A battery state of charge estimation method, device and storage medium

Abstract

The application discloses a battery state of charge estimation method, a computer device and a storage medium, comprising the steps of establishing a time sequence hybrid neural network model, obtaining at least one to-be-estimated discharge data of a to-be-estimated battery, processing each to-be-estimated discharge data by using the time sequence hybrid neural network model, performing physical constraint filtering processing on the processing result of the time sequence hybrid neural network model, and obtaining a state of charge estimation value of the to-be-estimated battery. By using the hybrid time sequence neural network model and the physical constraint filtering processing, the application can capture the time sequence characteristics in the to-be-estimated discharge data, thereby accurately estimating the state of charge estimation value and improving the estimation accuracy of the SOC. In addition, the application additionally applies the filtering with the physical constraint to the output of the time sequence hybrid neural network, thereby constraining and smoothing the output result of the time sequence hybrid neural network, and further improving the estimation accuracy of the model. The application is widely applied in the technical field of battery management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, and in particular to a battery state of charge estimation method, a computer device and a storage medium. BACKGROUND

[0002] Batteries are widely used to power devices such as consumer electronics, industrial electronics and electric devices. Especially under the demand of developing new energy, batteries are widely used in power storage, transportation and aerospace equipment and other fields. The battery management system is crucial to maintaining the safe and stable operation of the battery, and the detection and management of the state of charge (SOC) of the battery is the most critical. However, there is still a lack of technical means to directly measure the state of charge of the battery, and generally only the state of charge of the battery can be estimated by detecting other parameters of the battery. However, due to the influence of various factors on the state of charge of the battery, the current technology has the defects of large estimation deviation and unstable results. SUMMARY

[0003] In view of the technical problems of large estimation deviation and unstable results of the state of charge of the battery, the purpose of the present application is to provide a battery state of charge estimation method, a computer device and a storage medium.

[0004] In one aspect, the present application includes a battery state of charge estimation method, which comprises the following steps:

[0005] establishing a time sequence hybrid neural network model; the time sequence hybrid neural network model is used for time sequence feature capture of input data, learning of time dependence relationship of the captured time sequence features, and regression prediction of the learned time dependence relationship;

[0006] obtaining at least one to-be-estimated discharge data of a to-be-estimated battery;

[0007] processing each of the to-be-estimated discharge data using the time sequence hybrid neural network model;

[0008] performing physical constraint filtering processing on the processing result of the time sequence hybrid neural network model to obtain an estimated value of the state of charge of the to-be-estimated battery.

[0009] Further, the establishment of the time sequence hybrid neural network model comprises:

[0010] establish a time sequence convolutional neural network, a long short-term memory network and a full connection layer; the time sequence convolutional neural network is used for time sequence feature capturing of input data, the long short-term memory network is used for learning time dependence of the time sequence feature captured by the time sequence convolutional neural network, and the full connection layer is used for regression prediction of the time dependence learned by the long short-term memory network;

[0011] sequentially connect the time sequence convolutional neural network, the long short-term memory network and the full connection layer to obtain the time sequence hybrid neural network model.

[0012] Further, the time sequence hybrid neural network model is established, including:

[0013] obtain sample discharge data and sample state of charge data of a sample battery at a corresponding same time;

[0014] train the time sequence hybrid neural network model by using the sample discharge data and the sample state of charge data.

[0015] Further, the time sequence hybrid neural network model is trained by using the sample discharge data and the sample state of charge data, including:

[0016] establish a plurality of hyperparameter sets; each of the hyperparameter sets includes a plurality of hyperparameters;

[0017] perform a plurality of rounds of iteration processes until an iteration end condition is met;

[0018] return the hyperparameter set corresponding to the to-be-evaluated model with the best performance evaluation in the last round of the iteration process;

[0019] set the time sequence hybrid neural network model by using the returned hyperparameter set to obtain a trained time sequence hybrid neural network model;

[0020] wherein, any round of the iteration process includes the following steps:

[0021] set the time sequence hybrid neural network model by using each of the hyperparameter sets respectively to obtain a plurality of to-be-evaluated models; wherein, any to-be-evaluated model is obtained by setting the time sequence hybrid neural network model by using a corresponding one of the hyperparameter sets;

[0022] train each of the to-be-evaluated models by using the sample discharge data and the sample state of charge data respectively;

[0023] perform performance evaluation on each of the trained to-be-evaluated models respectively;

[0024] For at least one of the to-be-evaluated models with the worst performance evaluation result, update the hyperparameter set corresponding to each of the to-be-evaluated models.

