to the time slice that meets the model error threshold is obtained by SOC polynomial fitting, and the apparent is an accurate OCV value, which is used for training the cloud OCV correction deep learning model.

In step S7, the method for estimating the SOC by the unscented kalman filter algorithm specifically includes:

based on the Thevenin equivalent circuit model after parameter identification, constructing the circuit state matrix:

wherein C is _a Representing the maximum available capacity of the battery, updated by SOH;

based on the constructed circuit state matrix, estimating the SOC by an unscented Kalman filtering algorithm:

the first step, initializing a state quantity X and a covariance P;

second, the state quantity is initialized according to the formula (8)And calculating the value of each Sigma sampling point according to UT conversion:

Where n represents the state quantity dimension and λ represents the scale;

thirdly, performing one-step prediction on 2n+1 Sigma sampling points, wherein i=2n+1;

X ⁽ⁱ⁾ (k+1|k)＝A _k X ⁽ⁱ⁾ (k|k-1)+B _k u _k +w _k (10)

wherein A is _k And B _k Represents the system state matrix at the moment k, w _k Representing the measurement noise at time k;

further, calculating the mean value and covariance matrix prediction of the system state vector X at the last moment;

wherein mu represents a scaling scale parameter for reducing the estimation error of a nonlinear system, i represents an ith sampling point, m represents a mean value, c represents covariance, and a parameter beta is more than or equal to 0, and a controls the distribution of sampling points X;

further, according to the predicted value, carrying out UT conversion to obtain a new Sigma sampling point of the state quantity X;

step four, according to the new sampling point of the state vector X obtained in the previous step, bringing the new sampling point into an observation equation of the system, so as to obtain an observation value predicted by the system, wherein the observation value is shown in the following formula;

U _L (k+1|k)＝U _oc (k+1|k)-[0 1]X ⁽ⁱ⁾ (k+1|k)-I(k+1|k)·R ₀ +v (15)

wherein v represents the process noise of the observation equation;

fifthly, calculating and obtaining a prediction mean and covariance of the system observance according to the observance of the Sigma sampling points obtained in the last step;

wherein the observed value Y is U _L ，X _k R is measurement noise and is set to be a smaller value for the new sampling point obtained in the last step;

step six, the calculation result of the previous step is put into the above formula to update the gain matrix;

Seventh, the gain matrix updated in the last step is brought into the updated state quantity and covariance matrix;

and step S5, training and estimating the OCV correction estimation model according to the following specific modes:

step S51: deploying a TCN model in a cloud OCV correction estimation module in a cloud model training and data management platform;

step S52: the method comprises the steps that SOC, measurement terminal voltage, measurement terminal current and parameter identification results corresponding to short-time fragments exceeding a model error threshold received from a local battery terminal and SOC, model output terminal voltage, model output terminal current, parameter identification results and OCV corresponding to long-time fragments meeting the model error threshold are preprocessed in a battery database in a cloud model training and data management platform, and a preprocessed normal data set is obtained;

step S53: carrying out normalization processing on the normal data set to obtain a training set and a testing set of the cloud TCN model, wherein the training set is the SOC, the model output end voltage, the model output end current, the parameter identification result and the OCV corresponding to the preprocessed long-time segment which accords with the model error threshold, and the testing set is the SOC, the measuring end voltage, the measuring end current and the parameter identification result corresponding to the preprocessed time segment which exceeds the model error threshold;

Step S54: performing dimension lifting processing on the training set and the testing set, and processing the two-dimensional data set into a three-dimensional data set;

step S55: defining the network depth, the channel number, the convolution kernel size, the expansion factor size, the loss function, the optimizer type and the learning rate, the single training sample size and the training times of a TCN model deployed in a cloud;

step S56: training the TCN model deployed in the cloud based on the three-dimensional training set and the testing set of the TCN model, and storing the training model after the training model reaches a specified number of times;

step S57: and taking the time SOC to be estimated, the voltage of the measuring terminal, the current of the measuring terminal and the parameter identification result as feature vectors, and inputting the feature vectors into a deep learning model trained based on historical data to estimate the OCV.

