A digital closed-loop distributed predictive control method for battery simulation system

Abstract

The present invention discloses a digital closed-loop control method for a battery simulation system based on distributed model prediction, which is used for the evaluation of a battery management system BMS of an electric energy storage type. The method includes AD high-frequency acquisition of analog battery voltage, selecting a linear prediction filter LPC to remove high-frequency noise of the acquisition signal, and forming a denoised time series signal; extracting multivariate time-frequency features of the denoised time series signal as input, training a small sample domain transfer neural network to realize intelligent early warning of the health status of the battery simulation circuit; for a healthy circuit, using the sliding mean of the denoised time series signal as feedback, and using a distributed model predictive control DMPC algorithm to realize high-speed, high-precision synchronous tracking control of multi-channel analog voltage setting values, and the controller output is converted by DA as a simulated battery voltage, and finally realizing high-speed, high-precision simulation of the battery voltage. The present application ensures the accuracy and rapidity of voltage regulation under the premise of circuit health, and improves voltage simulation accuracy and circuit reliability.

Description

Technical Field

[0001] The present invention belongs to the field of big data intelligent control, and specifically relates to a digital closed-loop distributed predictive control method for a battery simulation system. Background Art

[0002] BMS for power storage needs to simulate batteries to test their various functions and performance indicators, among which single cell voltage testing is an important task of BMS. Battery simulation currently mainly adopts the differential amplifier simulator method. Affected by temperature and chip differences, the performance of different hardware varies greatly, and the accuracy of the simulated battery cannot be guaranteed. If the simulated hardware has unhealthy circuit conditions such as cold solder joints, aging, and damage, it cannot even guarantee the stable output of voltage. In order to overcome the above hardware simulation problems, the present invention proposes a digital closed-loop battery voltage prediction and control method based on high-frequency sampling, which can timely detect the abnormal health status of the circuit and only perform voltage regulation and control on healthy circuits. It is of great significance to ensure the stable operation of the circuit and extend the service life of the hardware, while ensuring the accuracy of voltage simulation and the rapidity of regulation. Summary of the invention

[0003] The object of the present invention is to provide a digital closed-loop distributed predictive control method for a battery simulation system, which can ensure fast and high-precision regulation of the simulated voltage of a healthy circuit.

[0004] In order to solve the above technical problems, the present invention provides a digital closed-loop distributed predictive control method for a battery simulation system, comprising:

[0005] Digital closed-loop control, in order to ensure the accuracy and efficiency of BMS performance testing for power energy storage, considering the shortcomings of the analog error amplifier-based method, which leads to inaccurate voltage and slow regulation due to hardware discreteness, a digital closed-loop voltage regulation based on AD high-speed sampling is constructed;

[0006] The high-speed AD acquisition data of the analog voltage forms a time series signal, and a data-driven linear prediction filter LPC is used to remove the high-frequency noise interference of the time series signal to form a denoised time series signal;

[0007] Calculate the multivariate secondary features of the denoised time series signal, calculate the spectrum value of the time series signal, use the spectral kurtosis algorithm to extract the abnormal high-frequency spectrum sequence as the first feature, calculate various time domain statistical information of the denoised time series signal as the second feature, and extract the Gaussian distribution form of the denoised time series signal as the third feature;

[0008] Based on three types of feature information, a small sample domain migration deep neural network is used to achieve health status warning of battery simulation circuits;

[0009] For the circuit in good health, the sliding mean of the denoised timing signal is calculated as the feedback of the simulated battery voltage, and the distributed model predictive control algorithm DMPC is used to synchronously track and control the set values ​​of multiple single battery voltages to achieve high-speed and high-precision analog voltage switching regulation, thereby ensuring the accuracy and speed of the power storage BMS test.