[0025] Further, the physically-constrained filtering processing on the processing result of the time-series hybrid neural network model obtains an estimated state of charge value of the to-be-estimated battery, including:

[0026] Determine a filter window size and a first time point;

[0027] Determine a plurality of second time points within a time length equivalent to the filter window size before the first time point; the first time point and each of the second time points are respectively a collection time point of a corresponding one of the to-be-estimated discharge data;

[0028] Obtain to-be-filtered data corresponding to each of the second time points;

[0029] Arrange each of the to-be-filtered data into a first time sequence according to time sequence;

[0030] For any to-be-filtered data, convert according to a physical constraint suffered by the to-be-estimated battery between the first time point and the second time point corresponding to the to-be-filtered data to obtain a corresponding conversion value;

[0031] Arrange each of the conversion values into a second time sequence according to the order of the first time sequence;

[0032] Add the processing result of the time-series hybrid neural network model on the to-be-estimated discharge data collected at the first time point to the end of the second time sequence to obtain a third time sequence;

[0033] Perform weighted smoothing filtering on the third time sequence to obtain the state of charge estimate value corresponding to the first time point.

[0034] Further, the obtaining of to-be-filtered data corresponding to each of the second time points includes:

[0035] For any second time point:

[0036] When the second time point has a corresponding state of charge estimate value, use the state of charge estimate value corresponding to the second time point as the to-be-filtered data corresponding to the second time point;

[0037] When the second time point does not have a corresponding state of charge estimate value, use the processing result of the time-series hybrid neural network model on the to-be-estimated discharge data collected at the second time point as the to-be-filtered data corresponding to the second time point.

[0038] Further, the conversion according to the physical constraint suffered by the battery to be estimated between the first time and the second time corresponding to the data to be filtered obtains a corresponding conversion value, comprising:

[0039] Obtaining a battery parameter of the battery to be estimated;

[0040] Taking the data to be filtered as an initial value, taking the second time as the lower limit of integration, taking the first time as the upper limit of integration, and performing ampere-hour integration according to the battery parameter to obtain the conversion value.

[0041] Further, the weighted smoothing filtering of the third time sequence obtains the state of charge estimation value corresponding to the first time, comprising:

[0042] According to the third time sequence, a Gaussian weight function is calculated to obtain a weight set;

[0043] According to the weight set, the third time sequence is weighted and averaged to obtain the state of charge estimation value corresponding to the first time.

[0044] On the other hand, the embodiment of the application also includes a computer device comprising a memory and a processor, the memory is used to store at least one program, and the processor is used to load at least one program to execute the battery state of charge estimation method in the embodiment.

[0045] On the other hand, the embodiment of the application also includes a computer readable storage medium, wherein a program executable by a processor is stored, and the program executable by the processor is used to execute the battery state of charge estimation method in the embodiment when executed by the processor.

[0046] The battery state of charge estimation method in the embodiment has the beneficial effects that: the battery state of charge estimation method in the embodiment uses a hybrid time series neural network model and a physically constrained filtering process, combines the two algorithms to improve the estimation ability and accuracy of the state of charge estimation value of the battery to be estimated, the hybrid time series neural network model can capture the time sequence characteristics in the discharge data to be estimated, thereby accurately estimating the state of charge estimation value, and the accuracy of the estimated SOC can be improved. In addition, the output of the time sequence hybrid neural network is additionally subjected to a physically constrained filtering process, which can constrain and smooth the output of the time sequence hybrid neural network, and further improve the accuracy and stability of the model estimation. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The schematic diagram of the steps of the battery state of charge estimation method in the embodiment;

[0048] Figure 2 The principle schematic diagram of training the time sequence hybrid neural network model in the embodiment;

[0049] Figure 3 A process diagram for obtaining a state of charge estimation value of a battery to be estimated using a trained temporal hybrid neural network model in the embodiment. DETAILED DESCRIPTION

[0050] Terminology explanation:

[0051] Data-driven model: Data-driven is a model that adapts to data, and the model changes according to different data. Typical data-driven models include deep data network models, etc.

[0052] Neural network: Neural network is a computational model that simulates the working mode of human brain, which is the basis of deep learning and machine learning. Neural network is composed of a large number of nodes (or called "neurons"), which are connected in the network, can process complex data input, and perform various tasks such as classification, regression, pattern recognition, etc.

[0053] Battery management system: Battery management system is an electronic device that can monitor and manage storage battery. Through the collection and calculation of parameters such as voltage, current, temperature and SOC, the charging and discharging process of the battery is controlled, and the protection and comprehensive performance of the battery are improved.

[0054] SOC: Sate of charge (state of charge), which is used to describe the current battery capacity, which can be described as "remaining available capacity / maximum available capacity ratio".

[0055] TCN: Temporal convolutional network (temporal convolutional neural network), TCN model is a time series prediction model based on convolutional neural network, which has certain advantages in processing time series data compared with CNN.