The SOH estimation model in step S6 is specifically trained and estimated as follows:

step S61: deploying a TCN model in a cloud SOH estimation module in a cloud model training and data management platform;

step S62: preprocessing the voltage, temperature, charge and discharge time and cycle times of a measuring end received from a local battery end in a battery database in a cloud model training and data management platform to obtain a preprocessed normal data set;

Step S63: carrying out normalization processing on the normal data set to obtain a training set and a testing set of a cloud TCN model;

step S64: performing dimension lifting processing on the training set and the testing set, and processing the two-dimensional data set into a three-dimensional data set;

step S65: defining the network depth, the channel number, the convolution kernel size, the expansion factor size, the loss function, the optimizer type and the learning rate, the single training sample size and the training times of a TCN model deployed in a cloud;

step S66: training the TCN model deployed in the cloud based on the three-dimensional training set and the testing set of the TCN model, and storing the training model after the training model reaches a specified number of times;

step S67: the voltage, temperature, charge and discharge time and cycle times of a measuring terminal at the moment to be estimated are taken as feature vectors to be input into a deep learning model trained based on historical data for SOH estimation;

the step S62 of preprocessing the data such as voltage, current, temperature and the like uploaded by the local battery terminal includes cleaning and denoising the data and compensating for the deficiency.

The training set and the testing set of the cloud-end OCV correction estimation model comprise input features and labels, namely SOC, measurement end voltage, measurement end current and OCV reference values corresponding to the features at the current moment;

The SOC calculation formula is as follows:

wherein SOC is ₀ Represents the initial value of the battery SOC, C _a Represents the maximum capacity of the battery, eta represents the coulombic efficiency, I represents the charge-discharge current of the battery at the current moment, t ₀ And t ₁ Respectively representing the charge and discharge starting time of the battery;

the training set and the testing set of the cloud SOH estimation model comprise input features and labels, namely measuring terminal voltage, temperature, charge and discharge time, cycle times and SOH corresponding to the features at the moment;

the SOH calculation formula is as follows:

step S66, the size of the three-dimensional data set is [ batch size, sampling time and characteristic quantity ], wherein the batch size is the number of data which are sampled from the data set each time and input into the deep learning model for training, the sampling time is the characteristic data sampling time corresponding to each SOH value, and the characteristic quantity is the sum of the types and the quantities of the voltage, the temperature, the charge and discharge time and the cycle number of the measuring terminal acquired in the unit sampling time;

step S66, saving the training model, specifically saving the parameters of the trained model, including the super parameters of the input layer, the hidden layer and the output layer which are initially defined, and the updated weight and threshold; the weight and the threshold are optimized and updated by an Adam optimizer; the maximum available capacity is updated by equation (23).

The invention provides a cloud-end-corrected OCV-based lithium battery state joint estimation method, which is based on a cloud-end interaction platform and comprises the steps of carrying out parameter identification and model error calculation on a local battery BMS end, dividing a time segment with a model error conforming to a threshold value and a time segment exceeding the threshold value according to a set error threshold value, extracting an SOC (system on chip), a model output end voltage, a model output end current, a parameter identification result and an OCV training cloud deep learning model corresponding to the time segment conforming to the threshold value, extracting an SOC, a measuring end voltage, a measuring end current and a parameter identification result corresponding to the time segment exceeding the threshold value, and inputting the OCV with larger correction estimation error into the trained deep learning model. And finally, returning the corrected OCV to a local battery terminal BMS (battery management system) terminal for correcting and updating the parameter identification result, inputting the corrected OCV and the parameter identification result into a unscented Kalman filter, carrying out SOH estimation at a cloud end, returning the corrected OCV and the parameter identification result to the local terminal for updating the maximum available capacity, and carrying out final SOC estimation.

Compared with the prior art, the invention has the following advantages: aiming at the technical difficulties and industry requirements of lithium batteries, the invention constructs a lithium battery state-of-charge estimation framework based on cloud-end correction OCV from the aspects of software and hardware by combining the Internet technology, the mobile communication technology and the artificial intelligence technology, fully utilizes the existing Internet technology, the mobile communication technology and the artificial intelligence technology, fully utilizes cloud platform computing resources, corrects OCV and estimates SOH in a data driving mode, reduces the requirement on high calculation force of an on-board BMS, achieves the aim of correcting battery model errors, and ensures the estimation accuracy of SOC.

