Deep learning driven dynamic battery state of charge estimation method and system

Abstract

The application relates to a deep learning driven battery residual capacity dynamic evaluation method and system, and relates to the technical field of battery capacity evaluation, which comprises the following steps: reading a current SOC area and a current prediction error confidence interval; monitoring and acquiring a multi-source operation data sequence, and performing operation state stability analysis to output a current operation stability coefficient; compensating for a preset standard prediction depth to generate an adaptive prediction depth; formulating an adaptive prediction scheme according to the current operation stability coefficient and the adaptive prediction depth, acquiring adaptive multi-source prediction data, calling a battery residual capacity prediction plug-in, performing battery residual capacity prediction on a target storage battery group, and outputting a battery residual capacity evaluation result. The application solves the problem that, in the traditional battery residual capacity evaluation, a unified evaluation scheme is adopted due to the fact that the actual operation scene is not adapted, and the problem that resource overconsumption or evaluation precision cannot match actual demand easily occurs.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of battery capacity evaluation, in particular to a deep learning driven dynamic battery residual capacity evaluation method and system. BACKGROUND

[0002] With the increasing demand for battery operation reliability in new energy and energy storage fields, the accuracy and resource efficiency of battery residual capacity evaluation have become key technical requirements.

[0003] Currently, traditional battery residual capacity evaluation methods do not adapt to actual operation scenarios, and mostly use a unified solution, which not only easily consumes excessive resources, but also may not meet the demand due to insufficient accuracy, affecting the stability of equipment operation and the rationality of decision-making. SUMMARY

[0004] The present application provides a deep learning driven dynamic battery residual capacity evaluation method and system, which improves the status quo of traditional battery residual capacity evaluation not adapting to actual operation scenarios and using a unified solution leading to excessive resource consumption or evaluation accuracy not matching actual demand.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] In a first aspect, the embodiments of the present application provide a deep learning driven dynamic battery residual capacity evaluation method, which comprises:

[0007] reading the current SOC region where the target battery pack is located, and matching to obtain the current prediction error confidence interval of battery residual capacity prediction;

[0008] monitoring and obtaining a multi-source operation data sequence of the target battery pack in a historical time zone, performing operation state stationarity analysis according to the multi-source operation data sequence, and outputting a current operation stationarity coefficient;

[0009] compensating for a preset standard prediction depth according to the current prediction error confidence interval and the current operation stationarity coefficient, and generating an adaptive prediction depth;

[0010] based on a preset prediction index space, formulating an adaptive prediction scheme according to the current operation stationarity coefficient and the adaptive prediction depth, collecting and obtaining adaptive multi-source prediction data according to the adaptive prediction scheme, and calling a battery residual capacity prediction plug-in constructed based on deep learning to perform battery residual capacity prediction on the target battery pack according to the adaptive multi-source prediction data, and outputting a battery residual capacity evaluation result.

[0011] In a second aspect, the embodiments of the present application provide a deep learning driven dynamic battery residual capacity evaluation system, which comprises:

[0012] The SOC and error reading module is configured to read a current SOC region where the target battery pack is located, and match a current prediction error confidence interval for obtaining a battery remaining capacity prediction;

[0013] The multi-source data stationary analysis module is configured to monitor and obtain a multi-source operation data sequence of the target battery pack in a historical time zone, perform operation state stationarity analysis according to the multi-source operation data sequence, and output a current operation stationarity coefficient;

[0014] The prediction depth compensation module is configured to compensate a preset standard prediction depth according to the current prediction error confidence interval and the current operation stationarity coefficient, and generate an adaptive prediction depth.

[0015] The prediction scheme execution module is configured to formulate an adaptive prediction scheme according to the current operation stationarity coefficient and the adaptive prediction depth based on a preset prediction index space, collect adaptive multi-source prediction data according to the adaptive prediction scheme, and call a battery remaining capacity prediction plug-in constructed based on deep learning to perform battery remaining capacity prediction on the target battery pack according to the adaptive multi-source prediction data, and output a battery remaining capacity evaluation result.

[0016] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0017] The present application provides a deep learning driven battery remaining capacity dynamic evaluation method and system, which realizes accurate and dynamic evaluation of the target battery pack remaining capacity by obtaining battery operation data in steps, formulating an adaptive prediction scheme, building a multi-model prediction plug-in, performing dynamic prediction and result optimization. First, the multi-source operation data sequence of the target battery pack in the historical time zone is monitored and obtained, the electrical characteristic fluctuation degree and the load characteristic fluctuation degree are calculated, and the current operation stationarity coefficient is determined. Then, the preset standard prediction depth is compensated in combination with the current prediction error confidence interval to generate an adaptive prediction depth. Then, based on the preset prediction index space, an adaptive prediction index set is selected and the number of adaptive selected models is determined to form an adaptive prediction scheme. At the same time, an adaptive battery capacity prediction plug-in containing N LSTM models is constructed with the adaptive prediction index as the constraint. Finally, the adaptive multi-source prediction data is collected according to the adaptive prediction scheme, the randomly selected model in the adaptive battery capacity prediction plug-in is called for prediction, and the result is fitted by mean value, and the battery remaining capacity evaluation result is output.

[0018] The technical scheme of the application solves the problems in the traditional battery residual capacity evaluation, such as data redundancy or information loss caused by fixed indexes, difficulty of a single model in adapting to different operating conditions, and great influence of the prediction accuracy by the environment and operating state, and avoids evaluation deviation caused by improper index selection, weak model generalization ability or mismatched prediction depth, and improves the accuracy, dynamic adaptability and reliability of the battery residual capacity evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 The flowchart of the deep learning driven battery residual capacity dynamic evaluation method provided by the embodiment of the present application is shown.

[0021] Figure 2 The structural diagram of the deep learning driven battery residual capacity dynamic evaluation system provided by the embodiment of the present application is shown.

[0022] In the drawings, the components represented by the numbers are described as follows:

[0023] The SOC and error reading module 01, the multi-source data stationary analysis module 02, the prediction depth compensation module 03, and the prediction scheme execution module 04. DETAILED DESCRIPTION

[0024] The present application provides a deep learning driven battery residual capacity dynamic evaluation method and system, which is used to solve the technical problems in the prior art that the traditional battery residual capacity evaluation is not adapted to the actual operating scene and uses a unified evaluation scheme, which is easy to cause excessive consumption of resources or evaluation accuracy that cannot match the actual demand.

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] In the description of the present application, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can be explicitly or implicitly included one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0027] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can recognize that the present application can be implemented without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope consistent with the principles and characteristics disclosed in the present application.

[0028] Embodiment one, as shown in the accompanying Figure 1 The present application provides a deep learning driven battery residual capacity dynamic evaluation method, which comprises the following steps:

[0029] S110: reading the current SOC region where the target battery group is located, and matching to obtain the current prediction error confidence interval of the battery residual capacity prediction;

[0030] In the embodiments of the present application, in the scenario where the battery residual capacity prediction needs to adapt to different operating states, in order to accurately obtain the prediction error range corresponding to different SOC regions, the SOC region where the target battery group is currently located needs to be determined first, and then the prediction error confidence interval in this region is calculated accordingly, so as to guarantee the accuracy and reliability of subsequent residual capacity evaluation.

[0031] Wherein, the SOC region is a numerical division interval of the battery state of charge, indicating the proportion range of the current residual capacity of the battery to the rated capacity. The difference in electrochemical characteristics of the battery in different intervals will directly affect the error performance of the residual capacity prediction.

[0032] In the method provided by the embodiments of the present application, the SOC region includes a high SOC region, an intermediate SOC region and a low SOC region, wherein the high SOC region is SOC greater than or equal to 80%, the intermediate SOC region is SOC greater than or equal to 20% and less than 80%, and the low SOC region is SOC less than 20%, and each SOC region is identified with a prediction error confidence interval.

[0033] Specifically, firstly, the current SOC value of the target battery pack is read and determined in real time to determine the SOC region where the target battery pack is located. For example, when it is detected that the current SOC of the target battery pack is 88%, it is determined that it is in the high SOC region; if the current SOC is 45%, it corresponds to the middle SOC region; if the current SOC is 18%, it belongs to the low SOC region.

[0034] Meanwhile, each SOC region is marked with a prediction error confidence interval to quantify the fluctuation range of the remaining capacity prediction result in the current SOC region.

