Cloud SOC capacity correction method based on neural network

Abstract

The invention provides a cloud SOC capacity correction method and system based on a neural network, and relates to the technical field of battery power management, and the method comprises the steps: obtaining a state-of-charge-open-circuit voltage curve of a battery pack at a current temperature and an identification parameter, and constructing an equivalent model of the battery pack, recording the average battery temperature of each single battery in the equivalent model at each moment; working condition parameters of the battery pack during actual operation at each moment are obtained according to a time sequence, all the working condition parameters and all the average battery temperatures serve as time sequence data and are processed into a plurality of input samples, and the input samples are input into a pre-constructed state-of-charge value prediction model to obtain a real-time predicted state-of-charge value; and reading the minimum single voltage in the battery pack and the total current of the battery pack, calculating to obtain a local charge state value and a real-time predicted charge state value, and carrying out weighted summation to obtain a final charge state estimation value. The method has the beneficial effect of better estimation precision in the initial discharge stage and the final discharge stage of the battery pack.

Description

Technical Field

[0001] The present invention relates to the technical field of battery power management, and in particular to a cloud-based SOC capacity correction method based on a neural network. Background Art

[0002] In a battery management system, it is crucial to accurately estimate the battery state of charge (SOC). This not only tells the user how much power is left in the battery, but also serves as the basis for battery charge and discharge management and balanced control management. Since SOC is affected by many factors such as temperature, current size and direction, accurate prediction becomes more difficult. Improving the accuracy of SOC estimation can help extend battery life and improve user experience.

[0003] At present, the most commonly used algorithm in battery management systems is the combination of the ampere-hour integration method and the open-circuit voltage method. However, this algorithm has problems such as difficulty in determining the initial value, large cumulative error, and difficulty in dealing with differences in battery packs.

[0004] In addition, there are two other charge state estimation methods in the prior art: Kalman filtering and neural network algorithm, but both have their own defects. Since the extended Kalman filter itself has the problem of high-order term error accumulation caused by the first-order Taylor expansion of the model observation equation under the premise of accurate model, the algorithm has a large error at the end of charging and discharging. Considering the plateau problem of lithium iron phosphate batteries, it is difficult for the neural network algorithm to learn a sufficiently accurate model during the plateau period, and it is difficult to maintain a high degree of accuracy at the beginning of discharge. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a cloud-based SOC capacity correction method based on a neural network, comprising: step S1, the cloud platform obtains the state of charge-open circuit voltage curve and identification parameters of the battery pack at the current temperature from a pre-constructed parameter library, then constructs an equivalent model of the battery pack based on the state of charge-open circuit voltage curve and the identification parameters, and records the average battery temperature of each single battery in the equivalent model at each moment; step S2, the cloud platform obtains the operating condition parameters of the battery pack during actual operation at each moment uploaded by the local battery management system in chronological order, takes each of the operating condition parameters and each of the average battery temperatures as time series data and processes them into multiple input samples, and inputs each of the input samples into a pre-constructed state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack; step S3, the local battery management system reads the minimum single cell voltage in the battery pack and the total current of the battery pack, calculates the local state of charge value according to the minimum single cell voltage and the total current, and performs weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value.

[0006] Preferably, the construction process of the parameter library includes: step A1, leaving the fully charged test battery group at a first test temperature to stand for a first standing time, and recording the open circuit voltage of the test battery group in the fully charged state as the terminal voltage corresponding to the initial charge state of the test battery group; step A2, discharging the test battery group according to a first proportional constant current pulse for a first discharge time, and leaving it to stand for a second standing time, then charging the test battery group according to the second proportional constant current pulse for a first charging time, then discharging the test battery group according to a third proportional constant current pulse, stopping the discharge and leaving it to stand for a third standing time after the charge state of the test battery group drops by a preset percentage, and then recording the test battery group The current open circuit voltage is recorded as the terminal voltage corresponding to the test battery group under the current state of charge; step A3, repeating step A2 multiple times until the test battery group is discharged, so as to obtain the state of charge-open circuit voltage curve and identification parameters of the test battery group under the first test temperature; step A4, setting the first test temperature to multiple different preset test temperatures in sequence, and repeating steps A1 to A3 after each setting of the first test temperature to obtain the state of charge-open circuit voltage curve and identification parameters corresponding to each preset test temperature, and adding all the state of charge-open circuit voltage curves and identification parameters to the parameter library.

[0007] Preferably, the construction process of the state of charge prediction model includes: step B1, the cloud platform collects the operating parameters of the battery pack at each moment, and obtains the state of charge label of the battery pack by integrating the operating parameters by ampere-hour, and uses each of the operating parameters and the corresponding state of charge label as model training data; step B2, the cloud platform adds a convolution layer after the input layer of the pre-constructed standard convolutional neural network, and sets multiple convolution kernels in the convolution layer and sets the sliding stride of the convolution layer to obtain an improved convolutional neural network; step B3, connecting the output of the improved convolutional neural network to the input of the long short-term memory network, connecting the output of the long short-term memory network module to the attention module to obtain an initial fusion model, inputting each of the operating parameters and the corresponding state of charge label in the model training data into the initial fusion model in clusters, and performing model training on the initial fusion model to obtain the state of charge prediction model.

