CN117031318A

CN117031318A - Battery performance prediction method, device, computer equipment and storage medium

Info

Publication number: CN117031318A
Application number: CN202211123613.5A
Authority: CN
Inventors: 沈雷; 王艺霖; 张睿欣; 张昊; 赵凯; 丁守鸿
Original assignee: Tencent Cloud Computing Beijing Co Ltd
Current assignee: Tencent Cloud Computing Beijing Co Ltd
Priority date: 2022-09-15
Filing date: 2022-09-15
Publication date: 2023-11-10

Abstract

The present application relates to a battery performance prediction method, apparatus, computer device, storage medium and computer program product. The method comprises the following steps: acquiring a multidimensional working condition information sequence of a battery to be tested in a past charge and discharge cycle; respectively inputting multidimensional working condition information sequences of charge and discharge cycles in the past into a pre-trained encoder, and outputting battery working condition characteristics of the charge and discharge cycles in the past through the encoder; and (3) merging the working condition characteristics of the battery in the past charge-discharge cycle, inputting the battery into a trained decoder, and outputting the predicted battery performance of the battery to be tested through the decoder. The method improves the accuracy of battery performance prediction.

Description

Battery performance prediction method, device, computer equipment and storage medium

Technical Field

The present application relates to the field of machine learning and battery technology, and in particular, to a battery performance prediction method, apparatus, computer device, storage medium, and computer program product.

Background

With the wide application of batteries in new energy vehicles/energy storage scenarios, accurate prediction of battery performance becomes increasingly important. The battery performance has close relation with the charge and discharge speed, the use environment and the like in the use process of the battery.

Taking the battery capacity prediction as an example, the battery capacity prediction technique may employ a data-driven method for prediction, which uses a statistical/machine learning/deep learning model to build a battery capacity model, predicting the remaining battery capacity. The method can integrate external influence factors, but depends on large-scale data, has higher requirement on data distribution, and can greatly influence the prediction result when input data is biased.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a battery performance prediction method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve the accuracy of battery performance prediction.

In a first aspect, the present application provides a battery performance prediction method. The method comprises the following steps:

acquiring a multidimensional working condition information sequence of a battery to be tested in a past charge and discharge cycle;

respectively inputting the multidimensional working condition information sequences of the charge and discharge cycles into a pre-trained encoder, and outputting battery working condition characteristics of the charge and discharge cycles through the encoder;

and after the battery working condition characteristics of the charge and discharge cycles are fused, inputting the battery working condition characteristics into a trained decoder, and outputting predicted battery performance of the battery to be tested through the decoder.

In a second aspect, the present application also provides a device for predicting battery performance. The device comprises:

the data acquisition module is used for acquiring a multidimensional working condition information sequence of the battery to be tested in the past charge and discharge cycle;

the coding module is used for respectively inputting the multidimensional working condition information sequences of the charge and discharge cycles into a pre-trained coder, and outputting the battery working condition characteristics of the charge and discharge cycles through the coder;

and the decoding module is used for inputting the battery working condition characteristics of the charge and discharge cycles into a trained decoder after being fused, and outputting the predicted battery performance of the battery to be tested through the decoder.

In a third aspect, the present application also provides a computer device. The computer device comprises a memory storing a computer program and a processor which when executing the computer program performs the steps of:

In a fourth aspect, the present application also provides a computer-readable storage medium. The computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

In a fifth aspect, the present application also provides a computer program product. The computer program product comprises a computer program which, when executed by a processor, implements the steps of:

According to the battery performance prediction method, the device, the computer equipment, the storage medium and the computer program product, the multi-dimensional working condition information sequence in the past charging cycle is input to the pre-trained encoder to obtain the battery working condition characteristics of the past charging and discharging cycle, and then the battery performance is predicted through the encoder, on one hand, the pre-trained encoder is used for carrying out characteristic characterization on the multi-dimensional working condition information of the past charging and discharging cycle, so that the influence on model precision when an abnormal signal or a biased signal is input is reduced, on the other hand, the battery performance is obtained based on the multi-dimensional working condition information sequence in the past charging cycle, the multi-dimensional working condition information is fused, the influence on the battery caused by the existing use condition can be fully considered, and the battery performance prediction precision is improved.