Drawings

The invention is described in further detail below with reference to the attached drawings and detailed description:

FIG. 1 is a schematic diagram of a lithium battery state joint estimation architecture based on cloud-end correction OCV according to an embodiment of the present invention;

fig. 2 is a schematic diagram of a lithium battery state joint estimation flow based on cloud-end correction OCV according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a cloud-end correction OCV and model parameter flow according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a unscented Kalman filtering algorithm for estimating SOC according to an embodiment of the invention;

FIG. 5 is a schematic diagram of the TCN-modified OCV structure in accordance with an embodiment of the present invention;

FIG. 6 is a schematic diagram of the TCN estimation SOH structure principle of the embodiment of the invention;

FIG. 7 is a schematic diagram of model error calculation and error time segment segmentation at 10deg.C according to an embodiment of the present invention;

FIG. 8 is a graph showing the comparison of the model output voltages before and after OCV correction at 10deg.C in accordance with an embodiment of the present invention;

FIG. 9 is a schematic diagram of the result of estimating the DST working condition SOC by the unscented Kalman filtering algorithm at 10 ℃ according to the embodiment of the invention;

FIG. 10 is a schematic diagram of model error calculation and error time segment segmentation at 25deg.C according to an embodiment of the invention;

FIG. 11 is a graph showing the comparison of the model output voltages before and after OCV correction at 25℃in accordance with an embodiment of the present invention;

FIG. 12 is a schematic diagram of the result of estimating the DST working condition SOC by the unscented Kalman filtering algorithm at 25 ℃ according to the embodiment of the invention;

fig. 13 is a schematic diagram of the result of TCN estimation SOH according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. It is apparent that the described embodiments are some, but not all embodiments of the present invention, and that all other embodiments obtained by persons of ordinary skill in the art without making creative efforts based on the embodiments in the present invention are within the protection scope of the present invention.

As shown in fig. 1, the present embodiment provides a method architecture for jointly estimating states of lithium batteries based on cloud-end correction OCV, which specifically includes the following contents:

the lithium battery state joint estimation method architecture based on cloud-end correction OCV comprises a local battery end BMS and a cloud model training and data management platform;

the local battery terminal BMS comprises a main control unit BCU and a slave control unit BMU. The main control unit BCU application layer comprises an SOC estimation module, an SOH estimation module, a charging management module and other management modules, and the hardware layer comprises a wireless communication module, an equalization control module, a communication interface and the like; the slave control unit BMU is used for monitoring and collecting the current, the voltage and the temperature of the battery module and issuing a management instruction of the master control unit BCU.

The cloud model training and data management platform comprises a lithium battery database, a cloud estimation module and a downlink control platform, and the platform is used for carrying out information interaction with a local battery terminal BMS mainly through a wireless communication module of the local battery terminal BMS.

In this embodiment, the wireless communication mode for information interaction between the main control unit BCU in the local battery terminal BMS and the cloud model training and data management platform is a 5G communication module, the data uplink frequency range based on the 5G communication module is 885-909MHz, and the data downlink frequency range is 930-954MHz. The local battery terminal BMS is simulated by adopting an Injetton Nano development board, and the specification is as follows: the CPU independent processor core number of 1.43GHz is 4, the NVIDIA Maxwell GPU processor core number is 128, the video memory is 4GB, the memory is 16GB, and 473 giga floating point calculation power can be provided. The battery database in the cloud model training and data management platform selects MySQL database, and the hardware specification of the MySQL database is as follows: the number of cores of the CPU independent processor with the memory of 500GB and the memory of 2.5GHz is 4, and the memory is 8GB; the cloud estimation module adopts an AI computing architecture based on a Pytorch environment, and the hardware specification of the cloud estimation module is as follows: the number of cores of the CPU independent processor with the storage memory of 256GB and 2.5GHz is 8, the number of cores of the GPU independent processor is 128, and the memory of 128GB.