[0008] Preferably, the step S2 includes: step S21, the cloud platform obtains the terminal voltage and the operating current of the battery pack during actual operation as the operating condition parameters in chronological order, and takes each of the operating condition parameters and each of the average battery temperatures in chronological order as time series data in clusters, and then processes the time series data through a sliding window to obtain input samples of several preset time steps; step S22, the cloud platform inputs each of the input samples into the pre-constructed state of charge value prediction model to obtain the real-time predicted state of charge value of the battery pack.

[0009] Preferably, the local battery management system has multiple alarm levels preset, the alarm levels include multiple low voltage alarm levels and multiple high voltage alarm levels, and the battery pack also includes a relay; then the step S3 includes: step S31, the local battery management system reads the minimum single cell voltage in the battery pack and the total current of the battery pack, and calculates the local state of charge value according to the minimum single cell voltage and the total current; step S32, the local battery management system determines whether the minimum single cell voltage is lower than the highest level of the low voltage alarm level: if so, then cut The relay is disconnected, and the local state of charge value is forcibly corrected to 0, and then the process goes to step S34; if not, the process goes to step S33; in step S33, the local battery management system determines whether the minimum single cell voltage is higher than the highest level of the overvoltage alarm level: if so, the relay is disconnected, and the local state of charge value is forcibly corrected to 100, and then the process goes to step S34; if not, the process goes to step S34; in step S34, the local battery management system performs a weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value.

[0010] Preferably, in the step S31, a recursive filter algorithm is used to calculate the local state of charge value according to the minimum cell voltage and the total current.

[0011] Preferably, step S2 also includes: step S23, the cloud platform reads the total current of the battery pack and the opening and closing state of the relay, and when determining that the battery pack is in a low-current charging and discharging state based on the total current and the opening and closing state, determines whether the low-current charging and discharging state exceeds a preset duration: if so, the state of charge value prediction model is calibrated using the state of charge-open circuit voltage curve at the current temperature; if not, a recursive filter algorithm is used to calculate the real-time predicted state of charge value based on each of the input samples, and the stored state of charge estimate of the battery pack at the previous moment is used to calculate the initial state of charge value of the recursive filter algorithm at the next moment through the ampere-hour empirical formula.

[0012] The present invention also provides a cloud-based SOC capacity correction system based on a neural network, which uses the cloud-based SOC capacity correction method as described above, including a cloud platform and a local battery management system connected to the cloud platform, wherein the cloud platform includes: a curve acquisition module, which is used to obtain the state of charge-open circuit voltage curve and identification parameters of the battery pack at the current temperature from a pre-built parameter library, and then construct an equivalent model of the battery pack based on the state of charge-open circuit voltage curve and identification parameters, and record the average battery temperature of each single battery in the equivalent model at each moment; a charge prediction module, which is connected to the curve acquisition module, and is used to obtain the local The operating condition parameters of the battery pack during actual operation at each moment of upload, each of the operating condition parameters and each of the average battery temperatures are used as time series data and processed into multiple input samples, and each of the input samples is input into a pre-built state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack; the local battery management system comprises: a state of charge calculation module, used to read the minimum single cell voltage in the battery pack and the total current of the battery pack, calculate the local state of charge value according to the minimum single cell voltage and the total current, and perform weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value.

[0013] Preferably, the cloud platform also includes a parameter library construction module, which is connected to the curve acquisition module, including: a first test unit, used to let the fully charged test battery group stand for a first standing time at a first test temperature, and record the open circuit voltage of the test battery group in the fully charged state as the terminal voltage corresponding to the initial charge state of the test battery group; a second test unit, connected to the first test unit, used to discharge the test battery group according to a first proportional constant current pulse for a first discharge time, and stand for a second standing time, then charge the test battery group according to the second proportional constant current pulse for a first charging time, then discharge the test battery group according to a third proportional constant current pulse, stop discharging and stand for a third standing time after the charge state of the test battery group drops by a preset percentage, and then record the current open circuit voltage of the test battery group as the test battery group. The terminal voltage corresponding to the current state of charge; a third test unit, connected to the second test unit, and used to repeat the test process of the second test unit multiple times until the test battery group is discharged, so as to obtain the state of charge-open circuit voltage curve and identification parameters of the test battery group at the first test temperature; a fourth test unit, connected to the first test unit, the second test unit and the third test unit, and used to set the first test temperature to multiple different preset test temperatures in sequence, and repeat the test process of the first test unit, the second test unit and the third test unit after each setting of the first test temperature to obtain the state of charge-open circuit voltage curve and identification parameters corresponding to each preset test temperature, and add all the state of charge-open circuit voltage curves and identification parameters to the parameter library.