[0056] LSTM: Long Short-Term Memory (Long Short-Term Memory), which is a time recurrent neural network, suitable for processing and predicting important events with relatively long interval and delay in time series.

[0057] Fully connected layer: Fully connected layer plays the role of "classifier" in the whole convolutional neural network. If the convolutional layer, pooling layer and activation function operations are to map the original data to the hidden layer feature space, the fully connected layer plays the role of mapping the learned "distributed feature representation" to the sample label space.

[0058] The neural network has the function of extracting feature information in complex data for classification or feature recognition, and therefore can be considered to use the neural network to identify some accurately measurable data of the battery to estimate the state of charge of the battery. However, if only a single neural network model is used for SOC estimation, or the time sequence characteristics of the battery SOC are not considered in the process of designing the model framework, the efficiency and accuracy of the model are prone to be low. Moreover, the SOC value estimated by the data-driven neural network model often fluctuates up and down, which is inconsistent with the actual SOC curve of the battery.

[0059] Based on the above principle, the battery state of charge estimation method provided in the embodiment is as follows. Figure 1 The battery state of charge estimation method comprises the following steps:

[0060] S1. Establishing a time sequence hybrid neural network model;

[0061] S2. Obtaining at least one to-be-estimated discharge data of a to-be-estimated battery;

[0062] S3. Processing each to-be-estimated discharge data using the time sequence hybrid neural network model;

[0063] S4. Performing physical constraint filtering processing on the processing result of the time sequence hybrid neural network model to obtain an estimated state of charge value of the to-be-estimated battery.

[0064] The battery state of charge estimation method in the embodiment can be executed by a computer.

[0065] In step S1, the time sequence hybrid neural network model has the performance of capturing time sequence characteristics of input data input to the time sequence hybrid neural network model, learning time dependence relationship of the captured time sequence characteristics, and making regression prediction on the learned time dependence relationship, and can effectively extract the time sequence characteristics of the input data. Specifically, in the training phase of the time sequence hybrid neural network model, the input data of the time sequence hybrid neural network model can be sample discharge data detected in the discharge process of a sample battery; in the phase of estimating the state of charge of the to-be-estimated battery using the time sequence hybrid neural network model, the input data of the time sequence hybrid neural network model is to-be-estimated discharge data detected in the discharge process of the to-be-estimated battery.

[0066] In the embodiment, the sample battery is a battery used to collect data to train the time sequence hybrid neural network model, and the to-be-estimated battery is a battery whose state of charge needs to be estimated using the trained time sequence hybrid neural network model. The sample battery and the to-be-estimated battery can both be lithium ion batteries.

[0067] In this embodiment, the time sequence hybrid neural network model established in step S1 includes a time sequence convolutional neural network, a long short-term memory network and a full connection layer, and the three parts are connected in sequence. The time sequence convolutional neural network is used for time sequence feature capturing of the input data, the long short-term memory network is used for learning time dependence relationship of the time sequence features captured by the time sequence convolutional neural network, and the full connection layer is used for regression prediction of the time dependence relationship learned by the long short-term memory network.

[0068] In this embodiment, after the time sequence hybrid neural network model is established in step S1, the time sequence hybrid neural network model can be trained. The principle of training the time sequence hybrid neural network model is as shown in Figure 2

[0069] Referring to Figure 2 , the discharge period of the sample battery can be measured by using an interpolation sampling method in the charge-discharge cycle period of the sample battery in a laboratory environment or the like. For example, for any time t0, the discharge voltage, discharge current and working temperature values of the sample battery are detected in a period of time before the time t0 (including times t, t+1, ……t0-1, etc., which can be at intervals of seconds), to obtain sample discharge data of the sample battery at the time t0 , and the sample state of charge data of the sample battery at the time t0 is determined by using a coulomb meter or the like , so as to form a pair of training data. Since the time t0 is arbitrary, a plurality of pairs of training data can be obtained based on the same principle. After removing abnormal values and standardizing the training data, the training data are divided into an SOC estimation model training set and an SOC estimation model verification set according to a ratio of 4:1.

[0070] In training the time sequence hybrid neural network model, taking the sample discharge data collected at the time t0 as an example, referring to Figure 2 , the sample discharge data collected at the time t0 is input into the time sequence hybrid neural network, the input data is subjected to time sequence feature capturing by the time sequence convolutional neural network TCN, the time sequence features captured by the time sequence convolutional neural network are subjected to learning of time dependence relationship by the long short-term memory network, and the time dependence relationship learned by the long short-term memory network is subjected to regression prediction by the full connection layer. The output of the full connection layer represents an estimation of the state of charge of the sample battery at the time t0 made by the time sequence hybrid neural network according to the input data , that is, an estimation value y" i , and the sample state of charge data of the sample battery at the time t0 is taken as a true value y i ​After inputting n sample discharge data, n estimated values will be obtained, and using n sample state of charge data as true values, the time series hybrid neural network model can be evaluated according to error functions such as

[0071] Root mean square error (RMSE):

[0072] And mean absolute error (MAE):

[0073] If the error function value is less than the threshold value, it indicates that the training of the time series hybrid neural network model is completed.