The specific implementation steps of this embodiment are as follows:

s1: the voltage, current and temperature acquisition module of the BMS at the local battery end acquires the voltage, current and temperature of the battery in real time and uploads the battery to a main control unit BCU (Jetson Nano development board);

s2: the method comprises the steps that a main control unit BCU of a local battery terminal BMS receives collected measurement terminal voltage, current, temperature, charge-discharge time and cycle times, packages the measurement terminal voltage, the current, the temperature, the charge-discharge time and the cycle times into data packets conforming to a TCP/IP transmission protocol, and a wireless communication module of the main control unit BCU in the local battery terminal BMS establishes socket communication with a cloud model training data management platform to upload the data packets to the cloud model training and data management platform;

s3: the method comprises the steps that a main control unit BCU of a local battery terminal BMS receives battery voltage, current and temperature data acquired by an acquisition module, parameter identification, model error calculation, error time segment data extraction and unscented Kalman filtering algorithm estimation SOC are carried out in an SOC estimation module, an SOC reference value, a measuring terminal voltage, a measuring terminal current, a parameter identification result and an SOC reference value, a model output terminal voltage, a model output terminal current, a parameter identification result and OCV data information which correspond to a long-time segment which accords with a model error threshold value and are output by the SOC estimation module are packaged into a data packet which accords with a TCP/IP transmission protocol, and a wireless communication module of the main control unit BCU in the local battery terminal BMS establishes socket communication with a cloud model training data management platform, and uploads the data packet to the cloud model training and data management platform;

S4: the battery database in the cloud model training and data management platform receives and analyzes the data packet, and performs preprocessing operations such as data cleaning, noise removal, value deficiency compensation and the like; uploading the training set and the testing set obtained by preprocessing to a cloud estimation module, respectively training an OCV correction estimation model and an SOH estimation model, and respectively carrying out OCV correction estimation and SOH estimation according to the uploading data at the current moment;

s5: and the OCV correction estimation result and the SOH estimation value obtained by the cloud model training and data management platform are returned to the energy storage battery end through the downlink control platform. And carrying out parameter identification in the SOC estimation module again by using the corrected OCV, updating the maximum available capacity in the SOH module by using the SOH value returned by the cloud, and finally estimating the SOC by using the unscented Kalman filtering algorithm.

In this embodiment, the SOC estimation module of the main control unit in the energy storage battery end BMS in step S3 includes three sub-modules of parameter identification, model error calculation, error time slice data extraction, and unscented kalman filter.

The parameter identification process specifically comprises the following steps:

the battery voltage, current, temperature and other data at different moments are collected through a local battery terminal BMS;

Preprocessing the acquired battery data, and removing abnormal data;

and carrying out parameter identification on the Thevenin equivalent circuit model by using the recursion least square method (FFRLS) with forgetting factors by using the preprocessed acquired data.

The state equation of the Thevenin equivalent circuit model is as follows:

further, the second equation in the state equation (1) is subjected to laplace transformation, bilinear transformation, and inverse transformation of Z to obtain:

equation (2) can be written as follows:

wherein,and θ (k) are respectively represented by the following formula:

where P represents the covariance matrix and λ represents the forgetting factor.

According to the predicted amount in (6)Inverse solving R from equation (3) ₀ ，R ₁ And C ₁ And completing the parameter identification.

The model error calculation and error time slice data extraction specifically comprise:

and carrying out error calculation on the Thevenin equivalent circuit model after parameter identification, setting a model error threshold, time sliding windows and step sizes, calculating the model error in each time sliding window, judging whether the model error exceeds the threshold, dividing a time segment conforming to the model error threshold and a time segment exceeding the model error threshold, and extracting the SOC, the voltage of a measuring end, the current of the measuring end, the parameter identification result and the SOC, the voltage of a model output end, the current of the model output end, the parameter identification result and the OCV corresponding to the long-time segment conforming to the model error threshold, which correspond to the time segment exceeding the model error threshold.

The OCV corresponding to the time slice meeting the model error threshold is obtained through SOC polynomial fitting, and can be regarded as an accurate OCV value for training the cloud OCV correction deep learning model.