[0035] In the method provided by the embodiment of the application, the calculation process of the prediction error confidence interval comprises:

[0036] In the high SOC region, the middle SOC region and the low SOC region, any SOC region is randomly selected as a first SOC region;

[0037] The current aging index of the target battery pack is obtained, and the current aging index is expanded according to a preset tolerance interval to obtain an aging index interval, wherein the current aging index is determined based on historical charge and discharge cycle times and historical deep discharge times;

[0038] The first SOC region and the aging index interval are used as constraints to search the historical operation records of the same type of battery pack of the target battery pack to obtain a sample predicted SOC value set and a sample real SOC value set under different operation conditions;

[0039] Based on the sample predicted SOC value set and the sample real SOC value set, the difference between each sample predicted SOC value and the corresponding sample real SOC value is taken as a sample error to obtain a sample error set, and the error average and the error standard deviation of the sample error set are calculated;

[0040] According to a preset confidence level, a corresponding standard score is determined, the error average minus the product of the standard score and the error standard deviation is taken as the lower limit of the confidence interval, and the error average plus the product of the standard score and the error standard deviation is taken as the upper limit of the confidence interval to obtain the first prediction error confidence interval of the first SOC region.

[0041] Specifically, one of the high SOC region, the middle SOC region and the low SOC region is randomly selected as a first SOC region. For example, the middle SOC region (20%-80%) is selected as the first SOC region, and the corresponding prediction error confidence interval can be calculated for this SOC region subsequently.

[0042] Further, a current aging index of the target battery pack is obtained, which is evaluated based on its historical number of charge-discharge cycles and number of deep discharge times. For example, if the target battery pack has 500 charge-discharge cycles and 30 deep discharge times, the current aging index is evaluated to be 0.6. The aging index interval is extended to 0.56-0.64 according to the preset tolerance interval ±0.04.

[0043] Further, the historical operation records of the same type of battery pack are retrieved with the selected first SOC region and the extended aging index interval as constraints to obtain a sample predicted SOC value set and a sample true SOC value set under different operation conditions.

[0044] The operation conditions include different temperatures, different charge-discharge rates, etc. For example, the operation data of the same type of battery under the conditions of 25°C, 1C charge-discharge rate, 30°C, 0.8C charge-discharge rate, etc. are retrieved, which are in the middle SOC region and have an aging index in the interval of 0.56-0.64. The sample predicted SOC values and sample true SOC values under the above conditions are extracted correspondingly.

[0045] Further, based on the obtained sample data, each sample predicted SOC value is subtracted by the corresponding sample true SOC value, and the difference is the sample error. Integrating all sample errors can form a sample error set.

[0046] For example, the predicted SOC value of a certain group of samples is 60%, and the true SOC value is 59.2%. The corresponding sample error is 0.8%. In addition, the predicted SOC value of another group of samples is 45%, and the true SOC value is 45.5%. The corresponding sample error is -0.5%.

[0047] Similarly, after collecting multiple sample errors, the error mean and error standard deviation of the sample error set are calculated. For example, the error mean is calculated to be 0.3%, and the error standard deviation is calculated to be 0.6%.

[0048] Finally, the corresponding standard score is determined according to the preset confidence level, and the upper and lower limits of the confidence interval are calculated. Specifically, if the preset confidence level is 95% and the corresponding standard score is 1.96, the lower limit of the confidence interval is -0.876%, and the upper limit of the confidence interval is 1.476%. Thus, the first prediction error confidence interval of the middle SOC region is [-0.876%, 1.476%].

[0049] Through the above steps, the prediction error confidence interval corresponding to the selected SOC region can be obtained, which provides a clear reference basis for the control of the subsequent remaining capacity prediction.

[0050] Further, after obtaining the first prediction error confidence interval corresponding to the selected SOC region, the interval needs to be accurately matched with the SOC region currently occupied by the target battery pack to determine the current prediction error confidence interval of the battery remaining capacity prediction.

[0051] Specifically, if the first SOC region calculated by the previous random selection is consistent with the current SOC region determined in real time by the target battery pack, for example, both are the middle SOC region (20%-80%), and the extended interval of the current aging index of the target battery pack is consistent with the interval of the aging index used when calculating the first prediction error confidence interval, then the first prediction error confidence interval calculated previously can be directly used as the current prediction error confidence interval of the target battery pack.

[0052] On the contrary, if the SOC region currently occupied by the target battery pack is different from the first SOC region selected previously, for example, the first prediction error confidence interval calculated previously is for the high SOC region (≥80%), while the target battery pack is actually in the low SOC region (<20%), it is necessary to re-calculate the prediction error confidence interval corresponding to the low SOC region according to the above calculation process, taking the low SOC region as the new first SOC region, combining the current aging index of the target battery pack and the extended aging index interval, searching the historical data of similar battery packs under different operating conditions in the low SOC region, corresponding aging index interval and re-calculating the prediction error confidence interval corresponding to the low SOC region. The interval is the current prediction error confidence interval of the target battery pack.

[0053] Through the above matching process, it can be ensured that the current prediction error confidence interval obtained is highly adapted to the actual operating state (current SOC region, current aging state) of the target battery pack, laying a foundation for subsequent adjustment of the remaining capacity prediction strategy based on the error interval and guaranteeing the prediction accuracy.

[0054] S120: monitoring and obtaining a multi-source operating data sequence of the target battery pack in a historical time zone, performing operating state stationarity analysis based on the multi-source operating data sequence, and outputting a current operating stationarity coefficient;

[0055] In the embodiments of the present application, in order to accurately analyze the operating stability of the target battery pack, the multi-source operating data sequence in the historical time zone needs to be monitored and obtained first, and then the operating state stationarity analysis is carried out based on these data and the current operating stationarity coefficient is outputted, so as to provide reliable state basis for the generation of adaptive prediction depth and the formulation of adaptive prediction scheme.

[0056] Specifically, first, a multi-source operation data sequence of the target battery pack in a historical time zone is monitored and acquired. The multi-source operation data sequence includes an electrical characteristic data sequence and a load characteristic data sequence, the electrical characteristic data sequence includes a voltage sequence, a current sequence and a power sequence, and the load characteristic data sequence includes a load current sequence and a load power sequence. Both types of sequence data need to be comprehensively collected to avoid incomplete judgment of the operation state due to data loss.

[0057] Further, index fluctuation calculation is carried out based on the acquired electrical characteristic data sequence. For the voltage sequence, the current sequence and the power sequence, corresponding fluctuation calculation methods are used to acquire a voltage fluctuation coefficient, a current fluctuation coefficient and a power fluctuation coefficient respectively. Then, according to the influence degree of each coefficient on the overall stability of the electrical characteristic, corresponding weights are given and weighted summation is performed to obtain an electrical characteristic fluctuation degree, so as to quantitatively reflect the operation fluctuation situation of the battery in the electrical layer.

[0058] Meanwhile, index fluctuation calculation is carried out based on the acquired load characteristic data sequence. For the load current sequence and the load power sequence, corresponding fluctuation calculation methods are used to acquire corresponding fluctuation coefficients respectively. Then, according to the influence weight of each fluctuation coefficient on the stability of the load characteristic, weighted summation is performed to obtain a load characteristic fluctuation degree, so as to quantitatively reflect the operation fluctuation situation of the battery in the load layer.

[0059] Finally, the electrical characteristic fluctuation degree and the load characteristic fluctuation degree are weighted and evaluated to determine a current operation smoothness coefficient. In the process of weighted evaluation, reasonable weight distribution needs to be set in combination with the focus of the battery application scene on the electrical characteristic and the load characteristic. The current operation smoothness coefficient is obtained by weighted summation of the reciprocals of the electrical characteristic fluctuation degree and the load characteristic fluctuation degree, and the numerical size thereof is negatively correlated with the electrical characteristic fluctuation degree and the load characteristic fluctuation degree.

[0060] This step converts the abstract operation state of the battery into a quantifiable current operation smoothness coefficient through multi-source operation data acquisition and fluctuation degree analysis, which provides a clear numerical reference for subsequent link to adjust the prediction strategy based on the operation stability.

[0061] The method provided in the embodiments of the present application comprises the following steps.

[0062] A multi-source operation data sequence of the target battery pack in a historical time zone is monitored and acquired, wherein the multi-source operation data sequence includes an electrical characteristic data sequence and a load characteristic data sequence, the electrical characteristic data sequence includes a voltage sequence, a current sequence and a power sequence, and the load characteristic data sequence includes a load current sequence and a load power sequence;

[0063] According to the voltage sequence, current sequence, power sequence, index fluctuation calculation is performed respectively to obtain voltage fluctuation coefficient, current fluctuation coefficient and power fluctuation coefficient, and weighted sum is obtained to obtain electrical characteristic fluctuation degree;

[0064] According to the load current sequence, load power sequence, index fluctuation calculation is performed respectively, and weighted sum is obtained to obtain load characteristic fluctuation degree;

[0065] According to the electrical characteristic fluctuation degree and load characteristic fluctuation degree, weighted evaluation is performed to determine the current running smoothness coefficient, wherein the current running smoothness coefficient is negatively correlated with the electrical characteristic fluctuation degree and the load characteristic fluctuation degree.

[0066] In the embodiments of the application, in order to avoid the accuracy deviation or resource waste of the prediction scheme formulated based on the fuzzy state judgment, it is necessary to monitor the historical running data, calculate the fluctuation index and finally determine the smoothness coefficient to provide a reliable state basis for subsequent adaptive prediction depth generation and prediction scheme formulation.