[0014] Preferably, the charge prediction module includes: a sample generation unit, used to obtain the terminal voltage and operating current of the battery pack during actual operation as the operating condition parameters in chronological order, and take each of the operating condition parameters and each of the average battery temperatures in chronological order as time series data in clusters, and then subject the time series data to sliding window processing to obtain a number of input samples with preset time steps; a charge prediction unit, connected to the sample generation unit, used to input each of the input samples into the pre-constructed state of charge value prediction model to obtain the real-time predicted state of charge value of the battery pack.

[0015] The above technical solution has the following advantages or beneficial effects: the state of charge of the battery pack is predicted by using a state of charge prediction model constructed based on a neural network model on the cloud platform, and the local state of charge is calculated in the local battery management system, and finally the two data obtained on the cloud and locally are weighted and summed to obtain the final charge estimation value. Compared with the charge estimation values ​​estimated by using only neural networks and only Kalman filters in the prior art, the correction method in this embodiment has better estimation accuracy at the initial and final stages of discharge of the battery pack. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a cloud-based SOC capacity correction method based on a neural network in a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the process of constructing a parameter library in a preferred embodiment of the present invention; Figure 3 A schematic diagram of the construction process of the state of charge prediction model in a preferred embodiment of the present invention; Figure 4 It is a schematic diagram of a charge state value prediction model in a preferred embodiment of the present invention; Figure 5 Schematic diagram of a sub-process of step S2 in a preferred embodiment of the present invention; Figure 6 Schematic diagram of a sub-process of step S3 in a preferred embodiment of the present invention; Figure 7 This is a schematic diagram of an equivalent model of a battery in a preferred embodiment of the present invention; Figure 8 The structure diagram of a cloud-based SOC capacity correction system based on a neural network in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment, and other embodiments may also fall within the scope of the present invention as long as they conform to the gist of the present invention.

[0018] In a preferred embodiment of the present invention, based on the above problems existing in the prior art, a cloud SOC capacity correction method based on a neural network is provided. Figure 1As shown, it includes: step S1, the cloud platform obtains the state of charge-open circuit voltage curve and identification parameters of the battery pack at the current temperature from a pre-built parameter library, and then builds an equivalent model of the battery pack based on the state of charge-open circuit voltage curve and the identification parameters, and records the average battery temperature of each single battery in the equivalent model at each moment; step S2, the cloud platform obtains the operating parameters of the battery pack when it is actually running at each moment uploaded by the local battery management system in chronological order, and uses each operating parameter and each average battery temperature as time series data and processes them into multiple input samples, and inputs each input sample into a pre-built state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack; step S3, the local battery management system reads the minimum single cell voltage and the total current of the battery pack, calculates the local state of charge value according to the minimum single cell voltage and the total current, and performs weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain the final state of charge estimation value.

[0019] Specifically, the present embodiment provides a cloud-based SOC capacity correction method, which predicts the state of charge of the battery pack by using a state of charge prediction model constructed based on a neural network model on the cloud platform, and calculates the local state of charge in a local battery management system. Finally, the two data obtained on the cloud and locally are weighted and summed to obtain the final charge estimation value. Compared with the charge estimation values ​​estimated by using only neural networks and only Kalman filters in the prior art, the correction method in this embodiment has better estimation accuracy at the initial and final stages of discharge of the battery pack.

[0020] Furthermore, before the charge estimation is performed, the method in this embodiment obtains a parameter library by constructing SOC-OCV curves (state of charge-open circuit voltage curves) at multiple temperatures and identification parameters of the battery model, so that more accurate parameters can be obtained at each temperature, such as Figure 2As shown, the construction process of the parameter library includes: step A1, leaving the fully charged test battery group at a first test temperature for a first resting time, and recording the open circuit voltage of the test battery group in the fully charged state as the terminal voltage corresponding to the initial charge state of the test battery group; step A2, discharging the test battery group according to a first proportional constant current pulse for a first discharge time, and leaving it for a second resting time, then charging the test battery group according to a second proportional constant current pulse for a first charging time, then discharging the test battery group according to a third proportional constant current pulse, stopping the discharge and leaving it for a third resting time after the charge state of the test battery group decreases by a preset percentage, and then recording the test battery group at a first charge time. The current open circuit voltage of the test battery pack is recorded as the terminal voltage corresponding to the current state of charge of the test battery pack; step A3, repeating step A2 multiple times until the test battery pack is discharged, so as to obtain the state of charge-open circuit voltage curve and identification parameters of the test battery pack at the first test temperature; step A4, setting the first test temperature to multiple different preset test temperatures in sequence, and repeating steps A1 to A3 after each setting of the first test temperature to obtain the state of charge-open circuit voltage curve and identification parameters corresponding to each preset test temperature, and adding all the state of charge-open circuit voltage curves and identification parameters to the parameter library.