The unscented Kalman filter algorithm for estimating the SOC specifically includes:

wherein C is _a Representing the maximum available capacity of the battery, updated by SOH.

the first step, initializing a state quantity X and a covariance P;

a second step of initializing the state quantity according to (9)And calculating the value of each Sigma sampling point according to UT conversion:

where n represents the state quantity dimension and λ represents the scale.

wherein A is _k And B _k Represents the system state matrix at the moment k, w _k Representing the noise of the measurement at time k.

wherein mu represents a scale parameter for reducing the estimation error of a nonlinear system, i represents an ith sampling point, m represents a mean value, c represents covariance, and the parameter beta is more than or equal to 0, and a controls the distribution of the sampling points X.

and fourthly, according to the new sampling point of the state vector X obtained in the last step, bringing the new sampling point into an observation equation of the system, thereby obtaining an observation value predicted by the system.

U _L (k+1|k)＝U _oc (k+1|k)-[0 1]X ⁽ⁱ⁾ (k+1|k)-I(k+1|k)·R ₀ +v (15)

Where v represents the process noise of the observation equation.

And fifthly, calculating and obtaining a prediction mean and covariance of the system observables according to the observables of the Sigma sampling points obtained in the last step.

/>

Wherein the observed value Y is U _L ，X _k For the new sample point obtained in the previous step, R is the measurement noise and is generally set to a smaller value.

in this embodiment, the deep learning model in step S4 is specifically trained and estimated as follows:

s41: deploying a TCN model in a cloud estimation module in a cloud model training and data management platform;

s42: preprocessing the OCV correction corresponding feature data and SOH estimation corresponding feature data received from a local battery terminal in a battery database in a cloud model training and data management platform to obtain a preprocessed normal data set;

S43: carrying out normalization processing on the normal data set to obtain a training set and a testing set of a cloud TCN model;

s44: performing dimension lifting processing on the training set and the testing set, and processing the two-dimensional data set into a three-dimensional data set;

s45: defining the network depth, the channel number, the convolution kernel size, the expansion factor size, the loss function, the optimizer type and the learning rate, the single training sample size and the training times of a TCN model deployed in a cloud;

s46: training the TCN model deployed in the cloud based on the three-dimensional training set and the testing set of the TCN model, and storing the training model after the training model reaches a specified number of times;

s47: and inputting the feature vector of the moment to be estimated into a deep learning model trained based on historical data to estimate OCV and SOH.

In this example, the feasibility of the method provided by the invention was verified by dynamic regime experiments at two temperatures, DST regime at 10 ℃ and 25 ℃, respectively. The SOH estimation method is verified in the public data set limited by the actual experimental conditions and experimental period, and the embodiment does not change the strategy of the invention, and only the accuracy of the estimation method is described.

The DST working conditions at 10 ℃ and 25 ℃ are tested by adopting ternary lithium batteries of US18650VTC6 model of SONY company, the rated capacity is 3Ah, and the upper and lower cut-off voltages are 4.2V and 2.5V respectively. The disclosed data set for verifying the SOH estimation accuracy adopts a performance degradation data set of the NASA lithium ion power battery, battery 5 data is used as a training set to train a cloud deep learning model, battery 7 data is used as a test set, the rated capacities of the battery 5 and the battery 7 are 2Ah, the battery 5 and the battery 7 are charged in a Constant Current (CC) mode of 1.5A at room temperature until the battery voltage reaches 4.2V, then the battery is continuously charged in a Constant Voltage (CV) mode until the charging current is reduced to 20mA, the battery is discharged in a Constant Current (CC) level of 2A until the battery voltages of the battery 5 and the battery 7 are respectively reduced to 2.7V and 2.2V, and the experiment is stopped when the battery rated capacity is reduced by 30% by repeating the charging and discharging processes.