[0067] Specifically, first, the multi-source running data sequence of the target battery pack in the historical time zone is monitored and obtained. The setting of the historical time zone needs to be determined in combination with the actual use scene of the battery and the data effectiveness requirement. For example, for the battery pack of the energy storage power station, the past 72 hours can be selected as the historical time zone, which covers the complete charging and discharging cycle of the battery and can also reflect the trend of the short-term running state.

[0068] In addition, for the power battery of a new energy vehicle, the time corresponding to the past 5 driving cycles can be selected as the historical time zone to ensure that the data can match the running characteristics in the actual use of the vehicle.

[0069] At the same time, during the data monitoring process, the integrity and accuracy of the multi-source running data sequence need to be ensured. The electrical characteristic data sequence needs to continuously collect real-time change data of voltage, current and power to form sequence information in the time dimension, for example, recording the voltage value once every 10 seconds, and continuously for 72 hours to form a voltage sequence containing 25920 data points. The load characteristic data sequence is the same, and the load current and load power data need to be continuously collected to form the corresponding sequence to avoid incomplete judgment of the running state due to data interruption or loss.

[0070] Further, after obtaining the multi-source running data sequence, index fluctuation calculation is performed according to the voltage sequence, current sequence and power sequence in the electrical characteristic data sequence. The index fluctuation calculation can use the ratio of standard deviation to mean value (coefficient of variation) as the core calculation method. For example, for the voltage sequence, first calculate the mean value and standard deviation of all data points in the sequence, then divide the standard deviation by the mean value to obtain the voltage fluctuation coefficient. The larger the coefficient, the more violent the voltage fluctuation in the historical time zone.

[0071] Similarly, the current fluctuation coefficient of the current sequence and the power fluctuation coefficient of the power sequence are calculated in the same way. Then, according to the demand weight of the stability of each electrical index in the battery application scenario, the three fluctuation coefficients are weighted and summed to obtain the electrical characteristic fluctuation degree.

[0072] Exemplarily, in the communication base station standby power supply scenario, the voltage stability is crucial for equipment power supply, and the voltage fluctuation coefficient can be given a weight of 40%, the current fluctuation coefficient and the power fluctuation coefficient can be given a weight of 30% respectively. If the voltage fluctuation coefficient of a certain base station battery pack is 0.02, the current fluctuation coefficient is 0.03, and the power fluctuation coefficient is 0.025, then the electrical characteristic fluctuation degree is 0.02 x 40% + 0.03 x 30% + 0.025 x 30% = 0.0245, and the size of this value directly reflects the running fluctuation situation in the electrical layer.

[0073] Meanwhile, the load current sequence and the load power sequence in the load characteristic data sequence are used to calculate the index fluctuation. Specifically, the calculation method is consistent with the electrical characteristic index, and the fluctuation degree is reflected by the coefficient of variation, and then the load characteristic fluctuation degree is obtained by weighting and summing the two indexes according to the dependence weight of the load type.

[0074] Exemplarily, in the industrial forklift battery scenario, the stability of the load current directly affects the power output of the forklift, and the load current fluctuation coefficient can be given a weight of 60%, and the load power fluctuation coefficient can be given a weight of 40%. If the load current fluctuation coefficient of a certain forklift battery is 0.05, and the load power fluctuation coefficient is 0.04, then the load characteristic fluctuation degree is 0.05 x 60% + 0.04 x 40% = 0.046, which quantifies the running fluctuation situation in the load layer.

[0075] Finally, the current running smoothness coefficient is determined by weighted evaluation according to the electrical characteristic fluctuation degree and the load characteristic fluctuation degree. Specifically, when weighting and evaluating, the weight of the electrical characteristic and the load characteristic should be set according to the focus of the battery application scenario, for example, in the energy storage power station scenario, the stability of the electrical characteristic has a greater impact on the grid-connected quality, and the reciprocal of the electrical characteristic fluctuation degree can be given a weight of 60%, and the reciprocal of the load characteristic fluctuation degree can be given a weight of 40%.

[0076] In addition, the current running smoothness coefficient is calculated by weighting and summing the reciprocals of the two after normalization processing. Specifically, first, the reciprocal of the electrical characteristic fluctuation degree and the reciprocal of the load characteristic fluctuation degree are mapped to the [0, 1] interval according to "reciprocal / maximum reciprocal of the same type of battery (value 100)", and then the final coefficient is calculated according to the scenario weight.

[0077] For example, if the electrical characteristic fluctuation degree of a certain energy storage battery pack is 0.0245 and the load characteristic fluctuation degree is 0.03, the reciprocal of the electrical characteristic fluctuation degree is about 40.82, and after normalization, it is 40.82 / 100 = 0.4082; the reciprocal of the load characteristic fluctuation degree is about 33.33, and after normalization, it is 33.33 / 100 = 0.3333. If the electrical characteristic normalization value is given a weight of 60% and the load characteristic normalization value is given a weight of 40% in the energy storage scenario, the current operation stability coefficient is 0.4082 x 60% + 0.3333 x 40% = 0.245 + 0.133 = 0.378.

[0078] Since the current operation stability coefficient is negatively correlated with the electrical characteristic fluctuation degree and the load characteristic fluctuation degree, if the electrical characteristic fluctuation degree of another battery pack is 0.05 (reciprocal 20) and the load characteristic fluctuation degree is 0.06 (reciprocal 16.67), the reciprocal of the electrical characteristic fluctuation degree after normalization is 20 / 100 = 0.2, and the reciprocal of the load characteristic fluctuation degree after normalization is 16.67 / 100 = 0.1667. The current operation stability coefficient is 0.2 x 60% + 0.1667 x 40% = 0.12 + 0.067 = 0.187, which is significantly lower than the former (0.378), indicating that the operation state of this battery pack is more unstable.

[0079] Through the above steps, the abstract operation state of the battery is converted into a quantifiable current operation stability coefficient, avoiding the limitations of single index judgment, and providing a clear numerical reference for subsequent adjustment of prediction depth and prediction scheme based on the current operation stability coefficient.

[0080] S130: compensating a preset standard prediction depth according to the current prediction error confidence interval and the current operation stability coefficient to generate an adaptive prediction depth;

[0081] In the embodiments of the present application, in order to avoid the insufficient accuracy caused by the unified use of the preset standard prediction depth, the standard prediction depth needs to be compensated based on the current prediction error confidence interval and the current operation stability coefficient to generate a prediction depth adapted to the actual scenario, so as to ensure that the subsequent prediction scheme can reasonably control resource consumption while ensuring accuracy.

[0082] Specifically, first, the interval width ratio of the current prediction error confidence interval to the preset standard prediction error confidence interval is taken as a first prediction depth compensation coefficient. The current prediction error confidence interval reflects the error range that the remaining capacity prediction may exist in the current scenario, and the preset standard prediction error confidence interval is a benchmark error range set based on a large amount of historical data and a general scenario. Through the interval width ratio of the two, the deviation degree of the current error situation relative to the benchmark can be quantified.

[0083] Further, a ratio of the preset standard operation stability coefficient and the current operation stability coefficient is set as a second prediction depth compensation coefficient. The preset standard operation stability coefficient is a reference coefficient set for the stable operation state of the battery, and the current operation stability coefficient reflects the stability degree of the actual operation of the battery. The ratio of the two coefficients can reflect the difference between the current operation stability and the reference state.

[0084] Further, the overall prediction depth compensation coefficient is obtained by weighted fusion based on the first prediction depth compensation coefficient and the second prediction depth compensation coefficient. In the weighted fusion process, reasonable weights are set in combination with the focus of the battery application scene on error control and operation stability, to ensure that the overall compensation coefficient can comprehensively reflect the influence of the error and the operation state on the prediction depth.

[0085] Finally, the product of the overall prediction depth compensation coefficient and the preset standard prediction depth is taken as the adaptive prediction depth. The preset standard prediction depth is a basic prediction precision set based on a conventional scene, and by multiplying the overall compensation coefficient, the final adaptive prediction depth can be dynamically adjusted according to the current error size and operation stability.

[0086] This step compensates the prediction depth by combining the error and the operation state in two dimensions, so that the generated adaptive prediction depth can match the current actual scene, providing a reasonable depth basis for subsequent development of an adaptive prediction scheme and residual capacity prediction.

[0087] The method provided in the embodiment of the application comprises the following steps:

[0088] A ratio of the interval width of the current prediction error confidence interval and the preset standard prediction error confidence interval is taken as a first prediction depth compensation coefficient;

[0089] A ratio of the preset standard operation stability coefficient and the current operation stability coefficient is taken as a second prediction depth compensation coefficient;

[0090] The overall prediction depth compensation coefficient is obtained by weighted fusion based on the first prediction depth compensation coefficient and the second prediction depth compensation coefficient;

[0091] The product of the overall prediction depth compensation coefficient and the preset standard prediction depth is taken as the adaptive prediction depth.