[0010] Optionally, the digital closed-loop voltage regulation based on AD high-speed sampling includes:

[0011] The BMS test process for power storage is described, the purpose of battery analog voltage is described, the basic principle of analog error amplifier is introduced, the closed-loop control principle is introduced, and a digital closed-loop voltage intelligent control method is proposed based on AD high-speed sampling timing signals.

[0012] Optionally, the performing denoising filtering on the high-frequency collected time series data includes:

[0013] The sliding average model is established using the all-pole model in the linear predictive coding (LPC) method. Considering the stationarity of the sampling signal, a data-driven method is used to realize the parameter estimation of the model's linear prediction coefficients. The sampling sequence signal is filtered based on the established linear prediction model to obtain a denoised time series signal.

[0014] Optionally, the extracting multivariate secondary features of the denoised time series signal includes:

[0015] The spectrum value of the denoised time series signal is calculated, and the abnormal spike spectrum sequence is extracted using the spectral kurtosis algorithm as the first feature. Various time domain statistical information of the denoised time series signal is calculated as the second feature, and the Gaussian distribution morphology of the denoised time series signal is extracted as the third feature.

[0016] Optionally, the evaluating the health status of the battery simulation circuit based on the multi-element fusion feature includes:

[0017] The first and second features are concatenated into a time-frequency sequence and input into the GRU deep time series neural network. At the same time, the third feature image is input into the VGG-16 convolutional neural network. The features extracted by the two neural networks are input into the Transformer large model to realize feature fusion. Considering the small sample characteristics of the real unhealthy circuit state, the ternary center loss (Triplet-cetner) is used as the training objective function, and the model outputs the health status of the battery simulation circuit.

[0018] Optionally, the high-speed and high-precision voltage control based on the distributed model predictive control algorithm DMPC includes:

[0019] For circuits in good health, the hardware differences of each simulated battery are taken into account, and a feedforward model is established for each battery to predict the DA analog voltage from the AD digital quantity. Each single-cell battery is used as a controlled subsystem, and a distributed acquisition and centralized control method is used to achieve multi-channel synchronous control, realize the rapidity and accuracy of battery voltage regulation, and improve the test efficiency of the power storage BMS.

[0020] The present invention discloses a digital closed-loop distributed predictive control method for a battery simulation system, which is used for the evaluation of a battery management system BMS for power storage. The method includes AD high-frequency acquisition of analog battery voltage, selecting a linear prediction filter LPC to remove high-frequency noise of the acquisition signal, and forming a denoised time series signal; extracting the multivariate time-frequency characteristics of the denoised time series signal as input, training a small sample domain transfer neural network to realize intelligent early warning of the health status of the battery simulation circuit; for a circuit in good health, using the sliding mean of the denoised time series signal as feedback, and using a distributed model predictive control DMPC algorithm to realize high-speed, high-precision synchronous tracking control of multi-channel analog voltage setting values, and the controller output is converted by DA as a simulated battery voltage, and finally realizing high-speed, high-precision simulation of the battery voltage. The present application ensures the accuracy and rapidity of battery voltage regulation, and improves the voltage simulation accuracy and circuit reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 A schematic flow chart of a digital closed-loop distributed predictive control method for a battery simulation system provided by the present invention;

[0023] Figure 2 The power storage BMS test principle diagram provided by the present invention;

[0024] Figure 3 The schematic diagram of the analog error amplifier provided by the present invention;

[0025] Figure 4 The digital closed-loop control principle diagram provided by the present invention;

[0026] Figure 5 This is the framework diagram of the circuit health status warning model;

[0027] Figure 6 The DMPC algorithm provided by the present invention realizes the synchronous tracking control principle diagram of multiple single-cell battery voltages. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, Figure 1 A flow chart of a digital closed-loop distributed predictive control method for a battery simulation system provided by the present invention, the method comprising:

[0030] S11: Introduction to BMS testing requirements and analog battery hardware circuits for power storage, and explanation of the principle of digital closed-loop voltage control based on high-frequency acquisition.