[0021] More specifically, in one of the preferred embodiments of the present invention, the parameter library is constructed according to the following process: a: Leave a fully charged single-cell battery at room temperature of 25°C (the first test temperature in this embodiment) for 1 hour (the first resting time in this embodiment) to reach a equilibrium state, and record the open circuit voltage OCV when the terminal voltage of the battery is 100% SOC (the initial state of charge in this embodiment).

[0022] b: The battery is discharged with a 1 / 2C constant current pulse (the first proportional constant current pulse in this embodiment, 1C current is the current required to discharge the battery to a residual capacity of 0 for one hour) for 10s (the first discharge time in this embodiment), the battery is left to stand for 40s (the second standing time in this embodiment), and then charged with a 1 / 2C constant current pulse (the second proportional constant current pulse in this embodiment, preferably equal to the first proportional constant current pulse, and may be unequal in other embodiments) for 10s (the first charging time in this embodiment, preferably equal to the first discharge time, and may be unequal in other embodiments).

[0023] c: Use a 1 / 3C constant current (the third proportional constant current pulse in this embodiment) to discharge the battery to 95% SOC (the preset proportional percentage in this embodiment is 5%, so the SOC ratio is discharged from 100% to 95% when the third proportional constant current pulse is discharged for the first time, and the SOC ratio is discharged from 95% to 90% when the third proportional constant current pulse is discharged for the second time, and so on), let the battery stand for 1 hour (the third static time in this embodiment) to reach a balanced state, and then record the current (95% SOC) open circuit voltage of the test battery pack as the corresponding terminal voltage of the test battery pack at the current state of charge (95% SOC for the first time, 90% for the second time, and so on); d: Repeat steps b and c until the battery is fully discharged (i.e., discharge to 0% SOC and stop, obtaining a state of charge-open circuit voltage curve starting from 100% SOC, with a step length of 5% SOC and an end point of 0% SOC).

[0024] e: respectively at 45 , 35 , 15 , 5 , -5 , -15 The above experiment was repeated at multiple different ambient temperatures to obtain the SOC-OCV curve and identification parameters under variable temperatures required for the experiment.

[0025] Furthermore, the average temperature data of the single cells in the battery pack is used as one of the feature inputs of the state of charge value prediction model, which enhances the applicability of the model at different temperatures. The state of charge value prediction model in this embodiment preferably adopts a fusion model CNN-LSTM-Attention fusion model based on a neural network model combined with a long short-term memory network and an attention mechanism, such as Figure 3 As shown, the construction process of the state of charge prediction model in this embodiment includes: step B1, the cloud platform collects the operating parameters of the battery pack at each moment, and obtains the state of charge label of the battery pack by integrating the operating parameters by ampere-hour, and uses each operating parameter and the corresponding state of charge label as model training data; step B2, the cloud platform adds a convolution layer after the input layer of the pre-constructed standard convolutional neural network, and sets multiple convolution kernels in the convolution layer and sets the sliding stride of the convolution layer to obtain an improved convolutional neural network; step B3, the output of the improved convolutional neural network is connected to the input of the long short-term memory network, and the output of the long short-term memory network module is connected to the attention module to obtain an initial fusion model, and each operating parameter and the corresponding state of charge label in the model training data are input into the initial fusion model in clusters, and the initial fusion model is trained to obtain a state of charge prediction model.

[0026] Specifically, the state of charge value prediction model in one of the preferred embodiments of the present invention adopts a CNN-LSTM-Attention fusion model based on a neural network model combined with a long short-term memory network and an attention mechanism, such as Figure 4 As shown in the figure, including input layer, convolution layer, recursive layer, attention layer and final output layer, the model construction process is as follows: a: Collect the operating parameters (including current and terminal voltage) of the battery pack used in the present invention, and label the state of charge at each moment through ampere-hour integration. The operating parameters refer to the data required for neural network training, and the labels are the training labels required for supervised training.

[0027] b: Taking clusters as units, the battery time series data of multi-dimensional feature variables is used as the input of the fusion model. The battery time series data includes terminal voltage, operating current and battery temperature, and is processed into several time steps by sliding windows. Input sample Lithium batteries can be abstracted as a complex nonlinear system. In addition to the relationship characteristics between each dimensional sequence and SOC, the interaction between variables will also affect the estimation of SOC. Therefore, a convolutional layer is added after the input layer of the model built based on the standard convolutional neural network to capture the spatial characteristics between multidimensional sequences. The convolutional layer is set The convolution kernel size is 3, and the sliding stride is 1, and the feature matrix is ​​calculated , among which time step Dimension Output , and as in The input of the LSTM network at this moment.