[0092] In the embodiment of the application, in order to make the precision degree of the residual capacity prediction of the battery fit the current actual operation scene, the prediction depth needs to be adjusted by multi-dimensional compensation coefficient calculation in combination with the current prediction error situation and the operation stability, to generate an adaptive prediction depth, providing a basis for subsequent development of a precise and efficient prediction scheme.

[0093] Specifically, first, the interval width ratio of the current prediction error confidence interval and the preset standard prediction error confidence interval is taken as the first prediction depth compensation coefficient. The current prediction error confidence interval is the error fluctuation range calculated based on the target battery pack current SOC region and aging state, and the preset standard prediction error confidence interval is the reference error range set based on a large number of similar batteries under standard working conditions. The interval width ratio can intuitively reflect the difference between the current error risk and the reference state.

[0094] For example, if the width of the current prediction error confidence interval is 2.5%, and the width of the preset standard prediction error confidence interval is 1.5%, the first prediction depth compensation coefficient is 2.5% / 1.5%≈1.67, which is greater than 1, indicating that the current error risk is higher than the reference, and the precision needs to be improved by adjusting the prediction depth to cover the error fluctuation.

[0095] Further, the ratio of the preset standard running smoothness coefficient and the current running smoothness coefficient is taken as the second prediction depth compensation coefficient. The preset standard running smoothness coefficient is a reference coefficient set for the battery in a stable running state, and is usually 0.8 (the closer the coefficient is to 1, the more stable the running is). The current running smoothness coefficient reflects the stability of the actual running of the battery.

[0096] For example, if the target battery pack current running smoothness coefficient is 0.5, which is lower than the preset standard running smoothness coefficient 0.8, the second prediction depth compensation coefficient is 0.8 / 0.5=1.6, which is greater than 1, indicating that the current running state is unstable, and the prediction depth needs to be deepened to reduce the interference of unstable factors on the prediction result.

[0097] On the contrary, if the target battery pack current running smoothness coefficient is 0.9, which is higher than the preset standard running smoothness coefficient 0.8, the second prediction depth compensation coefficient is 0.8 / 0.9≈0.89, which is less than 1, indicating that the running state is better than the reference, and the prediction depth can be appropriately reduced to save resources.

[0098] Further, after obtaining the first and second prediction depth compensation coefficients, the overall prediction depth compensation coefficient is obtained based on the weighted fusion of the two, to comprehensively balance the influence of the current error risk and the running stability on the prediction depth.

[0099] Specifically, the weight setting in the weighted fusion needs to focus on the error control and running stability in combination with the battery application scenarios. For example, in the new energy vehicle power battery scenario, the error control directly affects the judgment of the endurance mileage, and the first prediction depth compensation coefficient is given a weight of 60%, and the second prediction depth compensation coefficient is given a weight of 40%. In addition, in the communication base station standby power supply scenario, the running stability is more critical to the power supply reliability, and the first prediction depth compensation coefficient can be adjusted to 40% and the second prediction depth compensation coefficient can be adjusted to 60%.

[0100] Exemplarily, assuming that in the new energy vehicle scenario, the first prediction depth compensation coefficient is 1.67 and the second prediction depth compensation coefficient is 1.6, the overall prediction depth compensation coefficient is 1.67*60%+1.6*40%=1.642, which comprehensively reflects the common demand of error and running state on the prediction depth.

[0101] Finally, the product of the overall prediction depth compensation coefficient and the preset standard prediction depth is taken as the adaptive prediction depth. The preset standard prediction depth is the basic prediction precision set based on the conventional scenario, which is usually measured by “data sampling frequency*prediction time window”.

[0102] Exemplarily, the preset standard prediction depth is “sampled once every 5 seconds*predicts the future 30 minutes”. If the overall prediction depth compensation coefficient is 1.642, the adaptive prediction depth is “sampled once every 3 seconds*predicts the future 49.26 minutes” (the sampling frequency and the prediction window are enlarged by the coefficient to improve the precision); if the overall prediction depth compensation coefficient is 0.89, the adaptive prediction depth is “sampled once every 5.6 seconds*predicts the future 26.7 minutes” (the sampling frequency and the prediction window are reduced by the coefficient to reduce resource consumption).

[0103] Through the above steps, the generated adaptive prediction depth can accurately match the current error risk and running stability, avoiding both the problem of insufficient prediction accuracy when the error is too high and the problem of excessive resource consumption when the running is stable, and laying a reasonable precision foundation for subsequent adaptive prediction scheme based on the depth and battery remaining capacity prediction.

[0104] S140: Based on the preset prediction index space, an adaptive prediction scheme is formulated according to the current running smoothness coefficient and the adaptive prediction depth, adaptive multi-source prediction data is collected and acquired according to the adaptive prediction scheme, and a battery remaining capacity prediction plug-in based on deep learning is called to perform battery remaining capacity prediction on the target battery pack according to the adaptive multi-source prediction data, and output a battery remaining capacity evaluation result.

[0105] In the embodiments of the present application, in order to avoid insufficient prediction accuracy or resource waste caused by using fixed indicators and single models, an adaptive scheme is first formulated based on a preset indicator space and key parameters, then data is collected according to the scheme and a deep learning plug-in is called to complete prediction, so as to ensure that the final evaluation result is accurate and can match the efficiency requirements of actual application.

[0106] Specifically, first, a preset prediction indicator space of the target battery pack is acquired, which covers all prediction indicators related to residual capacity prediction. Then, the historical operation records of battery packs of the same type are combined, and the relevance of each prediction indicator in the space to the prediction result is evaluated one by one with the current SOC region as a constraint, to obtain the indicator relevance of each prediction indicator. The indicators are sorted in descending order of relevance to generate a prediction indicator sequence.

[0107] Further, according to the adaptive prediction depth, the number of indicators is determined by referring to a preset prediction depth-indicator number mapping table, a corresponding number of indicators are selected from the prediction indicator sequence to form an adaptive prediction indicator set.

[0108] At the same time, the number of adaptive selected models is calculated: the ratio of the current running smoothness coefficient to the average historical running smoothness coefficient in the historical time range of the target battery pack is multiplied by the standard selected model number, and the result is rounded off, and it is ensured that the number is not less than 2 and not greater than the total number of models in the battery residual capacity prediction plug-in. The adaptive prediction indicator set and the adaptive selected model number are used as an adaptive prediction scheme.

[0109] Further, after determining the adaptive prediction scheme, the corresponding adaptive multi-source prediction data of the target battery pack is collected according to the adaptive prediction indicator set in the adaptive prediction scheme, so as to ensure that the data can completely cover the information of the selected indicators.

[0110] Further, according to the adaptive prediction indicator set, an adaptive battery capacity prediction plug-in is matched and called, which includes not less than 10 battery capacity prediction models. Models are randomly selected from the adaptive battery capacity prediction plug-in according to the adaptive selected model number, and the collected adaptive multi-source prediction data is input into these models for residual capacity prediction. Finally, the prediction results output by multiple models are fitted by mean value to obtain and output the battery residual capacity evaluation result.

[0111] This step dynamically adjusts the number of prediction indicators and models, combines the multi-model prediction capability of the deep learning plug-in, ensures the adaptability of the prediction process to the actual scene, and improves the reliability of the evaluation result through mean value fitting of multiple results, thereby providing a complete and efficient implementation path for accurate judgment of battery residual capacity.

[0112] The step S140 in the method provided by the embodiments of the present application includes:

[0113] obtaining a preset prediction index space of the target battery pack, wherein the preset prediction index space comprises a plurality of prediction indexes;

[0114] based on historical operation records of the same type battery pack as the target battery pack, and taking the current SOC region as a constraint, evaluating the relevance of the plurality of prediction indexes and the battery residual capacity prediction result respectively to obtain a plurality of index correlation degrees;

[0115] arranging the plurality of prediction indexes in descending order of the index correlation degrees to generate a prediction index sequence;

[0116] based on a preset prediction depth-index number mapping table, determining an adaptive index number according to the adaptive prediction depth matching, and selecting the prediction indexes of the adaptive index number in the prediction index sequence as an adaptive prediction index set;

[0117] multiplying the ratio of the current operation smoothness coefficient and the average historical operation smoothness coefficient of the target battery pack within a historical time range by a standard selection model number and taking the integer part to obtain an adaptive selection model number, wherein the standard selection model number is 5, and the adaptive selection model number is greater than or equal to 2 and less than or equal to the total number of models in the battery residual capacity prediction plug-in;

[0118] taking the adaptive prediction index set and the adaptive selection model number as an adaptive prediction scheme.