[0031] According to the national standard GBT34131-2023, BMS for power energy storage must have voltage, current and temperature monitoring, battery charge and discharge control, circuit abnormality protection and battery balancing control functions, among which real-time monitoring of battery voltage is the key. In order to test the voltage monitoring performance of the BMS system, a high-performance battery simulator is designed to implement performance testing of the BMS system based on the simulated voltage.

[0032] The testing principle of BMS for power storage is as follows Figure 2 , the BMS system and the controller PC simultaneously collect the voltage value of each single-cell simulated battery. To test the correctness of the BMS system monitoring voltage, the controller sends a control quantity to adjust the battery simulator voltage to evaluate the rapidity and accuracy of the BMS system voltage monitoring; the controller can also send instructions to create battery simulator imbalance, disconnection, and short circuit anomalies. When the simulated battery pack is abnormal, the BMS system needs to give a balancing instruction or report the abnormal type in time. The BMS system outputs the results and sends them to the PC controller for correctness verification. It can be seen that the fast and high-precision adjustment of the battery voltage can improve the test efficiency. At the same time, to ensure the accuracy of the test results, it is necessary to simulate the healthy operation of the battery circuit.

[0033] Battery voltage analog regulation circuit such as Figure 3 In order to achieve fast and accurate detection of power energy storage BMS, the battery simulation system based on the hardware difference feedback method calculates the corresponding digital quantity according to the target setting value, converts the digital quantity into output voltage through DA, and makes a difference between the output voltage and the target voltage generated by the push-pull regulator. The difference is mainly the voltage offset caused by hardware differences. After the difference is amplified, the MOS tube opening amount is controlled, and the target voltage output by the regulator is controlled in the form of push-pull output. The temperature drift of the analog device and the hardware distribution difference affect the accuracy of the battery simulation voltage and the rapidity of voltage regulation.

[0034] In order to improve the voltage accuracy and the rapidity of regulation, a high-speed AD is directly used to collect the digital sequence value of the DA output after the push-pull regulator voltage, and the difference is made with the PC digital setting value. Based on the digital difference, an intelligent control algorithm is selected to control the DA output voltage to form a digital closed loop, bypassing the analog difference amplifier and MOS tube, and realizing high-speed and high-precision analog voltage digital closed-loop control. The principle of analog battery voltage digital closed-loop control is as follows: Figure 4 In addition, with high-speed digital signal acquisition, characteristic information related to the health status of DA converters, push-pull circuits and other conditioning circuits can be extracted to achieve hardware health status early warning of the battery simulator.

[0035] S12: Select a linear prediction filter LPC to remove high-frequency noise in the collected signal to form a denoised time series signal.

[0036] It should be noted that the linear prediction filter LPC is introduced in this application because high-frequency noise and white noise are inevitably mixed in during AD high-speed sampling. If the collected signal is directly used for digital closed-loop control and status warning, its effect will be affected.

[0037] The basic idea of ​​LPC is to obtain the prediction coefficients of the time series signal by solving the autoregressive AR model. The present invention adopts a full-pole model and optimizes the model parameters through a data-driven method. The time series signal can be reconstructed through these LPC coefficients, thereby achieving the goal of filtering out high-frequency noise or disordered white noise. Specifically, the following steps are included.

[0038] Step 1: Establish a full-pole model as shown in equation (1), where is the filter order, are the model parameters to be optimized. The PLC method is to Solve the model parameters under the condition of , That is, the data point sampled by AD.

[0039] (Formula 1)

[0040] Step 2: To solve the model parameters, the mean square error minimization method shown in formula (2) is used. The minimum sum of squares of the difference between the predicted value and the true value is used as the objective function. The time series data collected based on AD is used as the training set. The self-supervised training method is used. The previous 1-P collected data are used as input. The value of the P+1th data point is predicted in a rolling manner. The minimum value of formula (2) is used as the goal. The latest genetic algorithm is used to mine and obtain the LPC model parameters.