According to the method and steps of the embodiment, model error calculation and error time segment division are carried out on DST working conditions under two temperatures, OCV with larger error is corrected by determining data corresponding to the error threshold time segment as a training set, the model error threshold RMSE is set to be 50mV, the time sliding window is set to be one-fourth working condition duration, the step length is the same as that of the time sliding window, and if the threshold value is exceeded in the time sliding window, the OCV is considered to be corrected in the time segment. Model error calculation and error time segment segmentation are shown in fig. 7 and 10, the correction of the model error before and after the correction of the OCV is compared with that of fig. 8 and 11, the correction of the model root mean square error RMSE of a time segment part exceeding an error threshold under the working condition of 10 ℃ DST is reduced to 12.3mV, the correction of the model root mean square error RMSE under the working condition of the whole time segment DST is reduced to 51.11mV, and the correction is reduced to 25.70mV; the root mean square error correction of the partial model of the time segment exceeding the error threshold under the DST working condition at 25 ℃ is reduced to 12.6mV, the RMSE correction of the model under the DST working condition of the whole time segment is 37.98mV before correction, and the RMSE correction is reduced to 25.93mV; parameter identification is carried out again by the corrected OCV, the maximum available capacity of the battery is updated according to the SOH estimated value returned by the cloud, and final SOC estimation is carried out by using an unscented Kalman filter algorithm, wherein the estimated result is shown in fig. 9 and 12, the unscented Kalman filter algorithm estimates that the average absolute error MAE of the SOC is only 0.57% under the 10 ℃ DST working condition, and the unscented Kalman filter algorithm estimates that the average absolute error MAE of the SOC is only 0.22% under the 25 ℃ DST working condition; the estimated value is transmitted back to the maximum available capacity of the updated battery by using TCN to estimate SOH at the cloud, and the estimated result is shown in FIG. 13, wherein the absolute average error of SOH estimation is only 0.78%. As can be seen from the verification result of the embodiment of the invention, the error correction of the battery model and the joint estimation of the SOC and the SOH have better effects.

In summary, the invention aims at the technical difficulties and the industry demands of the lithium battery industry, combines the internet technology, the mobile communication technology and the artificial intelligence technology, constructs a lithium battery joint state estimation architecture based on cloud-end correction OCV from the software and hardware level, describes the method and the steps of lithium battery joint state estimation based on cloud-end correction OCV in detail, fully utilizes the advantages of the existing internet technology, the mobile communication technology and the artificial intelligence technology, plays the advantages of cloud computing resources and data driving algorithms, carries out OCV correction estimation and SOH estimation in a data driving mode, reduces the requirement on high calculation force of an on-board BMS, achieves the aim of correcting battery model errors at the same time, and ensures the estimation accuracy of SOC.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the invention in any way, and any person skilled in the art may make modifications or alterations to the disclosed technical content to the equivalent embodiments. However, any simple modification, equivalent variation and variation of the above embodiments according to the technical substance of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A lithium battery state joint estimation method based on cloud-end correction OCV is based on a cloud-end cooperative estimation architecture, and is characterized in that: the cloud-end collaborative estimation architecture comprises a local battery end BMS, a cloud model training platform and a data management platform; the cloud model training process includes the steps that firstly, parameter identification and model error calculation are conducted on a local battery terminal BMS, a time segment with a model error conforming to a threshold value and a time segment exceeding the threshold value are segmented according to a set error threshold value and a time sliding window, and the time segment is uploaded to the cloud; then OCV correction and SOH estimation are carried out on the cloud end, and the final estimation result is transmitted back to a BMS (battery management system) at a local battery end for model parameter updating and capacity updating; and finally estimating the SOC at the BMS of the local battery end by using an unscented Kalman filtering algorithm.

2. The cloud-terminal correction OCV-based lithium battery state joint estimation method of claim 1, wherein the method is characterized by comprising the following steps of: the local battery terminal BMS comprises a main control unit BCU and a slave control unit BMU; the application layer of the main control unit BCU comprises an SOC estimation module, an equalization control module, an SOH estimation module and a charging management module, and the hardware layer of the main control unit BCU comprises a wireless communication module and a communication interface; the slave control unit BMU is used for monitoring, collecting and issuing management instructions of the master control unit BCU, wherein the current, the voltage and the temperature of the battery module are monitored and collected;

3. The method for jointly estimating states of lithium batteries based on cloud-end correction OCV according to claim 2, which is characterized in that: the main control unit BCU of the local battery terminal BMS comprises a wireless communication module, wherein the wireless communication module is one of a Wi-Fi module, a Bluetooth module, a 4G/5G module and an NB-IoT module, the wireless communication module performs information interaction with a cloud model training and data management platform through a network, and the uplink frequency and the downlink frequency of data are determined by different frequency bands of the selected wireless communication module.