Drawings

FIG. 1 is a diagram of an application environment for a battery performance prediction method in one embodiment;

FIG. 2 is a flow chart of a method of battery performance prediction in one embodiment;

FIG. 3 is a flow diagram of the steps for training the encoder and decoder in one embodiment;

FIG. 4 is a flowchart illustrating steps for constructing positive and negative sample pairs based on multi-dimensional operating condition information of a plurality of batteries over a past charge-discharge cycle in one embodiment;

FIG. 5 is a flowchart illustrating the steps for constructing positive and negative pairs of samples based on multi-dimensional operating condition information of a plurality of batteries during a previous charge-discharge cycle in another embodiment;

FIG. 6 is a schematic diagram of an encoder according to one embodiment;

FIG. 7 is a block diagram showing the structure of a battery performance prediction apparatus in one embodiment;

fig. 8 is an internal structural diagram of a computer device in one embodiment.

Detailed Description

The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.

The battery performance prediction method provided by the embodiment of the application can be applied to an application environment shown in fig. 1. Wherein the device 102 communicates with the server 104 via a network. The data storage system may store data that the server 104 needs to process. The data storage system may be integrated on the server 104, may be placed on a cloud or other server, may store a multi-dimensional sequence of operating conditions in a last charge-discharge cycle, may store an encoder model, a decoder model, etc. The terminal reports the multidimensional working condition information sequence in the previous charging and discharging cycle to the server, and the server acquires the multidimensional working condition information sequence of the battery to be tested in the previous charging and discharging cycle; respectively inputting multidimensional working condition information sequences of charge and discharge cycles in the past into a pre-trained encoder, and outputting battery working condition characteristics of the charge and discharge cycles in the past through the encoder; and (3) merging the working condition characteristics of the battery in the past charge-discharge cycle, inputting the battery into a trained decoder, and outputting the predicted battery performance of the battery to be tested through the decoder. The device 102 may be a device powered by a battery, including but not limited to a new energy automobile, a mobile phone, a computer, an intelligent voice interaction device, an intelligent home appliance, a vehicle-mounted terminal, an aircraft, etc. The server 104 may be implemented as a stand-alone server or as a server cluster of multiple servers.

The device 102 can also be used for realizing a battery performance prediction method, and the device can be used for obtaining a multidimensional working condition information sequence of the battery to be tested in a past charge and discharge cycle; respectively inputting multidimensional working condition information sequences of charge and discharge cycles in the past into a pre-trained encoder, and outputting battery working condition characteristics of the charge and discharge cycles in the past through the encoder; and (3) merging the working condition characteristics of the battery in the past charge-discharge cycle, inputting the battery into a trained decoder, and outputting the predicted battery performance of the battery to be tested through the decoder.

In one embodiment, as shown in fig. 2, a battery performance prediction method is provided, wherein the battery performance may be at least one of battery life, battery capacity, battery remaining charge time, battery discharge characteristics and internal resistance, and battery storage performance. Taking the server in fig. 1 as an example, the method comprises the following steps:

step 202, obtaining a multidimensional working condition information sequence of the battery to be tested in a past charge and discharge cycle.

A battery is a device that converts chemical energy into electrical energy. In this embodiment, the battery performance is predicted based on the multidimensional operating condition information of the battery in the historical charge-discharge cycle, and therefore, the battery in this embodiment is a rechargeable battery. Among them, common rechargeable batteries can be classified into different kinds according to electrode materials and electrolyte properties. For example, zinc series batteries, such as zinc-manganese batteries, zinc-silver batteries, nickel series batteries, lead series batteries, lithium-magnesium batteries, and the like, can be included in the division of materials for the positive and negative electrodes used in the batteries. Alkaline cells may be included, depending on the type of electrolyte, cells with the electrolyte being predominantly aqueous potassium hydroxide.