[0119] collecting adaptive multi-source prediction data of the target battery pack according to the adaptive prediction index set;

[0120] matching and calling an adaptive battery capacity prediction plug-in according to the adaptive prediction index set, wherein the adaptive battery capacity prediction plug-in comprises N battery capacity prediction models, and N is greater than or equal to 10;

[0121] randomly selecting the same number of prediction models from the N battery capacity prediction models of the adaptive battery capacity prediction plug-in according to the adaptive selection model number, performing battery residual capacity prediction according to the adaptive multi-source prediction data respectively, and performing mean fitting on a plurality of prediction results to obtain a battery residual capacity evaluation result.

[0122] In the embodiments of the application, in order to accurately match the index selection and model use of the battery residual capacity prediction with the current operation state and prediction demand, the adaptive prediction index is screened through the preset prediction index space, the number of models is determined in combination with the operation smoothness coefficient, and then data is collected and an adaptive battery capacity prediction plug-in is called for prediction according to the adaptive prediction scheme, so as to ensure that the evaluation result is accurate and efficient.

[0123] Specifically, a preset prediction index space of the target battery pack is first acquired. The preset prediction index space is set by prediction indexes from three dimensions of the battery pack as a whole, individual cells, and external environment, and can comprehensively cover key factors affecting the remaining capacity prediction.

[0124] In the method provided by the embodiments of the present application, the plurality of prediction indexes include group-level parameters, individual cell parameters, and environment parameters, wherein the target battery pack includes a plurality of individual cells, the group-level parameters at least include terminal voltage and total current of the battery pack, the individual cell parameters at least include cell voltage, pole temperature, and internal resistance of the individual cell, and the environment parameters at least include environment temperature and environment humidity.

[0125] The group-level parameters are from the macroscopic level of the battery pack, the terminal voltage of the battery pack is directly related to the overall energy reserve state, the total current reflects the energy output intensity, and the combination of the two can quickly judge the basic operation performance of the battery pack.

[0126] In addition, the individual cell parameters focus on the characteristics of individual cells, the voltage difference of the individual cell can reflect the consistency of the cell, the pole temperature can warn the local overheating risk, and the internal resistance of the cell reflects the aging degree of the cell.

[0127] In addition, the environment parameters pay attention to external influencing factors, the environment temperature can significantly change the chemical reaction rate of the battery, and the environment humidity can affect the insulation performance and service life of the battery. The inclusion of the two types of parameters can make the prediction model more suitable for actual use scenarios, and the three types of parameters complement each other from different dimensions to jointly constitute a comprehensive and accurate prediction index basis.

[0128] Further, based on the historical operation records of the same type of battery pack as the target battery pack, the relevance of the plurality of prediction indexes and the battery remaining capacity prediction result is evaluated respectively with the current SOC region as a constraint to obtain a plurality of index relevance degrees.

[0129] The same type of battery pack refers to a battery pack of the same model and similar use scenario as the target battery pack, and the historical operation records include index data and corresponding real remaining capacity values at different time nodes. In addition, the current SOC region is used as a constraint because the influence of the same index on the remaining capacity varies at different SOC regions. For example, in the low SOC region (<20%), the internal resistance of the individual cell has a more significant influence on the remaining capacity, and in the high SOC region (≥80%), the relevance of the terminal voltage of the battery pack is stronger.

[0130] Specifically, the correlation degree is determined by calculating the correlation coefficient (i.e. Pearson correlation coefficient) of each index with the true remaining capacity during the evaluation. The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation between the index and the prediction result. If the evaluation in the intermediate SOC region (20%-80%) shows that the correlation coefficient of the battery pack terminal voltage with the remaining capacity is 0.91, the correlation coefficient of the single battery voltage is 0.87, and the correlation coefficient of the environmental temperature is 0.73, the corresponding index correlation degrees are 0.91, 0.87, and 0.73, respectively.

[0131] Further, the prediction indexes are arranged in descending order of index correlation degree to generate a prediction index sequence. Continuing the evaluation results in the above-mentioned intermediate SOC region, if the index correlation degrees from high to low are "battery pack terminal voltage (0.91) → single battery voltage (0.87) → environmental temperature (0.73) → pole temperature (0.68) → battery pack total current (0.65) → single battery internal resistance (0.62) → environmental humidity (0.58)", the prediction index sequence is the above-mentioned order, so as to ensure that the indexes with stronger correlation and more critical to prediction are in the front of the sequence.

[0132] Further, based on a preset prediction depth-index number mapping table, the adaptive index number is determined according to the adaptive prediction depth matching, and the prediction indexes in the prediction index sequence are selected as the adaptive prediction index set.

[0133] The preset prediction depth-index number mapping table is prepared in combination with a large number of prediction experiments. The deeper the adaptive prediction depth, the higher the prediction accuracy required, and the more the matching indexes. For example, when the adaptive prediction depth is marked as depth 1 (the highest accuracy requirement), 6 indexes are matched; when the depth is 2, 4 indexes are matched; and when the depth is 3 (the basic accuracy requirement), 2 indexes are matched.

[0134] For example, if the current adaptive prediction depth is depth 2, the first 4 indexes in the above-mentioned sequence, i.e. "battery pack terminal voltage, single battery voltage, environmental temperature, and pole temperature", are selected to form the adaptive prediction index set, which not only ensures that the indexes cover the key influencing factors, but also avoids excessive redundant indexes increasing the data acquisition cost.

[0135] Meanwhile, the ratio of the current running smoothness coefficient to the average historical running smoothness coefficient of the target battery pack in a historical time range is multiplied by the standard selection model number and rounded to obtain the adaptive selection model number.

[0136] The historical time range can be set to the past 30 days, and the average daily running smoothness coefficient in this period is calculated as the average historical running smoothness coefficient. The standard selection model number is fixed at 5, and the adaptive selection model number needs to satisfy ≥2 and ≤ the total number of models in the battery remaining capacity prediction plug-in (N≥10).

[0137] Exemplarily, the current running smoothness coefficient is 0.5, the historical running smoothness coefficient mean is 0.4, the ratio of the two is 1.25, multiplied by 5 is 6.25, and the integer part is 6; if the total number of models in the plug-in is 12 (satisfying N≥10), 6 is within the range of 2-12, so the number of models selected by adaptation is 6; if the current running smoothness coefficient is 0.2, the historical mean is 0.5, the ratio is 0.4, multiplied by 5 is 2, and the integer part is 2, that is, the number of models selected by adaptation is 2, to ensure that the number of models can support multiple result fitting to improve precision, and will not waste computing resources due to too many models.

[0138] Further, the adaptation prediction index set and the number of models selected by adaptation are used as an adaptation prediction scheme, which clearly defines the range of subsequent data collection and the number of prediction models used, providing clear guidance for the prediction process.

[0139] Further, after determining the adaptation prediction scheme, the adaptation multi-source prediction data of the target battery pack is collected according to the adaptation prediction index set. Taking the adaptation prediction index set of "battery pack terminal voltage, single battery voltage, environmental temperature, and pole temperature" as an example, the terminal voltage data of the battery pack (recorded every 5 seconds), the voltage data of each single battery (recorded every 10 seconds) are collected in real time through the battery management system (BMS), and the pole temperature (recorded every 30 seconds) and the environmental temperature (recorded every 1 minute) are collected through the temperature sensor, to ensure that the collected data correspond to the adaptation prediction index one by one, and the data sampling frequency can meet the prediction accuracy requirement.

[0140] Further, the adaptation battery capacity prediction plug-in is matched and called according to the adaptation prediction index set, the data features are accurately extracted through the model in the plug-in that is adapted to the adaptation prediction index set, and then the subsequent prediction results can be ensured to meet the actual demand.

[0141] The method provided in the embodiment of the application comprises the following steps:

[0142] The historical running records of the same type of battery pack are searched with the adaptation prediction index set as a constraint, a sample multi-source data set is collected, and the historical battery residual capacity corresponding to different sample multi-source data at the same time is taken as a sample battery residual capacity to obtain a sample battery residual capacity set;

[0143] The sample multi-source data set and the sample battery residual capacity set are taken as training data, the training data is equally divided into N parts, and is randomly selected with replacement to obtain N sample training sets;

[0144] The N sample training sets are used to train deep learning models respectively until convergence to obtain N battery capacity prediction models, and an adaptation battery capacity prediction plug-in is constructed by combination.

[0145] Specifically, first, the historical operation records of the same type of battery pack are retrieved as constraints to adapt the prediction indicator set, the sample multi-source data set is collected, and the sample battery remaining capacity set is obtained. Among them, the same type of battery pack is consistent with the model of the target battery pack and has the same use scene, and its historical operation records contain long-term accumulated indicator data and real remaining capacity values at corresponding time nodes.

[0146] In addition, taking the adaptive prediction indicator set as a constraint means that only data related to the adaptive prediction indicator is collected. For example, if the adaptive prediction indicator set is "battery pack terminal voltage, single battery voltage, and environment temperature", only the time series data of these three types of indicators is extracted from the historical operation records, and irrelevant indicators such as pole temperature and environment humidity are excluded to avoid redundant data interference with model training.