[0028] c: Use LSTM structure (Long Short-Term Memory) in the recursive layer to convert the feature matrix output by the convolutional layer into As the input of the LSTM layer. The LSTM layer can capture the contextual characteristics of the sequence data in the time dimension, that is, the impact of historical information on the prediction of the SOC value at the current moment. The hidden state of each moment is calculated by the network , and use it as the output of the recurrent layer.

[0029] d: Finally, the attention mechanism is used to process the output of the recursive layer and finally connected to the fully connected layer to obtain the state of charge prediction value output of the final model. Finally, the operating parameters and corresponding labels in step a are input into the constructed initial fusion model for model training, and the model training parameters are saved to obtain the final state of charge value prediction model.

[0030] More specifically, after the final state of charge value prediction model of the present invention is obtained through training, the real-time state of charge prediction value can be predicted according to the process of step S2, such as Figure 5 As shown, specifically, it includes: step S21, the cloud platform obtains the terminal voltage and working current of the battery pack during actual operation as operating parameters in chronological order, and uses clusters of each operating parameter and each average battery temperature in chronological order as time series data, and then processes the time series data through a sliding window to obtain input samples of several preset time steps; step S22, the cloud platform inputs each input sample into a pre-constructed state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack.

[0031] In a preferred embodiment of the present invention, the local battery management system has multiple alarm levels preset, the alarm levels include multiple low voltage alarm levels and multiple high voltage alarm levels, and the battery pack also includes a relay; Figure 6As shown, step S3 includes: step S31, the local battery management system reads the minimum cell voltage and the total current of the battery pack in the battery pack, and calculates the local state of charge value according to the minimum cell voltage and the total current; step S32, the local battery management system determines whether the minimum cell voltage is lower than the highest level of the voltage too low alarm level: if so, the relay is cut off, and the local state of charge value is forcibly corrected to 0, and then turns to step S34; if not, turn to step S33; step S33, the local battery management system determines whether the minimum cell voltage is higher than the highest level of the voltage too high alarm level: if so, the relay is cut off, and the local state of charge value is forcibly corrected to 100, and then turns to step S34; if not, turn to step S34; step S34, the local battery management system performs a weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value; the weighted summation here can be calculated in the following way: the real-time predicted state of charge value is represented as A, and the sum of the real-time predicted state of charge value is The weight coefficient is a, the local state of charge value is expressed as B, the weight coefficient of the local state of charge value is b, and the final state of charge estimation value is expressed as C; then the weighted calculation formula here can be expressed as: aA+bB=C; a and b respectively represent the importance of A and B to C, and in this embodiment, it is preferably set to the sum of a and b to 1, for example, a=0.4, b=0.6, or a=0.3, b=0.7, etc.; in other embodiments, considering that the battery aging degree may affect the accuracy of the real-time prediction of the state of charge value, the coefficients a and b can be adjusted in combination with the battery aging degree c (usually obtained based on the usage time / standard service life, and can also be obtained based on other existing technologies), and the adjustment method is: set the coefficient reference d, a=d*(1-c), b=1-a, and then perform weighted calculation to obtain the final state of charge estimation value; A, B, C, a, b, and d here are only used as symbols and have nothing to do with the calculation formula in the subsequent embodiments; the above is only used as an example and does not mean that it is limited to the above weighting method.

[0032] In this embodiment, in step S31, a recursive filter algorithm is used to calculate the local state of charge value according to the minimum cell voltage and the total current.

[0033] Specifically, in this embodiment, the local battery management system adopts the EKF algorithm (EKF stands for Extended Kalman Filter, a highly efficient recursive filter) to estimate the local state of charge value. The local battery management system reads the minimum single cell voltage and total current in the battery pack as input, and uses the calculated SOC as the SOC of the overall battery pack.

[0034] like Figure 7As shown, the equivalent model of the battery constructed in the above embodiment (a second-order RC Thevenin battery simulation model, constructed using the prior art, which will not be repeated here), the discrete state space equation and observation equation of the battery model are as follows:

[0035]

[0036] in:

[0037]

[0038]

[0039] in: is the rated capacity of the battery. is the voltage of the two RC modules in the battery model shown in the figure.

[0040] After the first-order Taylor expansion of formula (2), we can obtain:

[0041] in:

[0042]

[0043]

[0044] set up is the system noise covariance matrix, is the observation noise covariance matrix, is the covariance matrix of the system state quantity, and the calculation process of EKF is as follows: Step 1: Initialization and , k=0 Step2:

[0045] Step 3:

[0046] Step 4:

[0047] Step5:

[0048] Step 6:

[0049] in:

[0050] Finally, the SOC value of the battery with the lowest single cell voltage is taken as the local state of charge value of the entire battery pack.