[0074] In this embodiment, refer to Figure 2 During the training of the time series hybrid neural network model, a hyperparameter optimization algorithm based on cluster training can be applied. Specifically, it includes the following steps:

[0075] S101. Establish a plurality of hyperparameter sets; each hyperparameter set includes a plurality of hyperparameters;

[0076] S102. Perform a plurality of rounds of iteration processes until the iteration end condition is met;

[0077] S103. Return the hyperparameter set corresponding to the to-be-evaluated model with the best performance evaluation in the last round of iteration processes;

[0078] S104. Use the returned hyperparameter set to set the time series hybrid neural network model to obtain a trained time series hybrid neural network model.

[0079] In step S101, potential hyperparameters such as learning rate, network layer number and neuron can be selected to form a hyperparameter set in the format (learning rate, network layer number, neuron). Since learning rate, network layer number and neuron can take different values respectively, each combination of values can form a hyperparameter set, so multiple hyperparameter sets can be formed.

[0080] In step S102, the iteration round number L and the update rate rate are set. Taking the kth iteration process as an example, the time series hybrid neural network model is set using the super parameter set updated by the (k-1)th iteration process, and a plurality of to-be-evaluated models are obtained. For example, a plurality of super parameter sets (learning rate 1, network layer number 1, neuron 1), (learning rate 2, network layer number 2, neuron 2), (learning rate 3, network layer number 3, neuron 3) and the like are obtained after the (k-1)th iteration process, a time series hybrid neural network model with super parameters (learning rate 1, network layer number 1, neuron 1) and the like is established, thereby obtaining a to-be-evaluated model 1, a time series hybrid neural network model with super parameters (learning rate 2, network layer number 2, neuron 2) and the like is established, thereby obtaining a to-be-evaluated model 2, a time series hybrid neural network model with super parameters (learning rate 3, network layer number 3, neuron 3) and the like is established, thereby obtaining a to-be-evaluated model 3, and a plurality of to-be-evaluated models are obtained. Then, the sample discharge data and the sample state of charge data are used to train the to-be-evaluated models, and a plurality of trained to-be-evaluated models, such as a trained to-be-evaluated model 1, a trained to-be-evaluated model 2, a trained to-be-evaluated model 3 and the like, are obtained. Then, the performance of the trained to-be-evaluated models is evaluated, for example, the root mean square error and the mean absolute error of the trained to-be-evaluated model 1, the root mean square error and the mean absolute error of the trained to-be-evaluated model 2, the root mean square error and the mean absolute error of the trained to-be-evaluated model 3 and the like are obtained, and a batch of to-be-evaluated models with the worst performance evaluation result (the largest error) and a batch of to-be-evaluated models with the best performance evaluation result (the smallest error) are selected. Assuming that the to-be-evaluated model with the best performance evaluation result is the to-be-evaluated model 3 and the to-be-evaluated model with the worst performance evaluation result is the to-be-evaluated model 5, the super parameters (learning rate 5, network layer number 5, neuron 5) of the to-be-evaluated model 5 with the worst performance evaluation result can be updated using the update rate rate, for example, (learning rate 5, network layer number 5, neuron 5) is set to a random number to complete the update, or the super parameters (learning rate 5, network layer number 5, neuron 5) are fine-tuned using the super parameters (learning rate 3, network layer number 3, neuron 3) of the to-be-evaluated model 3 with the best performance evaluation result, so that (learning rate 5, network layer number 5, neuron 5) is updated to (learning rate 5, network layer number 3, neuron 5) and the like to complete the update.

[0081] In the above process, if the kth iteration process is the first iteration process, (learning rate 1, network layer number 1, neuron 1) and the like can be a random number.

[0082] After the kth iteration process is performed, L is set to L-1, and it is determined whether L>0. If yes, the (k+1)th iteration process is performed, and if L≤0, the next iteration process is not performed, and step S103 is performed.

[0083] In step S103, the hyperparameter set corresponding to the model to be evaluated with the best performance evaluation in the last iteration process is returned. Assuming that the model to be evaluated with the hyperparameter set of (learning rate 4, network layer number 3, neuron 5) has the best performance evaluation, in step S104, the hyperparameter set returned in step S103 is used to set the time series hybrid neural network model, that is, a time series hybrid neural network model with parameters such as (learning rate 4, network layer number 3, neuron 5) is established as the trained time series hybrid neural network model.