[0147] During specific collection, the time continuity and correspondence of sample data need to be ensured. The "terminal voltage 12.5V, single battery voltage 3.1V, and environment temperature 25℃" collected at a certain time are taken as a group of sample multi-source data, and the real remaining capacity 85% detected by professional equipment at that time is recorded as the corresponding sample battery remaining capacity. In this way, thousands or even tens of thousands of groups of data are accumulated to form the sample multi-source data set and the sample battery remaining capacity set, so as to ensure that the amount of training data can support the model to learn the correlation between indicators and remaining capacity.

[0148] Further, the sample multi-source data set and the sample battery remaining capacity set are taken as training data, the training data is equally divided into N parts, and random selection with replacement is performed to obtain N sample training sets.

[0149] Among them, the value of N needs to satisfy ≥10, for example, set N=12, first divide all training data (assuming a total of 12000 groups) into 12 parts, each containing 1000 groups of "sample multi-source data-remaining capacity" corresponding data.

[0150] Further, 12 sample training sets are generated by random selection with replacement (i.e. Bootstrap sampling), wherein each training set still randomly selects 12000 groups of data from the original 12000 groups of data (allowing repeated selection) to ensure that each training set not only retains the overall distribution characteristics of the original data, but also has certain differences. The models trained based on different training sets can cover more data scenarios and improve the generalization ability of the plug-in as a whole.

[0151] For example, the first training set may contain more sample data in the low SOC area, and the second training set contains more sample data in the high SOC area, avoiding the model bias caused by uneven data distribution in a single training set.

[0152] Finally, N battery capacity prediction models (N is greater than or equal to 10) are obtained by training the deep learning model to convergence using N sample training sets, respectively, and a battery capacity prediction plug-in is constructed by combination to accurately capture the dynamic correlation between the adaptive prediction index and the remaining capacity.

[0153] Specifically, a long short-term memory network (LSTM) is used as the basic framework of the deep learning model. Through the synergistic effect of the input gate, the forgetting gate, and the output gate, the LSTM can effectively process the time series data characteristics of the adaptive prediction index and avoid the gradient disappearance problem of traditional recurrent neural networks (RNN).

[0154] When building the network architecture, the network layer number, the number of hidden layer neurons, the input and output dimensions, and other parameters need to be set reasonably in combination with the sample data characteristics and the prediction accuracy requirements. For example, if the adaptive prediction index set contains 3 types of indexes (terminal voltage, single cell voltage, and environmental temperature), the input dimension of the LSTM network is set to 3, corresponding to receiving 3 index data at a single time step; the hidden layer is set to 2 layers, and the number of neurons in each layer is set to 64, which ensures that the network can learn the characteristics sufficiently and avoids the complexity of the architecture leading to low training efficiency; the output dimension is set to 1, which directly outputs the remaining capacity prediction value at the corresponding time.

[0155] Meanwhile, a fully connected layer and a Dropout layer (Dropout probability is set to 0.2) are added after the LSTM layer. The fully connected layer is used to integrate the feature vectors output by the LSTM layer, and the Dropout layer is used to prevent overfitting during training, further improving the model's generalization ability.

[0156] During the model training phase, for each of the N sample training sets, an independent training process is used. First, the sample data is preprocessed, i.e., the sample multi-source data is mapped to the [0, 1] interval according to the Min-Max normalization method, eliminating the influence of different index dimension differences on training; the preprocessed time series data is divided into input sequences according to the "time window = 20", i.e., each input contains 20 consecutive time step index data, corresponding to the remaining capacity value at the end of the time window, so that the model learns the mapping relationship from the index time series change to the remaining capacity.

[0157] In addition, during specific training, the root mean square error (RMSE) is selected as the loss function to quantify the difference between the model's predicted value and the true value of the sample battery's remaining capacity. The Adam optimizer (initial learning rate set to 0.001) is used to update the network parameters, the batch size is set to 32, the training rounds are set to 100 rounds, and the early stopping mechanism (Patience = 10) is introduced. If the RMSE on the validation set does not decrease for 10 consecutive rounds and the model's prediction accuracy on the test set no longer improves, the training is stopped, and the model is determined to have converged.

[0158] Further, after completing the independent training of the N sample training sets according to the above process, N converged LSTM battery capacity prediction models can be obtained, and each model has different prediction performances in different working conditions due to the subtle differences in the training data.

[0159] Finally, the N LSTM models obtained are integrated and packaged to build an adaptive battery capacity prediction plug-in. The plug-in is provided with a model calling interface, and the number of models can be selected according to subsequent adaptation, and a corresponding number of models are randomly selected to participate in prediction. At the same time, a data preprocessing module is built in, which can automatically standardize and time window divide the input adaptive multi-source prediction data, without additional data processing operations, to ensure the convenience of plug-in calling.

[0160] Through the above construction process, the adaptive battery capacity prediction plug-in can not only accurately match the characteristics of the adaptive prediction index set, but also improve the reliability and generalization ability of prediction relying on multi-model design, to ensure that the plug-in can output high-quality remaining capacity prediction results based on adaptive multi-source prediction data when called subsequently.

[0161] Further, the same number of prediction models are randomly selected from the N battery capacity prediction models in the adaptive battery capacity prediction plug-in according to the number of selected models, and battery remaining capacity prediction is performed according to the adaptive multi-source prediction data, and the mean fitting of multiple prediction results is performed to obtain the battery remaining capacity evaluation result.

[0162] Specifically, when selecting battery capacity prediction models, the corresponding number of models is extracted from N (for example, 12) LSTM models based on the number of selected models (for example, 6 calculated in the foregoing) through the random selection algorithm built in the adaptive battery capacity prediction plug-in, and the probability of each model being selected is equal each time to avoid distortion of the prediction results due to model selection bias.

[0163] For example, if the number of selected models is 6, the adaptive battery capacity prediction plug-in will randomly select 6 models numbered 2, 5, 7, 9, 10 and 11 from 12 LSTM models, and these models are trained based on different sample training sets.

[0164] Among them, some models have better prediction accuracy in the high SOC area (≥80%) (error <1.8%), some models are sensitive to environmental temperature fluctuations (error <2.2% in low temperature environment), and some models are good at capturing capacity changes caused by subtle differences in single cell voltage. The combination of multiple models can cover the working condition scenarios that the target battery pack may face.

[0165] Further, after obtaining the selected battery capacity prediction models, the preprocessed adaptive multi-source prediction data is input into each model for independent prediction to avoid the deviation caused by the single model perspective limitation.

[0166] Specifically, first, the data preprocessing module built in the battery capacity prediction plug-in will perform Min-Max standardization on the collected adaptive multi-source prediction data, which is consistent with the model training stage, and map the data to the [0, 1] interval.

[0167] Further, the standardized time series data is divided into input sequences according to the "time window = 20", for example, if 40 consecutive time steps of index data are collected, 2 groups of input sequences will be generated (1-20 steps, 21-40 steps), each sequence corresponds to output a remaining capacity prediction value, and the average of the two prediction values is taken as the final prediction result of the model.

[0168] Taking the battery capacity prediction model No. 5 as an example, after inputting two groups of time series data, the output prediction values are 82.3% and 81.9%, respectively, and the final prediction result of the model is (82.3%+81.9%) / 2=82.1%; the battery capacity prediction model No. 7 outputs prediction values of 81.8% and 82.4%, respectively, and the final prediction result is 82.1%; the battery capacity prediction model No. 9 outputs prediction values of 82.5% and 82.0%, respectively, and the final prediction result is 82.25%, and so on. The six models will output their respective final prediction results, which are assumed to be 82.1%, 81.9%, 82.1%, 82.25%, 81.8%, and 82.0%, respectively.

[0169] Further, after obtaining the prediction results of all selected models, the results are calculated by mean fitting. Specifically, before fitting, possible outliers need to be removed, and then the effective prediction results are summed and divided by the number of models to obtain the final battery remaining capacity evaluation result. Among them, if the prediction result of a model deviates from other results by more than 5% (based on the statistical threshold of prediction error of similar batteries), it is determined as an outlier and is removed, and if there is no outlier, all are involved in the calculation.

[0170] Taking the prediction results of the above six models as an example, the sum is:

[0171] 82.1%+81.9%+82.1%+82.25%+81.8%+82.0%=492.15%, divided by 6 is 82.025%, rounded to two decimal places is 82.03%, which is the remaining capacity evaluation result of the target battery pack.

[0172] Through the above steps, the prediction demand under different working conditions can be covered by relying on the differentiated advantages of multiple models, and the random error of a single model can be offset by mean fitting, so that the remaining capacity evaluation result is closer to the true value. At the same time, a model is randomly selected each time for prediction, avoiding the long-term deviation that may exist in a fixed model combination, ensuring that stable and accurate remaining capacity evaluation results can be output regardless of whether the target battery pack is in a high / low SOC region or a high / low temperature environment, and providing reliable data support for battery operation and maintenance, range determination and other practical applications.