[0041] (Formula 2)

[0042] Step 3: Perform real-time prediction on the sampled data based on the optimized parameters, and filter the data according to the difference between the predicted result and the true value of the sampled data. When the difference is greater than a certain set threshold, it means that the sampled value is abnormal. The abnormal data is removed, and the missing data is filled using the 3-spline interpolation method. The filtered sampled data is the denoised time series signal.

[0043] S13: Extracting multivariate secondary features from denoised time series signals.

[0044] It should be noted that the denoised timing signal contains characteristic information of the circuit health status, and various signal processing methods are needed to extract the corresponding features of different states. The feature extraction process mainly includes the following contents.

[0045] The first feature extraction is to calculate the spectrum value of the denoised time series signal by the FFT algorithm. If the time series contains an unhealthy circuit state, the spectrum value of the corresponding state will appear abnormally high in the spectrum value; the spectral kurtosis algorithm calculates the high-order statistics of the kurtosis value of each spectral line. The spectral kurtosis mainly extracts the transient abnormal surge in the time series signal, and can effectively identify the transient surge and its distribution in the frequency band. The present invention calculates the first-order discrete difference of the time series signal to obtain the slope of the change. The slope is equivalent to the first-order derivative and represents the law of data change; at the same time, the second-order discrete difference is calculated to obtain the curvature of the time series change. The curvature is equivalent to the second-order derivative and represents the positive and negative extreme points. For the spectrum sequence signal, when the slope is greater than zero, it means that the spectrum curve is increasing. The larger the value, the faster the spectrum changes, that is, the larger the spectrum kurtosis, and the positive and negative curvature of the corresponding position represents the direction of the spectrum surge value, and vice versa. When the slope and curvature are both 0, it means that the sampled data has no surge and belongs to a stable DC signal. The difference calculation formulas are shown in formula (3). The second-order difference is the discrete difference of the first-order difference sequence. The first five values ​​with the largest absolute values ​​of slope and curvature are obtained to represent the abnormal surge spectrum as the first feature.

[0046] (Formula 3)

[0047] The second feature extraction is to calculate various statistical information of the time series signal, which mainly includes two parts: (1) Gaussian kernel 1-4 order features, the first order feature is the mean, the second order feature is the variance, the third order feature is the skewness, and the fourth order feature is the kurtosis; (2) basic statistical features, calculate various statistical information such as maximum value, minimum value, median, 95 and 05 quantiles, effective value, standard deviation, peak value, margin, etc. The vector formed by concatenating the four values ​​of the 1-4 order Gaussian kernel and all the time domain information of the basic statistics is used as the second feature.

[0048] Based on the above discussion, in an optional embodiment of the present application, the acquisition frequency is set to 10KHz, 10,000 data points per second are selected as training data, and the first feature and the second feature of 10,000 data points per second are calculated to form a time series feature training set.

[0049] The third feature extraction is to establish a Gaussian distribution anomaly detection model for the data points in the time series signal. The multivariate Gaussian model is shown in formula (4). The multivariate Gaussian normal distribution has two important parameters, namely, the mean vector With the covariance matrix , where the mean value changes the center position, and the covariance matrix affects the shape of the Gaussian distribution. The larger the diagonal value, the flatter the Gaussian distribution, and vice versa. In order to obtain these two core parameters, the present invention uses a data-driven parameter optimization algorithm to mine the Gaussian model parameters that best fit the true distribution of time series data, thereby obtaining the Gaussian distribution shape of the time series signal as the third feature.

[0050] (Formula 4)

[0051] Based on the above discussion, in an optional embodiment of the present application, 10,000 data points per second are also selected as training data, the least squares objective function is selected, the ant colony algorithm is used to find the optimal parameters of the Gaussian function, and a distribution diagram of the time series data points collected every second is drawn. A Gaussian distribution test is used to determine whether it conforms to the Gaussian distribution and then save it as a distribution morphology feature training set with the same number of samples as the time series feature training set.