[0084] The trained time series hybrid neural network model is used to obtain the state of charge estimation value of the battery to be estimated, and the process is as shown in Figure 3

[0085] In step S2, the estimated discharge data of the battery to be estimated to be obtained can have the same format as the sample discharge data. For example, assuming that it is desired to detect the state of charge estimation value of the battery to be estimated at the current time t0, the discharge voltage, discharge current and working temperature values at the current time t0 and a period of time before the current time t0 (including times t, t+1, … t0-1, etc., which can be spaced apart by seconds) can be obtained, thereby obtaining the estimated discharge data of the battery to be estimated at time t0

[0086] In step S2, the estimated discharge data of the battery to be estimated at the current time t0 is input into the time series hybrid neural network model. Since the trained time series hybrid neural network model has the performance of extracting the state of charge feature from the discharge data, the processing result of the time series hybrid neural network model on the estimated discharge data can represent the state of charge estimation value of the battery to be estimated at time t0.

[0087] In the present embodiment, if the processing result of the time series hybrid neural network model on the estimated discharge data at the current time t0 is directly used as the state of charge estimation value of the battery to be estimated at time t0, since the current time t0 is arbitrary, as time elapses, the time t0 changes constantly, and each time the estimated discharge data measured can be different, each time the estimated discharge data measured can be processed using the time series hybrid neural network model in step S3, thereby causing the state of charge estimation value at different times to jump greatly, thereby not conforming to the actual SOC change of the battery to be estimated. Therefore, in the present embodiment, step S4 can be performed to realize smoothing.

[0088] ​Specifically, in the embodiment, when the step S4, that is, the physically constrained filtering processing on the processing result of the time sequence hybrid neural network model is performed to obtain the state of charge estimation value of the battery to be estimated, the following steps can be performed:

[0089] S401. Determine the filter window size and the first time;

[0090] S402. Determine a plurality of second times within a time length equivalent to the filter window size before the first time;

[0091] S403. Obtain the respective second corresponding to-be-filtered data of each second time;

[0092] S404. Arrange the to-be-filtered data according to the time sequence into a first time sequence;

[0093] S405. For any to-be-filtered data, a corresponding conversion value is obtained according to the physical constraint suffered by the battery to be estimated between the first time and the second time corresponding to the to-be-filtered data;

[0094] S406. Arrange the conversion values according to the order of the first time sequence into a second time sequence;

[0095] S407. Add the processing result of the time sequence hybrid neural network model on the to-be-estimated discharge data collected at the first time to the end of the second time sequence to obtain a third time sequence;

[0096] S408. Perform weighted smoothing filtering on the third time sequence to obtain the state of charge estimation value corresponding to the first time.

[0097] In step S401, the first time is set as the current time t0, and the filter window size is n, that is, a time window capable of accommodating n time points is established.

[0098] In step S402, n-1 time points are pushed back from the first time, that is, the current time t0, to determine t0-1, t0-2, …, t0-n+2, t0-n+1, and arrange these time points in the order of time from early to late, actually forming a time window including n time points t0-n+1, t0-n+2, …, t0-2, t0-1, t0, wherein t0-n+1, t0-n+2, …, t0-2, t0-1 before the first time t0 are respectively referred to as second times. At the first time and each second time, the to-be-estimated discharge data at that time is collected from the battery to be estimated.

[0099] In step S402, n-1 time points are pushed back from the first time, that is, the current time t0, to determine t0-1, t0-2, …, t0-n+2, t0-n+1, and arrange these time points in the order of time from early to late, actually forming a time window including n time points t0-n+1, t0-n+2, …, t0-2, t0-1, t0, wherein t0-n+1, t0-n+2, …, t0-2, t0-1 before the first time t0 are respectively referred to as second times. At the first time and each second time, the to-be-estimated discharge data at that time is collected from the battery to be estimated.

[0100] In step S403, the data to be filtered corresponding to each second time is obtained. Specifically, taking the second time t0-1 before the first time t0 as an example, before the current time is moved to the first time t0, the second time t0-1 is actually the last “current time”, and the second time t0-1 is actually the “first time” in history. If S401-S408 are performed with the second time t0-1 as the “first time” in history, the weighted smoothing filtering is actually performed with the second time t0-1 as the “first time”, and the state of charge estimation value corresponding to the second time t0-1 is obtained, which is used as the data to be filtered corresponding to the second time t0-1 If S401-S408 are not performed with the second time t0-1 as the “first time” in history, that is, the weighted smoothing filtering is not performed with the second time t0-1 as the “first time”, the discharge data to be estimated collected at the second time t0-1 (and the time before the second time t0-1) can be obtained (wherein V represents the discharge voltage, I represents the discharge current, T represents the working temperature, and the subscript represents the sampling time), the discharge data to be estimated collected at the second time t0-1 The input is processed by the trained time sequence hybrid neural network, and the processing result of the time sequence hybrid neural network is used as the data to be filtered corresponding to the second time t0-1