4. The method for jointly estimating states of lithium batteries based on cloud-end correction OCV according to claim 2, which is characterized in that: the SOH estimation module in the main control unit BCU of the local battery terminal BMS is used for receiving the SOH estimation value transmitted back by the cloud end in a downlink mode, updating the maximum available capacity of the battery, and inputting the updated maximum available capacity into the SOC estimation module for estimating the SOC;

5. The method for jointly estimating states of lithium batteries based on cloud-end correction OCV according to claim 2, which is characterized in that: the training and estimating modes of the deep learning model in the cloud estimating module are as follows:

step seven: and taking the characteristic data corresponding to the OCV and SOH of the moment to be estimated as the characteristic vector, and inputting the characteristic vector into a deep learning model trained based on historical data to carry out correction estimation of the OCV and estimation of the SOH.

6. The method for jointly estimating states of lithium batteries based on cloud-end correction OCV according to claim 5, which is characterized in that: the specific implementation steps of the lithium battery state joint estimation method comprise:

The deep learning model deployed by the cloud estimation module is a TCN model.

The TCN model comprises an input layer, a hidden layer and an output layer, wherein the hidden layers are connected through a residual error module, the residual error module comprises an expansion causal convolution, weight regularization, a ReLU function, a Dropout layer and one-time residual error operation, the extraction of time sequence features is realized through the convolution module, and the gradient disappearance or gradient explosion problem can be avoided simultaneously by using the residual error module;

7. The cloud-terminal correction OCV-based lithium battery state joint estimation method of claim 6, wherein the method comprises the following steps of: the parameter identification process in the step S2 comprises the steps of collecting battery voltage, current and temperature data at different moments through a slave control unit BMU in a local battery terminal BMS; preprocessing the acquired battery data, and removing abnormal data; carrying out parameter identification on the Thevenin equivalent circuit model by using the recursive least square FFRLS with forgetting factors by using the preprocessed acquired data;

equation (2) can be written as follows:

wherein,and θ (k) are respectively represented by the following formula:

8. The cloud-terminal correction OCV-based lithium battery state joint estimation method of claim 6, wherein the method comprises the following steps of: in step S4, the OCV corresponding to the time slice that meets the model error threshold is obtained by SOC polynomial fitting, and the apparent is an accurate OCV value, which is used for training the cloud OCV correction deep learning model.

9. The cloud-terminal correction OCV-based lithium battery state joint estimation method of claim 6, wherein the method comprises the following steps of: in step S7, the method for estimating the SOC by the unscented kalman filter algorithm specifically includes:

the first step, initializing a state quantity X and a covariance P;

where n represents the state quantity dimension and λ represents the scale;

X ⁽ⁱ⁾ (k+1|k)＝A _k X ⁽ⁱ⁾ (k|k-1)+B _k u _k +w _k (10)

U _L (k+1|k)＝U _oc (k+1|k)-[0 1]X ⁽ⁱ⁾ (k+1|k)-I(k+1|k)·R ₀ +v (15)

Wherein v represents the process noise of the observation equation;

P(k+1|k+1)＝P(k+1|k)-K(k+1|k)P _YKYk K ^T (k+1|k) (21)；

10. The cloud-terminal correction OCV-based lithium battery state joint estimation method of claim 8, wherein the method comprises the following steps of: the SOH estimation model in step S6 is specifically trained and estimated as follows:

the SOC calculation formula is as follows:

the SOH calculation formula is as follows:

step S66, the size of the three-dimensional data set is [ batch size, sampling time and characteristic quantity ], wherein the batch size is the number of data which are sampled from the data set each time and input into the deep learning model for training, the sampling time is the characteristic data sampling time corresponding to each SOH value, and the characteristic quantity is the sum of the types and the quantities of the voltage, the temperature, the charge and discharge time and the cycle number of the measuring terminal acquired in the unit sampling time; step S66, saving the training model, specifically saving the parameters of the trained model, including the super parameters of the input layer, the hidden layer and the output layer which are initially defined, and the updated weight and threshold; the weight and the threshold are optimized and updated by an Adam optimizer; the maximum available capacity is updated by equation (23).