[0051] Furthermore, after each calculation of the local state of charge value, the terminal voltage is used to determine whether the battery is in a fully charged or discharged state, and the local state of charge value is forcibly set for correction according to the situation. Specifically: the minimum voltage of the single cell is read through the cluster control; two types of three-level warnings are set, the first type: one-level warning for low voltage, two-level warning for low voltage, and three-level warning for low voltage. When the voltage reaches the three-level warning limit, the relay is automatically cut off and the SOC is forced to be set to 0; the second type: one-level warning for overvoltage, two-level warning for overvoltage, and three-level warning for overvoltage. When the voltage reaches the three-level warning limit, the relay is automatically cut off and the SOC is forced to be set to 100.

[0052] In a preferred embodiment of the present invention, step S2 also includes: step S23, the cloud platform reads the total current of the battery pack and the opening and closing state of the relay, and when it is determined that the battery pack is in a low-current charging and discharging state based on the total current and the opening and closing state, it is determined whether the low-current charging and discharging state exceeds a preset duration: if so, the state of charge value prediction model is calibrated using the state of charge-open circuit voltage curve at the current temperature; if not, a recursive filter algorithm is used to calculate a real-time predicted state of charge value based on each input sample, and the stored state of charge value of the battery pack at the previous moment is used to calculate the initial state of charge value of the recursive filter algorithm at the next moment through the ampere-hour empirical formula.

[0053] To be more specific, the cluster control reads the value of the overall current and the opening and closing states of the relay to determine whether the battery is in a low-current charge and discharge state. If the battery is in this state for no longer than the set duration, the EKF algorithm is used to calculate the real-time predicted state of charge value based on each input sample. When calculating the real-time predicted state of charge value, the stored state of charge estimate at the previous moment is used to calculate the EKF with the ampere-hour empirical formula (the current value is set through experience) to assign the EKF an initial SOC value for the next moment. If the duration is exceeded, the state of charge-open circuit voltage curve at the current temperature is used to calibrate the model and clear the accumulated time of the low-current charge and discharge state to restart the timing.

[0054] The cloud-based SOC capacity correction method of the present invention obtains the SOC-OCV curves and parameter curves of the battery model at multiple temperatures, and uses the average temperature data of the battery pack single cells as one of the feature inputs of the CNN-LSTM-Attention fusion model, thereby enhancing the applicability of the model at different temperatures; the difference in calculation accuracy at different stages of battery charging and discharging is combined by weighted combination of the two calculation methods, thereby improving the accuracy of the estimation of the battery pack state of charge.

[0055] The present invention also provides a cloud SOC capacity correction system based on a neural network, which uses the cloud SOC capacity correction method as described above. Figure 8 As shown, it includes a cloud platform and a local battery management system connected to the cloud platform, such as the cloud platform 1 includes: a curve acquisition module 11, which is used to obtain the state of charge-open circuit voltage curve and identification parameters of the battery pack at the current temperature from a pre-constructed parameter library, and then construct an equivalent model of the battery pack based on the state of charge-open circuit voltage curve and the identification parameters, and record the average battery temperature of each single battery in the equivalent model at each moment; a charge prediction module 12, connected to the curve acquisition module 11, is used to obtain the operating condition parameters of the battery pack when it is actually running at each moment uploaded locally in chronological order, and each operating condition parameter and each average battery temperature are used as time series data and processed into multiple input samples, and each input sample is input into a pre-constructed state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack; the local battery management system 2 includes: a state of charge calculation module 21, which is used to read the minimum single cell voltage and the total current of the battery pack in the battery pack, calculate the local state of charge value according to the minimum single cell voltage and the total current, and perform weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain the final state of charge estimation value.

[0056] In a preferred embodiment of the present invention, the cloud platform 1 also includes a parameter library construction module 13, which is connected to the curve acquisition module 11, including: a first test unit 131, which is used to let the fully charged test battery group stand for a first standing time at a first test temperature, and record the open circuit voltage of the test battery group in the fully charged state as the terminal voltage corresponding to the initial charge state of the test battery group; a second test unit 132, connected to the first test unit 131, which is used to discharge the test battery group according to a first proportional constant current pulse for a first discharge time, and stand for a second standing time, then charge the test battery group according to a second proportional constant current pulse for a first charging time, then discharge the test battery group according to a third proportional constant current pulse, stop discharging and stand for a third standing time after the charge state of the test battery group drops by a preset percentage, and then record the current open circuit voltage of the test battery group as The terminal voltage corresponding to the current state of charge of the test battery group; the third test unit 133, connected to the second test unit 132, is used to repeat the process of the second test unit 132 multiple times until the test battery group is discharged, so as to obtain the state of charge-open circuit voltage curve and identification parameters of the test battery group at the first test temperature; the fourth test unit 134, connected to the first test unit 131, the second test unit 132 and the third test unit 133, is used to set the first test temperature to multiple different preset test temperatures in sequence, and repeat the process of the first test unit 131, the second test unit 132 and the third test unit 133 after each setting of the first test temperature to obtain the state of charge-open circuit voltage curve and identification parameters corresponding to each preset test temperature, and add all the state of charge-open circuit voltage curves and identification parameters to the parameter library.