[0101] Based on the above principle, the data to be filtered corresponding to the second time t0-n+1 can be obtained The data to be filtered corresponding to the second time t0-n+2 The data to be filtered corresponding to the second time t0-2 In step S404, the data to be filtered corresponding to each second time is arranged in chronological order of the second time, thereby obtaining the first time sequence

[0102]

[0103] In step S405, for each data to be filtered in the first time sequence , a conversion value is obtained by converting the data to be filtered according to the physical constraint suffered by the battery to be estimated between the first time and the second time corresponding to the data to be filtered.

[0104] For example, for the data to be filtered corresponding to the second time t0-n+1 , the meaning is equivalent to the estimation result of the state of charge of the battery to be estimated at the second time t0-n+1, and therefore the data to be filtered corresponding to the second time t0-n+1 can be converted according to the physical constraint suffered by the battery to be estimated between the second time t0-n+1 and the first time t0 to a corresponding conversion value.

[0105] In this embodiment, when step S405 is performed, the physical constraint considered for the battery to be estimated can be the discharge effect of the battery to be estimated, i.e., assuming that the battery to be estimated continuously discharges from the second time t0-n+1 to the first time t0, the influence of the physical effect on the battery to be estimated in this process.

[0106] Specifically, in this embodiment, when step S405 is performed, the physical constraint considered is the ampere-hour integral. For example, for the to-be-filtered data corresponding to the second time t0-n+1 When the ampere-hour integral method is applied, the initial value is , the lower limit of integration is t0-n+1, the upper limit of integration is the first time t0, and in the case where the battery parameter of the battery to be estimated is the capacity C, the following can be calculated

[0107]

[0108] where i(t) is a function of the discharge current of the battery to be estimated with respect to time t. is the to-be-filtered data corresponding to the second time t0-n+1 The corresponding conversion value is The meaning of

[0109] Based on the same principle, the to-be-filtered data corresponding to the second time t0-n+2 can be calculated according to the formula

[0110]

[0111] The corresponding conversion value is The corresponding conversion value is The conversion value corresponding to the second time t0-n+3 is continuously calculated The conversion value corresponding to the second time t0-2 is The conversion value corresponding to the second time t0-1 is and so on. In step S406, these conversion values are arranged into a second time sequence

[0112]

[0113] In step S407, the processing result of the time series hybrid neural network model on the to-be-estimated discharge data collected at the first time is obtained is obtained is added to the end of the second time sequence, so as to obtain a third time sequence

[0114]

[0115] The length of the third time sequence is consistent with the filter window size, both being n. The first n-1 values in the third time sequence have the meaning of respectively being the pre-estimation of the state of charge of the battery at the first time t0 made at the second time t0-n+1, t0-n+2, …, t0-2, t0-1, etc. The last value in the third time sequence is the actual estimation result of the time series hybrid neural network model.

[0116] In step S408, the third time sequence is subjected to weighted smoothing filtering to obtain the state of charge estimation value corresponding to the first time.

[0117] In the embodiment, Gaussian weighted smoothing filtering can be performed in step S408. The Gaussian weighted smoothing filtering is to use a Gaussian weight function in the weighted average moving filtering to give a higher weight to the data points closer to the current time point and a lower weight to the data points farther from the current time point. The purpose of this is to emphasize the recent data points more and weaken the influence of the distant data points on the average value, thereby achieving the smoothing effect.

[0118] Before the Gaussian weighted smoothing filtering is performed, a weight set is calculated according to the Gaussian weight function. The formula of the Gaussian weight function is: where w(i) is the weight of the i-th point, c is the current point, and σ is the standard deviation of the Gaussian function. In practice, the width of the weight function can be controlled by adjusting σ to balance the smoothing effect and the response speed to the data points. Specifically, in step S408, c=t0, and i=t0-n+1, t0-n+2, …, t0-2, t0 are taken in turn, so as to calculate the weight set where represents the weight of in the third time sequence. represents the weight of . represents the weight of . represents the weight of .

[0119] In step S408, each weight in the weight set is multiplied by the corresponding value in the third time sequence, and the sum is divided by the sum of the weights, so as to obtain the SOC value of the first time, i.e. the current time t0, after being subjected to the filtering constraint i.e.

[0120] the calculated is the state of charge estimation value of the battery to be estimated corresponding to the first time t0. Since the first time t0 can be arbitrarily selected or changed over time, the state of charge estimation value of the battery to be estimated at any time or any current time can be calculated.