The charge-discharge cycle of the battery refers to the process by which the battery completes one 100% complete discharge/charge. For example, a battery is 100% of available batteries, and is charged to 100% by using 0% of light at a time, and is charged and discharged once. At this time, the number of charge and discharge cycles of the battery was 1, and the number of charges was 1. For another example, a battery has 100% charge, 20% for the first time, then 100% for charging, 80% for the second time, and 100% for recharging, and the two uses together are a complete charge-discharge cycle. At this time, the number of battery cycles was 1, and the number of charging was 2. Accordingly, the charge-discharge cycle information may include at least one of the number of charge-discharge cycles and the number of charge times.

The multi-dimensional working condition information sequence is specifically working condition information of multiple dimensions of each time point of one charge-discharge cycle, wherein the multiple dimensions can be at least two of charge-discharge voltage, current, temperature, charge-discharge duration and load of the battery. It will be appreciated that the operating mode information required for different battery performance varies in dimension. Taking the battery capacity as an example, the multiple dimensions include battery charge-discharge voltage, current, temperature, charge-discharge duration, and load.

Step 204, respectively inputting the multidimensional working condition information sequences of the charge and discharge cycles in the past into a pre-trained encoder, and outputting the working condition characteristics of the battery of the charge and discharge cycles in the past through the encoder.

The encoder is used for extracting the characteristic expression of the multi-dimensional working condition information. The encoder is obtained by performing supervised learning training according to a multi-dimensional information sequence of the battery in the marked battery history charge-discharge cycle by utilizing a machine learning model in advance. Among them, the machine learning model of the encoder may employ Transformer, BR, RNN or the like.

And inputting the past multidimensional working condition information sequence into a pre-trained encoder to obtain the working condition characteristics of the battery of each charge and discharge cycle.

The traditional battery performance prediction method comprises three steps:

a prediction method based on a physical electrochemical model, which establishes a mathematical formula or an empirical model to describe the aging behavior of a battery according to the electrochemical reaction mechanism inside the battery, is generally composed of a series of algebraic and differential equations. The model-based method is characterized in that the established battery capacity prediction model is only specific to a specific system. In a practical application scenario, due to the change of environment and use conditions, the battery is in dynamic change, and it is difficult to describe the battery use condition by using a single system.

The method is based on a statistical theory and a machine learning theory, and directly utilizes historical data to build a prediction model without depending on a specific physical model. But depends on large-scale data, has higher requirement on data distribution, and can have larger influence on the prediction result when input data is biased.

Based on a fusion algorithm, the model is combined with a data driving method and the advantages of the model and the data driving method are generated, the method can improve the robustness of prediction, but the calculation complexity is improved, and the calculation flow is complex.

Similar to the battery performance prediction method of the embodiment, the prediction method based on data driving aims at the defects of the method, and in the embodiment, the multidimensional working condition information sequence in the charge and discharge cycle is subjected to characteristic expression through the encoder, so that the influence on the model prediction result when input data is biased is reduced.

And 206, inputting the battery working condition characteristics of the charge and discharge cycles into a trained decoder after fusing, and outputting predicted battery performance of the battery to be tested through the decoder.

The decoder takes the battery working condition characteristics of the battery in the past charge and discharge cycles as input and is used for predicting the battery performance. After the encoder is utilized to obtain the characteristic expression of the multidimensional working condition information sequence in the past charge and discharge cycle, all the historical working condition characterization information is fused through a downstream decoder, so that the influence of the historical use condition on the battery performance is more comprehensively considered, and the high-precision prediction on the battery performance can be realized.

According to the battery performance prediction method, the multi-dimensional working condition information sequence in the past charging cycle is input to the pre-trained encoder to obtain the battery working condition characteristics of the past charging and discharging cycle, and then the battery working condition characteristics of the past charging and discharging cycle are predicted through the encoder.

In another embodiment, the battery performance of the battery to be tested is predicted by the decoder output after the battery working condition characteristics of the charge and discharge cycles are fused and input into the trained decoder, and the method comprises the following steps: the battery working condition characteristics of charge and discharge cycles in the past are spliced to obtain battery characteristics; the battery characteristics are input to a trained decoder, through which the predicted battery performance is output.