[0057] In a preferred embodiment of the present invention, the charge prediction module 12 includes: a sample generation unit 121, which is used to obtain the terminal voltage and operating current of the battery pack during actual operation as operating parameters in chronological order, and use the operating parameters and average battery temperatures in chronological order as time series data in clusters, and then process the time series data through a sliding window to obtain input samples of several preset time steps; a charge prediction unit 122, which is connected to the sample generation unit 121, and is used to input each input sample into a pre-constructed state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack.

[0058] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of this specification and illustrations should be included in the protection scope of the present invention.

Claims

1. A cloud SOC capacity correction method based on neural network, characterized in that: include: Step S1, the cloud platform obtains the state of charge-open circuit voltage curve and identification parameters of the battery pack at the current temperature from a pre-constructed parameter library, and then constructs an equivalent model of the battery pack based on the state of charge-open circuit voltage curve and the identification parameters, and records the average battery temperature of each single battery in the equivalent model at each moment; Step S2, the cloud platform obtains the operating condition parameters of the battery pack during actual operation at each moment uploaded by the local battery management system in chronological order, takes each of the operating condition parameters and each of the average battery temperatures as time series data and processes them into multiple input samples, and inputs each of the input samples into a pre-constructed state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack; Step S3, the local battery management system reads the minimum single cell voltage in the battery pack and the total current of the battery pack, calculates the local state of charge value according to the minimum single cell voltage and the total current, and performs weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value.

2. The cloud SOC capacity correction method according to claim 1, characterized in that: The construction process of the parameter library includes: step A1, leaving the fully charged test battery group at a first test temperature to stand for a first standing time, and recording the open circuit voltage of the test battery group in a fully charged state as the terminal voltage corresponding to the initial state of charge of the test battery group; step A2, discharging the test battery group according to a first proportional constant current pulse for a first discharge time, and leaving it to stand for a second standing time, then charging the test battery group according to a second proportional constant current pulse for a first charging time, then discharging the test battery group according to a third proportional constant current pulse, stopping the discharge and leaving it to stand for a third standing time after the state of charge of the test battery group drops by a preset percentage, and then recording the current of the test battery group. The open circuit voltage is recorded as the terminal voltage corresponding to the current state of charge of the test battery group; step A3, repeating step A2 multiple times until the test battery group is discharged, so as to obtain the state of charge-open circuit voltage curve and identification parameters of the test battery group at the first test temperature; step A4, setting the first test temperature to multiple different preset test temperatures in sequence, and repeating steps A1 to A3 after each setting of the first test temperature to obtain the state of charge-open circuit voltage curve and identification parameters corresponding to each preset test temperature, and adding all the state of charge-open circuit voltage curves and identification parameters to the parameter library.

3. The cloud SOC capacity correction method according to claim 1, characterized in that: The construction process of the state of charge prediction model includes: step B1, the cloud platform collects the operating parameters of the battery pack at each moment, and obtains the state of charge label of the battery pack by integrating the operating parameters by ampere-hour, and uses each of the operating parameters and the corresponding state of charge label as model training data; step B2, the cloud platform adds a convolution layer after the input layer of the pre-constructed standard convolutional neural network, and sets multiple convolution kernels in the convolution layer and sets the sliding stride of the convolution layer to obtain an improved convolutional neural network; step B3, connecting the output of the improved convolutional neural network to the input of the long short-term memory network, connecting the output of the long short-term memory network module to the attention module to obtain an initial fusion model, inputting each of the operating parameters and the corresponding state of charge label in the model training data into the initial fusion model in clusters, and performing model training on the initial fusion model to obtain the state of charge prediction model.

4. The cloud SOC capacity correction method according to claim 1, characterized in that: The step S2 includes: step S21, the cloud platform obtains the terminal voltage and the operating current of the battery pack during actual operation as the operating condition parameters in chronological order, and takes each of the operating condition parameters and each of the average battery temperatures in chronological order as time series data in clusters, and then processes the time series data through a sliding window to obtain input samples of a plurality of preset time steps; step S22, the cloud platform inputs each of the input samples into the pre-constructed state of charge value prediction model to obtain the real-time predicted state of charge value of the battery pack.