[0121] In the embodiment, by performing the physically constrained filtering processing on the processing result of the time sequence hybrid neural network model, the state of charge estimation value of the battery to be estimated within a period of time can be smoothed and filtered according to the physical constraint on the battery to be estimated, so as to reduce the up and down fluctuations of the plurality of state of charge estimation values detected over time, and make the detected state of charge estimation value closer to the actual SOC curve of the battery to be estimated.

[0122] In the embodiment, by using the hybrid time sequence neural network model and the physically constrained filtering processing, the estimation ability and accuracy of the state of charge estimation value of the battery to be estimated are improved in combination with the two algorithms. The hybrid time sequence neural network model is composed of a time sequence convolutional neural network TCN, a long short-term memory network LSTM, and a fully connected layer to estimate SOC, which can improve the accuracy of estimating SOC. In addition, the physically constrained filtering (specifically, Gaussian weighted filtering) is additionally applied to the output of the hybrid time sequence neural network, which can constrain and smooth the output result of the hybrid time sequence neural network, and further improve the estimation accuracy of the model.

[0123] The battery state of charge estimation method in the embodiment can be implemented by writing a computer program for executing the battery state of charge estimation method in the embodiment, and writing the computer program into a computer device or a storage medium. When the computer program is read out and run, the battery state of charge estimation method in the embodiment is executed, so as to achieve the same technical effects as the battery state of charge estimation method in the embodiment.

[0124] It should be noted that, unless otherwise specified, when a certain feature is referred to as being "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. In addition, the up, down, left, right and other descriptions used in the disclosure are only relative to the relative positional relationship of the components of the disclosure in the drawings. The singular forms "a", "an" and "the" used in the disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. In addition, unless otherwise defined, all technical and scientific terms used in the embodiment have the same meaning as generally understood by those skilled in the art. The terms used in the embodiment are only used to describe the specific embodiments, and are not intended to limit the present application. The term "and / or" used in the embodiment includes any combination of one or more related listed items.

[0125] It should be understood that, although the terms first, second, third, etc. can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could also be termed a second element, and, similarly, a second element could also be termed a first element without departing from the scope of the present disclosure. The use of any and all examples, or exemplary language (e.g., "such as", "for instance", etc.) provided herein is intended merely to better illuminate embodiments of the present disclosure and does not pose a limitation on the scope of the present disclosure unless otherwise claimed.

[0126] It will be recognized that embodiments of the present application can be realized by a computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The methods can be implemented in a computer program product tangibly embodied in a non-transitory computer-readable storage medium having computer instructions that are configured to cause a computer to operate in a specific and predefined manner described in the specific embodiments and drawings. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with a computer system. However, the program can be implemented in assembly or machine language if desired. In any case, the language can be a compiled or interpreted language. Furthermore, the program can be capable of running on a specially programmed integrated circuit for this purpose.

[0127] Furthermore, the operations of the processes described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The processes described herein (or variations and / or combinations thereof) can be implemented under the control of one or more computer systems configured with executable instructions (e.g., computer programs, one or more computer programs, or one or more applications) to perform the operations of the processes, by hardware, or combinations thereof. The computer programs include a plurality of instructions executable by one or more processors.

[0128] Further, the methods can be implemented in any type of computing platform operably connected to the appropriate, including but not limited to a personal computer, mini-computer, mainframe, workstation, network or distributed computing environment, separate or integrated computer platforms, or in communication with charged particle tools or other imaging devices, and the like. Aspects of the present application can be implemented in machine readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage media, RAM, ROM, and the like, such that it can be read by a programmable computer to configure and operate the computer to perform the processes described herein when the storage medium or device is read by the computer. In addition, the machine readable code, or portions thereof, can be transmitted over wired or wireless networks. The present embodiments of the application include these and other different types of non-transitory computer readable storage media when such media include instructions or programs that implement the above steps in conjunction with a microprocessor or other data processor. The present application also includes the computer itself when programmed in accordance with the methods and techniques of the present application.

[0129] The computer program can be applied to input data to perform the functions of the present embodiments, thereby transforming the input data to generate output data that is stored to non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In preferred embodiments of the present application, the transformed data represents a physical and tangible object, including a particular visual depiction of the physical and tangible object produced on a display.

[0130] The above merely preferred embodiments of the present application and are not intended to limit the present application thereto. The present application is not limited to the above-described embodiments, but any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the present application. The technical scheme and / or embodiments of the present application can have various modifications and changes within the scope of the present application.