[0173] Through the above specific embodiments, the application achieves the following technical effects:

[0174] The application proposes a deep learning driven battery remaining capacity dynamic evaluation method. First, the current SOC region of the target battery pack is read, the corresponding current prediction error confidence interval is obtained, and the fluctuation range of the prediction error in the current scenario is determined. Then, the multi-source operation data sequence in the historical time zone is monitored and obtained, the current operation smoothness coefficient is output by analyzing the fluctuation degree of the electrical characteristics and the load characteristics, and the current prediction error confidence interval and the current operation smoothness coefficient are combined to compensate the preset standard prediction depth, generate a prediction depth that adapts to the current scenario, and form an adaptive prediction scheme. At the same time, an adaptive battery capacity prediction plug-in containing N LSTM models is constructed. Finally, the adaptive multi-source prediction data is collected according to the adaptive prediction scheme, the model randomly selected from the adaptive battery capacity prediction plug-in is called for prediction, the prediction results are mean fitted, and the battery remaining capacity evaluation result is output, realizing the accurate and dynamic evaluation of the battery remaining capacity of the target battery pack in different SOC regions and different operation stability scenarios.

[0175] The method provided by the application solves the problems in the traditional battery remaining capacity evaluation, such as information redundancy or loss caused by fixed indicators, difficulty of a single model in adapting to complex working conditions, and mismatch between prediction depth and actual scenario, improves the accuracy and dynamic adaptability of battery remaining capacity evaluation, and provides technical support for battery operation and maintenance and range determination in new energy vehicle and energy storage power station scenarios.

[0176] Embodiment two, as shown in the accompanying Figure 2 Based on the inventive concept of the deep learning driven battery remaining capacity dynamic evaluation method provided in embodiment one, the application further provides a deep learning driven battery remaining capacity dynamic evaluation system, which specifically includes:

[0177] The SOC and error reading module 01 is configured to read a current SOC region in which the target battery pack is located, and match a current prediction error confidence interval for obtaining a battery remaining capacity prediction.

[0178] The multi-source data stationary analysis module 02 is configured to monitor a multi-source operation data sequence of the target battery pack in a historical time zone, perform operation state stationarity analysis according to the multi-source operation data sequence, and output a current operation stationarity coefficient.

[0179] The prediction depth compensation module 03 is configured to compensate a preset standard prediction depth according to the current prediction error confidence interval and the current operation stationarity coefficient, and generate an adaptive prediction depth.

[0180] The prediction scheme execution module 04 is configured to formulate an adaptive prediction scheme according to the current operation stationarity coefficient and the adaptive prediction depth based on a preset prediction index space, collect adaptive multi-source prediction data according to the adaptive prediction scheme, and call a battery remaining capacity prediction plug-in constructed based on deep learning to perform battery remaining capacity prediction on the target battery pack according to the adaptive multi-source prediction data, and output a battery remaining capacity evaluation result.

[0181] In one embodiment, the SOC and error reading module 01 is further configured to:

[0182] The SOC region includes a high SOC region, an intermediate SOC region and a low SOC region, wherein the high SOC region is SOC greater than or equal to 80%, the intermediate SOC region is SOC greater than or equal to 20% and less than 80%, and the low SOC region is SOC less than 20%, and each SOC region is identified with a prediction error confidence interval.

[0183] Further, the SOC and error reading module 01 further includes:

[0184] In the high SOC region, the middle SOC region and the low SOC region, any SOC region is randomly selected as a first SOC region; a current aging index of the target battery pack is obtained, and the current aging index is expanded according to a preset tolerance interval to obtain an aging index interval, wherein the current aging index is determined based on historical charge and discharge cycle times and historical deep discharge times; the first SOC region and the aging index interval are used as constraints to retrieve historical operation records of the same type of battery pack of the target battery pack to obtain a sample predicted SOC value set and a sample real SOC value set under different operation conditions; based on the sample predicted SOC value set and the sample real SOC value set, a difference between each sample predicted SOC value and a corresponding sample real SOC value is taken as a sample error to obtain a sample error set, and an error average value and an error standard deviation of the sample error set are calculated; a corresponding standard score is determined according to a preset confidence level, a lower limit of a confidence interval is obtained by subtracting the product of the standard score and the error standard deviation from the error average value, and an upper limit of the confidence interval is obtained by adding the product of the standard score and the error standard deviation to the error average value, to obtain a first prediction error confidence interval of the first SOC region.

[0185] In one embodiment, the multi-source data stationary analysis module 02 is further used for:

[0186] The multi-source operation data sequence of the target battery pack in the historical time zone is monitored, wherein the multi-source operation data sequence includes an electrical characteristic data sequence and a load characteristic data sequence, the electrical characteristic data sequence includes a voltage sequence, a current sequence and a power sequence, and the load characteristic data sequence includes a load current sequence and a load power sequence; index fluctuation calculation is performed according to the voltage sequence, the current sequence and the power sequence respectively to obtain a voltage fluctuation coefficient, a current fluctuation coefficient and a power fluctuation coefficient, and a weighted sum is obtained to obtain an electrical characteristic fluctuation degree; index fluctuation calculation is performed according to the load current sequence and the load power sequence respectively, and a weighted sum is obtained to obtain a load characteristic fluctuation degree; a current operation stationary coefficient is determined by weighted evaluation according to the electrical characteristic fluctuation degree and the load characteristic fluctuation degree, wherein the current operation stationary coefficient is negatively correlated with the electrical characteristic fluctuation degree and the load characteristic fluctuation degree.

[0187] In one embodiment, the prediction depth compensation module 03 is further used for:

[0188] The interval width ratio of the current prediction error confidence interval and a preset standard prediction error confidence interval is taken as a first prediction depth compensation coefficient; the ratio of a preset standard running stationary coefficient and the current running stationary coefficient is taken as a second prediction depth compensation coefficient; the first prediction depth compensation coefficient and the second prediction depth compensation coefficient are weighted and fused to obtain an overall prediction depth compensation coefficient; and the product of the overall prediction depth compensation coefficient and a preset standard prediction depth is taken as an adaptive prediction depth.

[0189] In one embodiment, the prediction scheme execution module 04 is further configured to:

[0190] A preset prediction index space of the target battery pack is obtained, wherein the preset prediction index space includes a plurality of prediction indexes; based on the historical operation records of the same type battery pack as the target battery pack, the relevance of the plurality of prediction indexes and the battery residual capacity prediction result is evaluated respectively with the current SOC region as a constraint to obtain a plurality of index correlation degrees; the plurality of prediction indexes are arranged in descending order of index correlation degrees to generate a prediction index sequence; based on a preset prediction depth-index number mapping table, the adaptive index number is determined according to the adaptive prediction depth matching, and the first adaptive index number of prediction indexes in the prediction index sequence is selected as an adaptive prediction index set; the ratio of the current running stationary coefficient and the historical running stationary coefficient average of the target battery pack in the historical time range is multiplied by a standard selection model number and rounded to obtain an adaptive selection model number, wherein the standard selection model number is 5, the adaptive selection model number is greater than or equal to 2 and less than or equal to the total number of models in the battery residual capacity prediction plug-in; the adaptive prediction index set and the adaptive selection model number are taken as an adaptive prediction scheme. The adaptive multi-source prediction data of the target battery pack is collected according to the adaptive prediction index set; the adaptive battery capacity prediction plug-in is matched and called according to the adaptive prediction index set, wherein the adaptive battery capacity prediction plug-in includes N battery capacity prediction models, and N is greater than or equal to 10; the same number of prediction models are randomly selected from the N battery capacity prediction models of the adaptive battery capacity prediction plug-in according to the adaptive selection model number, the battery residual capacity is predicted according to the adaptive multi-source prediction data, and the mean value fitting is performed on a plurality of prediction results to obtain a battery residual capacity evaluation result.

[0191] Further, the prediction scheme execution module 04 further includes:

[0192] The several prediction indexes include group-level parameters, single-body parameters and environment parameters, wherein the target battery group includes a plurality of single-body batteries, the group-level parameters at least include terminal voltage and total current of the battery group, the single-body parameters at least include battery voltage, pole temperature and internal resistance of the single-body battery, and the environment parameters at least include environment temperature and environment humidity.

[0193] Further, the prediction scheme execution module 04 further includes:

[0194] The historical operation records of the same type battery group are retrieved with the adaptive prediction index set as a constraint, a sample multi-source data set is collected, and the historical battery residual capacity corresponding to different sample multi-source data at the same time is taken as a sample battery residual capacity to obtain a sample battery residual capacity set; the sample multi-source data set and the sample battery residual capacity set are taken as training data, the training data is equally divided into N parts, and is randomly selected with replacement to obtain N sample training sets;

[0195] The N sample training sets are used to train the deep learning model to convergence respectively to obtain N battery capacity prediction models, and an adaptive battery capacity prediction plug-in is constructed by combination.