[0052] S14: A small sample domain adaptive deep network is used to evaluate the health of battery simulation circuits based on multi-dimensional fusion features.

[0053] It should be noted that the three types of features are mixed features of timing and graphics. The timing signal can use the GRU model, and the graphic feature can use the CNN model. However, the three types of features need to be fused to reflect the comprehensive characteristics of the circuit state. Therefore, the GRU and CNN model data features are input into the Transformer model again to achieve feature adaptation and finally obtain the circuit health information. Considering the circuit differences, the model needs to be trained on domain transfer to increase the generalization diagnostic ability of the model. In addition, considering the small sample characteristics of the circuit abnormal state, the ternary center loss is used as the objective function. The framework of the health evaluation model is as follows: Figure 5 , mainly including the following steps.

[0054] Step 1: Set up labeled source domain data and unlabeled target domain data Step 1: Set up labeled source domain data and unlabeled target domain data , number of training iterations V, batch size n.

[0055] Step 2: Establish the ternary center loss objective function as shown in formula (5), where: Represents the number of samples in a batch size, represents the common feature extractor, represents a domain-specific feature extractor, Represents the source domain Fault type in , is the class center of the same fault, The ternary center loss is a metric learning-based optimization function. First, a class center is learned for each type of fault feature in the data set. The training makes the fault features of the same type of samples close to their class center, while ensuring that the distance from the sample feature to its class center is smaller than the distance to the center of different classes, and the distance difference is greater than the set boundary a. Finally, the intra-class distance of the feature is reduced and the inter-class distance is increased.

[0056] (Formula 5)

[0057] Step 3: The vector formed by concatenating the first and second features in the time series feature training set is input into the GRU time series model as the feature time series, and the distribution morphology feature training set is sent to the CNN morphology recognition model. The last layer of feature values ​​of the GRU and CNN models are concatenated again and sent to the Transformer model to achieve feature adaptation. The last layer of features of the Transformer model is trained by domain adaptation to form an evaluation vector.

[0058] Step 4: The present invention uses domain adversarial learning to train the feature extractor so that it can extract domain-invariant features of each source domain and target domain. Define the domain label of the source domain as 0 and the domain label of the target domain as 1, then the domain classification loss of each specific domain discriminator is: where K is the number of specific domain discriminators, , and represents the number of samples in the kth source domain and the number of samples in the target domain, Denotes the kth domain-specific discriminator, let and are the parameters of the feature extractor F and the domain discriminator D in the sub-network respectively. In order to make the network extract domain-invariant features, minimize the loss of the domain discriminator Parameters for training the domain discriminator , maximizes the loss of the domain discriminator To update the parameters of the feature extractor In order to extract the cross-domain core features of health status, data from multiple source domains are used. Each source domain has an independent sub-classifier. Cosine similarity is used to align all classifier outputs, and the mean of all classifier output features is used as the evaluation vector of different health status.

[0059] (Formula 6)

[0060] Step 5: The KNN method selects the type with the greatest similarity to the input sample from the evaluation vectors of all states. This paper adopts the KNN algorithm instead of the Euclidean distance measurement, matches the evaluation vector extracted from the sample to be tested with the evaluation vector set extracted from the multi-source domain training set samples, and selects the sample with the largest number among the several sample vectors closest to the vector to be tested as the state type.

[0061] Based on the above discussion, in an optional embodiment of the present application, the circuit health evaluation process may include:

[0062] In an optional embodiment of S14, we select 10,000 data points collected in 1 second as the sample length, and collect data from circuits in four states: good, moderate, poor, and severely poor health at different voltages to form a multi-source domain training set. The cross-domain features of the samples sampled per second at different voltages for each state are extracted through a small sample domain adaptive training model to form a state evaluation vector set, and then the collected data of the circuit to be tested is extracted through a small sample domain adaptive training model to extract the vector to be evaluated, and finally the KNN nearest neighbor algorithm is used to match the health state type of the circuit to be tested.