5. The cloud SOC capacity correction method according to claim 1, characterized in that: The local battery management system is preset with multiple alarm levels, including multiple low voltage alarm levels and multiple high voltage alarm levels. The battery pack also includes a relay. Then the step S3 includes: step S31, the local battery management system reads the minimum cell voltage in the battery pack and the total current of the battery pack, and calculates the local state of charge value according to the minimum cell voltage and the total current; step S32, the local battery management system determines whether the minimum cell voltage is lower than the highest level of the low voltage alarm level: if so, cut off the The relay is turned on and the local state of charge value is forcibly corrected to 0, and then the process goes to step S34; if not, go to step S33; in step S33, the local battery management system determines whether the minimum single cell voltage is higher than the highest level of the overvoltage alarm level: if so, the relay is turned off and the local state of charge value is forcibly corrected to 100, and then the process goes to step S34; if not, go to step S34; in step S34, the local battery management system performs a weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value.

6. The cloud SOC capacity correction method according to claim 5, characterized in that: In the step S31, a recursive filter algorithm is used to calculate the local state of charge value according to the minimum cell voltage and the total current.

7. The cloud SOC capacity correction method according to claim 5, characterized in that: The step S2 also includes: step S23, the cloud platform reads the total current of the battery pack and the opening and closing state of the relay, and when determining that the battery pack is in a low-current charging and discharging state based on the total current and the opening and closing state, determines whether the low-current charging and discharging state exceeds a preset duration: if so, the state of charge value prediction model is calibrated using the state of charge-open circuit voltage curve at the current temperature; if not, a recursive filter algorithm is used to calculate the real-time predicted state of charge value based on each of the input samples, and the stored state of charge estimate of the battery pack at the previous moment is used to calculate the initial state of charge value of the recursive filter algorithm at the next moment through the ampere-hour empirical formula.

8. A cloud-based SOC capacity correction system based on neural network, characterized in that: The cloud SOC capacity correction method as described in any one of claims 1 to 7 is applied, comprising a cloud platform and a local battery management system connected to the cloud platform, wherein the cloud platform comprises: a curve acquisition module, for acquiring a state of charge-open circuit voltage curve and identification parameters of a battery pack at a current temperature from a pre-built parameter library, and then constructing an equivalent model of the battery pack based on the state of charge-open circuit voltage curve and identification parameters, and recording the average battery temperature of each single battery in the equivalent model at each moment; a charge prediction module, connected to the curve acquisition module, for acquiring the state of charge-open circuit voltage curve and identification parameters of each moment uploaded locally in chronological order. The operating parameters of the battery pack during actual operation are taken as time series data and processed into multiple input samples, and each input sample is input into a pre-built state of charge value prediction model to obtain a real-time predicted state of charge value of the battery pack; the local battery management system comprises: a state of charge calculation module, which is used to read the minimum single cell voltage in the battery pack and the total current of the battery pack, calculate the local state of charge value according to the minimum single cell voltage and the total current, and perform weighted summation of the real-time predicted state of charge value and the local state of charge value to obtain a final state of charge estimation value.

9. The cloud-based SOC capacity correction system according to claim 8, characterized in that: The cloud platform also includes a parameter library construction module, which is connected to the curve acquisition module, including: a first test unit, which is used to let the fully charged test battery group stand for a first standing time at a first test temperature, and record the open circuit voltage of the test battery group in the fully charged state as the terminal voltage corresponding to the initial charge state of the test battery group; a second test unit, which is connected to the first test unit, and is used to discharge the test battery group according to a first proportional constant current pulse for a first discharge time, and stand for a second standing time, then charge the test battery group according to a second proportional constant current pulse for a first charging time, then discharge the test battery group according to a third proportional constant current pulse, stop discharging and stand for a third standing time after the charge state of the test battery group drops by a preset percentage, and then record the current open circuit voltage of the test battery group as the current state of charge of the test battery group a terminal voltage corresponding to the state of charge; a third test unit, connected to the second test unit, and used to repeat the test process of the second test unit multiple times until the test battery group is discharged, so as to obtain the state of charge-open circuit voltage curve and identification parameters of the test battery group at the first test temperature; a fourth test unit, connected to the first test unit, the second test unit and the third test unit, and used to set the first test temperature to multiple different preset test temperatures in sequence, and repeat the test process of the first test unit, the second test unit and the third test unit after each setting of the first test temperature to obtain the state of charge-open circuit voltage curve and identification parameters corresponding to each preset test temperature, and add all the state of charge-open circuit voltage curves and identification parameters to the parameter library.

10. The cloud-based SOC capacity correction system according to claim 8, characterized in that: The charge prediction module includes: a sample generation unit, which is used to obtain the terminal voltage and operating current of the battery pack during actual operation as the operating condition parameters in chronological order, and use the operating condition parameters and the average battery temperatures in chronological order as time series data in clusters, and then obtain a number of input samples with preset time steps through sliding window processing of the time series data; a charge prediction unit, connected to the sample generation unit, which is used to input each of the input samples into the pre-constructed state of charge value prediction model to obtain the real-time predicted state of charge value of the battery pack.

Images