Claims

1. A method for estimating the state of charge of a battery, characterized in that, The battery state-of-charge estimation method includes: A temporal hybrid neural network model is established; the temporal hybrid neural network model is used to capture temporal features of input data, learn time dependencies of the captured temporal features, and perform regression prediction on the learned time dependencies; Obtain at least one discharge data point of the battery to be estimated; The time-series hybrid neural network model is used to process each of the discharge data to be estimated; The processing results of the temporal hybrid neural network model are subjected to physically constrained filtering to obtain an estimate of the state of charge of the battery to be estimated. The step of performing physically constrained filtering on the processing results of the temporal hybrid neural network model to obtain an estimate of the state of charge of the battery to be estimated includes: Determine the filter window size and the first time step; Within a time period equivalent to the size of the filter window before the first time point, multiple second time points are determined; the first time point and each of the second time points are respectively the acquisition time of one of the discharge data to be estimated; Obtain the data to be filtered corresponding to each of the second time points; Arrange the data to be filtered into a first time series according to the time order; For any of the data to be filtered, a transformation value is obtained by converting the data based on the physical constraints experienced by the battery to be estimated between the first time and the second time corresponding to the data to be filtered. Arrange the transformed values ​​into a second time series according to the order of the first time series; The processing result of the time-series hybrid neural network model on the discharge data to be estimated collected at the first time moment is added to the end of the second time series to obtain the third time series; The third time series is subjected to weighted smoothing filtering to obtain the estimated value of the state of charge corresponding to the first time moment.

2. The battery state-of-charge estimation method according to claim 1, characterized in that, The establishment of the temporal hybrid neural network model includes: A temporal convolutional neural network, a long short-term memory network, and a fully connected layer are established. The temporal convolutional neural network is used to capture temporal features of the input data, the long short-term memory network is used to learn the time dependencies of the temporal features captured by the temporal convolutional neural network, and the fully connected layer is used to perform regression prediction on the time dependencies learned by the long short-term memory network. The temporal hybrid neural network model is obtained by sequentially connecting the temporal convolutional neural network, the long short-term memory network, and the fully connected layer.

3. The battery state-of-charge estimation method according to claim 1, characterized in that, The establishment of the temporal hybrid neural network model includes: Obtain sample discharge data and sample state of charge data of the sample battery at the same time point; The temporal hybrid neural network model is trained using the sample discharge data and the sample state of charge data.

4. The battery state-of-charge estimation method according to claim 3, characterized in that, The step of training the temporal hybrid neural network model using the sample discharge data and the sample state of charge data includes: Establish multiple hyperparameter sets; each hyperparameter set includes multiple hyperparameters. Execute multiple iterations until the iteration termination condition is met; Return the set of hyperparameters corresponding to the model with the best performance evaluation in the last round of the iteration; The temporal hybrid neural network model is configured using the returned set of hyperparameters to obtain the trained temporal hybrid neural network model; The iterative process in any round includes the following steps: The temporal hybrid neural network model is configured using each of the aforementioned hyperparameter sets to obtain multiple models to be evaluated; wherein, any one of the models to be evaluated is obtained by configuring the temporal hybrid neural network model using a corresponding hyperparameter set. The sample discharge data and the sample state of charge data are used to train each of the models to be evaluated. The performance of each of the trained models to be evaluated is evaluated separately. For at least one of the models to be evaluated with the worst performance evaluation result, the hyperparameter set corresponding to each of the models to be evaluated is updated.

5. The battery state-of-charge estimation method according to claim 1, characterized in that, The step of obtaining the data to be filtered corresponding to each of the second time moments includes: For any of the second time moments: When there is a corresponding state of charge estimate at the second time, the state of charge estimate at the second time is used as the data to be filtered at the second time. If there is no corresponding state of charge estimate at the second time, the processing result of the discharge data to be estimated collected at the second time by the time-series hybrid neural network model shall be used as the data to be filtered at the second time.

6. The battery state-of-charge estimation method according to claim 1, characterized in that, The step of transforming the battery to be estimated between the first time point and the second time point corresponding to the data to be filtered, to obtain a corresponding transformed value, includes: Obtain the battery parameters of the battery to be estimated; Using the data to be filtered as the initial value, the second time point as the lower limit of integration, and the first time point as the upper limit of integration, the conversion value is obtained by performing ampere-hour integration based on the battery parameters.

7. The battery state-of-charge estimation method according to claim 1, characterized in that, The step of performing weighted smoothing filtering on the third time series to obtain the estimated state of charge corresponding to the first time moment includes: The Gaussian weighting function is calculated based on the third time series to obtain the weight set; The third time series is weighted and averaged according to the weight set to obtain the estimated state of charge corresponding to the first time moment.

8. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store at least one program, and the processor being used to load at least one program to execute the battery state-of-charge estimation method according to any one of claims 1-7.

9. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the battery state-of-charge estimation method according to any one of claims 1-7.

Images