[0063] S15: High-speed and high-precision voltage control is achieved based on the distributed model predictive control algorithm DMPC.

[0064] It should be noted that in the problem of digital closed-loop control of battery simulator voltage, the goal of distributed MPC is to achieve tracking control of multiple single-cell battery voltages by reasonably allocating control tasks and information exchange. Traditional centralized control methods have a large computational load when dealing with large systems. Distributed MPC decomposes complex systems into multiple subsystems. The specific process is as follows Figure 6 , mainly including the following steps.

[0065] Step 1: Consider each single-cell battery in the battery simulation system as an independent subsystem and set a target voltage value for each single-cell battery.

[0066] Step 2: Digital closed-loop analog voltage control is a linear system. The digital quantity is directly linearly related to the output analog quantity. However, due to external interference such as DA, AD conversion and push-pull output circuit differences, the voltage output deviates. In order to overcome the control deviation caused by external interference, it is necessary to first establish a step response mathematical model for each single-cell battery as shown in formula (7), where represents the impact of future input on output, G is the weight matrix, and Represents the impact of historical input on output, and y(t) represents the output value at the current moment.

[0067] (Formula 7)

[0068] Step 3: In formula (7) The control vector is the control variable to be solved, that is, the digital quantity output to the DA converter. The MPC solution problem based on the step response model is a quadratic optimization problem. The optimization objective function is shown in Equation (8), which is a multi-objective function. and Represent the target value and the predicted value respectively, N is the prediction step length, is the control step length, and r is the weight ratio. In the rolling optimization process, in addition to requiring the predicted value sequence to track the set target, it is also hoped that the control variable is as small as possible to avoid overshoot. In addition, to ensure the accuracy of the prediction, the real sampling value will be used to continuously feedback and correct the prediction result at each sampling moment.

[0069] (Formula 8)

[0070] Step 4: During the BMS balancing test, a certain range of differences is required between the voltages of different batteries. At this time, the voltage of each single-cell battery is no longer the focus of control, and the voltage difference between the batteries becomes the core of control. This requires collaborative optimization between the various sub-controllers to ensure information communication between different single-cell batteries while achieving subsystem control, thereby achieving the goal of distributed control and centralized management, which is impossible to achieve with general centralized control algorithms and decentralized predictive control algorithms.

[0071] Based on the above discussion, in an optional embodiment of the present application, the battery voltage distributed model predictive control process may include:

[0072] In an optional embodiment of S15, we set the interception length to 30, the prediction step to 7, the control step to 5, the sampling interval to 10 milliseconds, and according to the influence of the digital quantity on the analog quantity, the maximum digital change is set to 20, and the corresponding output analog quantity is 0.1V.

[0073] This application collects the simulated battery voltage as the original sequence, obtains the denoised time series signal after LPC preprocessing, extracts the multivariate time-frequency features of the denoised time series signal as input, and trains the small sample domain transfer neural network to realize the intelligent early warning of the health status of the battery simulation circuit; for the healthy circuit, the sliding mean of the denoised time series signal is used as the feedback quantity, and the distributed model predictive control DMPC algorithm is used to realize the high-speed and high-precision synchronous tracking control of the multi-channel analog voltage setting value. The controller output is converted by DA as the simulated battery voltage, and finally realizes the high-speed and high-precision simulation of the battery voltage. This application ensures the accuracy and rapidity of battery voltage regulation and the early warning of the health status of the hardware circuit, and improves the voltage simulation accuracy and circuit reliability.