[0196] It should be noted that the above sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above describes the specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0197] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0198] The present application and the drawings are only exemplary descriptions of the present application, and are considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalents, the present application intends to include these modifications and changes.

Claims

1. A deep learning-driven dynamic evaluation method for remaining battery capacity, characterized in that the method... include: Read the current SOC region of the target battery pack and match it to obtain the current prediction error confidence interval of the remaining battery capacity prediction; The target battery pack is monitored and acquired in a historical time zone using multi-source operating data sequences. Based on the multi-source operating data sequences, an operational stability analysis is performed, and the current operational stability coefficient is output. Based on the current prediction error confidence interval and the current operational stability coefficient, the preset standard prediction depth is compensated to generate an adapted prediction depth; Based on the preset prediction index space, an adaptation prediction scheme is formulated according to the current operating stability coefficient and the adaptation prediction depth. Adaptation multi-source prediction data is collected according to the adaptation prediction scheme, and a battery remaining capacity prediction plugin built based on deep learning is called. The remaining battery capacity of the target battery pack is predicted according to the adaptation multi-source prediction data, and the battery remaining capacity evaluation result is output.

2. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 1, characterized in that, The SOC region includes a high SOC region, an intermediate SOC region, and a low SOC region. The high SOC region is defined as having an SOC greater than or equal to 80%, the intermediate SOC region is defined as having an SOC greater than or equal to 20% and less than 80%, and the low SOC region is defined as having an SOC less than 20%. Each SOC region is also marked with a prediction error confidence interval.

3. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 2, characterized in that, The calculation process for the prediction error confidence interval includes: Randomly select any one of the high SOC region, intermediate SOC region and low SOC region as the first SOC region; The current aging index of the target battery pack is obtained, and the current aging index is expanded according to a preset tolerance range to obtain an aging index range, wherein the current aging index is determined based on the historical charge-discharge cycle count and the historical deep discharge count. Using the first SOC region and aging index range as constraints, retrieve historical operating records of similar battery packs of the target battery pack to obtain a set of predicted SOC values ​​and a set of actual SOC values ​​under different operating conditions. Based on the set of predicted SOC values ​​and the set of actual SOC values, the difference between the predicted SOC value and the corresponding actual SOC value of each sample is taken as the sample error, and the set of sample errors is obtained. The mean error and standard deviation of the set of sample errors are then calculated. The corresponding standard score is determined according to the preset confidence level. The lower limit of the confidence interval is obtained by subtracting the product of the standard score and the standard deviation of the error from the average error. The upper limit of the confidence interval is obtained by adding the product of the standard score and the standard deviation of the error to the average error. The first prediction error confidence interval of the first SOC region is obtained.

4. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 1, characterized in that, The system monitors and acquires multi-source operational data sequences of the target battery pack within a historical time zone. Based on these multi-source operational data sequences, it performs operational stability analysis and outputs the current operational stability coefficient, including: The target battery pack is monitored and acquired in a historical time zone using multi-source operating data sequences, wherein the multi-source operating data sequences include electrical characteristic data sequences and load characteristic data sequences. The electrical characteristic data sequences include voltage sequences, current sequences, and power sequences, and the load characteristic data sequences include load current sequences and load power sequences. The index volatility is calculated based on the voltage sequence, current sequence, and power sequence to obtain the voltage fluctuation coefficient, current fluctuation coefficient, and power fluctuation coefficient, and then the electrical characteristic volatility is obtained by weighted summation. The load characteristic fluctuation is obtained by calculating the index fluctuation based on the load current sequence and the load power sequence, and then summing them by weight. The current operational stability coefficient is determined based on a weighted evaluation of the electrical characteristic fluctuation and the load characteristic fluctuation, wherein the current operational stability coefficient is negatively correlated with the electrical characteristic fluctuation and the load characteristic fluctuation.

5. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 1, characterized in that, Based on the current prediction error confidence interval and the current operational stability coefficient, the preset standard prediction depth is compensated to generate an adapted prediction depth, including: The ratio of the width of the current prediction error confidence interval to the width of the preset standard prediction error confidence interval is used as the first prediction depth compensation coefficient. The ratio of the preset standard operating stability coefficient to the current operating stability coefficient is set as the second prediction depth compensation coefficient; The overall prediction depth compensation coefficient is obtained by weighted fusion of the first prediction depth compensation coefficient and the second prediction depth compensation coefficient. The product of the overall prediction depth compensation coefficient and the preset standard prediction depth is used as the adaptive prediction depth.

6. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 1, characterized in that, Based on a preset prediction index space, an adaptation prediction scheme is formulated according to the current operational stability coefficient and the adaptation prediction depth, including: Obtain a preset prediction index space for the target battery pack, wherein the preset prediction index space includes several prediction indices. Based on the historical operating records of similar battery packs of the target battery pack, and with the current SOC region as a constraint, the correlation between the prediction results of the several prediction indicators and the remaining battery capacity is evaluated to obtain the correlation degree of the several indicators. The prediction indicators are arranged in descending order of their correlation to generate a prediction indicator sequence. Based on the preset prediction depth-index quantity mapping table, the number of adaptive indicators is determined according to the adaptive prediction depth matching, and the predicted indicators of the previous number of adaptive indicators are selected from the prediction indicator sequence as the adaptive prediction indicator set. The ratio of the current operating stability coefficient to the average historical operating stability coefficient of the target battery pack within the historical time range is multiplied by the number of standard selected models and rounded to obtain the number of adaptive selected models. The number of standard selected models is 5, and the number of adaptive selected models is greater than or equal to 2 and less than or equal to the total number of models in the battery remaining capacity prediction plugin. The set of adaptation prediction indicators and the number of adaptation selection models are used as the adaptation prediction scheme.

7. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 6, characterized in that, The predicted parameters include group-level parameters, individual cell parameters, and environmental parameters. The target battery pack includes multiple individual cells. The group-level parameters include at least the battery pack's terminal voltage and total current. The individual cell parameters include at least the individual cell's battery voltage, terminal temperature, and internal resistance. The environmental parameters include at least the ambient temperature and ambient humidity.

8. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 6, characterized in that, According to the aforementioned adaptation prediction scheme, multi-source prediction data is collected, and a battery remaining capacity prediction plugin based on deep learning is invoked. Based on the multi-source prediction data, the remaining battery capacity of the target battery pack is predicted, and the battery remaining capacity assessment result is output, including: The multi-source prediction data for the target battery pack is collected and obtained according to the set of adaptation prediction indicators. The adaptive battery capacity prediction plugin is called according to the adaptive prediction index set, wherein the adaptive battery capacity prediction plugin includes N battery capacity prediction models, where N is greater than or equal to 10. According to the number of models selected for adaptation, the same number of prediction models are randomly selected from the N battery capacity prediction models of the adapted battery capacity prediction plugin. The remaining battery capacity is predicted according to the adapted multi-source prediction data, and the mean of multiple prediction results is fitted to obtain the remaining battery capacity assessment result.

9. The deep learning-driven dynamic evaluation method for remaining battery capacity according to claim 8, characterized in that, The method for constructing the adaptive battery capacity prediction plugin includes: Using the set of adaptation prediction indicators as constraints, retrieve historical operation records of similar battery packs, collect multi-source sample datasets, and use the historical battery remaining capacity corresponding to different multi-source sample data at the same time as the sample battery remaining capacity to obtain the sample battery remaining capacity set. The multi-source dataset and the remaining battery capacity dataset are used as training data. The training data is divided into N equal parts and randomly selected with replacement to obtain N sample training sets. The deep learning models are trained to convergence using the N sample training sets to obtain N battery capacity prediction models, which are then combined to construct an adaptive battery capacity prediction plugin.

10. A deep learning-driven dynamic evaluation system for remaining battery capacity, characterized in that, The system is used to execute the deep learning-driven dynamic evaluation method for battery remaining capacity according to any one of claims 1-9, and the system comprises: The SOC and error reading module is used to read the current SOC region of the target battery pack and match it to obtain the current prediction error confidence interval of the remaining battery capacity prediction. The multi-source data stability analysis module is used to monitor and acquire the multi-source operating data sequence of the target battery pack in the historical time zone, perform operating status stability analysis based on the multi-source operating data sequence, and output the current operating stability coefficient. The prediction depth compensation module is used to compensate the preset standard prediction depth based on the current prediction error confidence interval and the current operating stability coefficient, and generate an adapted prediction depth. The prediction scheme execution module is used to formulate an adaptation prediction scheme based on a preset prediction index space, the current operating stability coefficient, and the adaptation prediction depth; collect and acquire adaptation multi-source prediction data according to the adaptation prediction scheme; and call a battery remaining capacity prediction plugin built based on deep learning to predict the battery remaining capacity of the target battery pack according to the adaptation multi-source prediction data, and output the battery remaining capacity assessment result.

Images