[0074] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Specific cases are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A digital closed-loop distributed predictive control method for a battery simulation system, characterized in that: include: Construct a digital closed-loop voltage regulation circuit based on AD high-speed sampling; The high-speed AD acquisition data of the analog voltage of the single battery forms a time series, and a linear predictive coding LPC is used to remove the high-frequency noise interference of the time series signal to form a denoised time series signal; Calculate the spectrum value of the denoised time series signal, use the spectral kurtosis algorithm to extract the abnormal spike spectrum sequence as the first feature, calculate the skewness, margin, variance and mean time domain statistical information of the denoised time series signal as the second feature, and extract the Gaussian distribution form of the denoised time series signal data points as the third feature; Based on three kinds of feature information, a small sample domain transfer deep neural network is trained to realize the health status warning of the battery simulation circuit; For healthy circuits, the sliding mean of the denoised timing signal is calculated as the feedback of the simulated battery voltage, and the distributed model predictive control algorithm DMPC is used to synchronously track and control the set values ​​of multiple single battery voltages to achieve high-speed and high-precision analog voltage switching regulation, thereby ensuring the accuracy and speed of the power storage BMS test.

2. A digital closed-loop distributed predictive control method for a battery simulation system as claimed in claim 1, characterized in that: Construct a digital closed-loop voltage regulation loop based on AD high-speed sampling, including: A digital closed-loop control method based on AD high-frequency sampling is proposed. The high-speed AD is directly used to collect the digital value of the DA output after passing through the push-pull regulator voltage, and the difference is made with the PC digital setting value. The DA conversion output is controlled based on the difference to form a digital closed loop, bypassing the difference amplifier and MOS tube to achieve high-speed and high-precision voltage simulation.

3. A digital closed-loop distributed predictive control method for a battery simulation system as claimed in claim 1, characterized in that: Remove high-frequency noise interference from timing signals, including: The full-pole model in the linear predictive coding (LPC) method is adopted to establish a sliding average model. Considering the stationarity of the sampling signal, a data-driven method is used to realize the parameter estimation of the model's linear prediction coefficients. The time series signal is filtered based on the established linear prediction model to obtain a denoised time series signal.

4. A digital closed-loop distributed predictive control method for a battery simulation system as claimed in claim 1, characterized in that: Calculate three features of denoised time series signals, including: Perform FFT transformation on the denoised time series signal to obtain spectrum sequence data, calculate the first-order and second-order differences of the spectrum sequence, obtain the slope and curvature of the spectrum sequence respectively, and find the sequence formed by the five maximum points of the slope and curvature as the first feature; calculate the skewness, margin, variance and mean time domain statistics of the denoised time series signal as the second feature, and extract the Gaussian distribution morphology of the denoised time series signal as the third feature based on the big data statistical method.

5. A digital closed-loop distributed predictive control method for a battery simulation system as claimed in claim 1, characterized in that: Early warning of battery simulation circuit health is provided based on three types of fusion features, including: The first and second features are concatenated into a time-frequency sequence and input into the GRU deep time series neural network. At the same time, the third feature image is input into the VGG-16 convolutional neural network. The features extracted by the two neural networks are input into the Transformer large model to realize feature fusion. Considering the small sample characteristics of the real unhealthy circuit state, the triplet center loss is used as the training objective function, and the model outputs the health status of the battery simulation circuit.

6. A digital closed-loop distributed predictive control method for a battery simulation system as claimed in claim 1, characterized in that: The distributed model predictive control algorithm DMPC is used to achieve high-speed and high-precision voltage control, including: For healthy circuits, BMS monitoring needs to test multiple different voltage values. A multi-voltage set point synchronous tracking control method based on the DMPC algorithm is proposed. Considering the hardware differences of each simulated battery, a feedforward model is established for each battery. All battery voltages are taken as subsystems, and distributed acquisition and centralized control methods are used to achieve multi-channel synchronous control. Ultimately, the rapidity and accuracy of battery voltage regulation are guaranteed, and the testing efficiency of power storage BMS is improved.

Images