The battery working condition characteristics of the charge and discharge cycles of the past times are battery working condition characteristics of the charge and discharge cycles of all the past times of the battery to be predicted. And respectively inputting the multidimensional working condition information sequences of each historical charge and discharge cycle into a pre-trained encoder to obtain the battery working condition characteristics of each charge and discharge cycle. And then splice the battery working condition characteristic of charge-discharge cycle of past time, as the input of the demoder, transversely splice the battery working condition characteristic of charge-discharge cycle t of input past time, get the battery characteristic, as the input of the demoder model:

Feature＝Concat(feature ₁ ,...,feature _t )

The battery characteristics are processed by a decoder to obtain predicted battery performance.

The machine model structure of the decoder can adopt BP network, LSTM or RNN. Taking Long Short-Term Memory (LSTM) as an example, mapping of battery characteristics to battery performance is accomplished using Long Short-Term Memory networks. A Long Short-Term Memory network (LSTM) is an RNN network that addresses time series and sequence problems, with many achievements in speech recognition and machine translation. The LSTM can control the information flow by introducing the functions of an input gate, a forgetting gate, a unit candidate gate, an output gate and the like, so that forgetting of information at a position with a longer time step due to disappearance of a long sequence gradient is reduced, and efficient grabbing and memorizing of the long sequence information are realized. The layer structure of the LSTM exists in the form of a chain repetition module, called LSTM cell.

The prediction of the decoder is specifically to forward propagate battery characteristics in the decoder, taking the decoder as an LSTM as an example, the LSTM can effectively alleviate the gradient disappearance problem caused by too far apart time steps, thereby realizing the extraction and storage of long-time sequence information. LSTM has four components that interact in a special way, which are forget gate f, input gate i, cell candidates A gate and an output gate o. They control the level of cell state updates, the level of control cell state forgets, adding information to cell state, the level of control cell state to hidden state, respectively.

Details of the forward propagation of neurons at time t, the LSTM shown is a single layer LSTM sequence-to-sequence regression network. Neuron receives h _t-1 And c _t-1 At time t-1, the resulting x _t-1 Input, x _t-1 And h _t c _t Will be calculated in the next timing and output by forward propagationThe values of the four components can be obtained from the equation.

f _t ＝σ(W _f h _t-1 +U _f x _t +b _f )＝σ(zf _t )

i _t ＝σ(W _i h _t-1 +U _i x _t +b _i )＝σ(zi _t )

o _t ＝σ(W _o h _t-1 +U _o x _t +b _o )＝σ(zf _o )

Where W represents a recursive weight from neuron to neuron, U represents an input to neuron input weight, and b represents a bias. Let the number of neuronal nodes be m. Thentan h (·) and sigmoid represent activation functions and are respectively as follows:

wherein c _t ，h _t And (3) withCan be obtained by the following formulas:

h _t ＝o _t ⊙tanh(c _t )

in this embodiment, the battery characteristics are obtained by splicing the battery working condition characteristics of the charge and discharge cycles of the past, and the battery characteristics are predicted by the encoder, so that the prediction of the battery performance is obtained by using the multi-dimensional working condition information sequence in the charge cycle of the past, the multi-dimensional working condition information is fused, the influence of the existing use condition on the battery can be fully considered, and the accuracy of the battery performance prediction is improved.

In another embodiment, the way the encoder and decoder are trained, as shown in fig. 3, comprises:

and step 302, training the encoder based on the constraint of the charge-discharge cycle loss prior term by using the positive sample pair and the negative sample pair to obtain a trained encoder.

The training of the encoder can be based on unsupervised learning training, such as training the encoder by using a contrast representation learning method. The method of unsupervised learning directly uses the data itself as the supervision information without manually labeled category label information, and learns the feature expression of the sample data, and is applied to downstream tasks (the downstream tasks in this embodiment are the decoder predicting the battery performance). The mode omits the manual labeling process, and the training efficiency is effectively improved.

The training of the encoder can also be obtained based on supervised learning training. This approach requires manual pre-construction of positive and negative pairs of samples.

The goal of training the encoder with positive and negative pairs of samples is to increase the similarity between positive pairs of samples and to increase the similarity between negative pairs of samples.

In order to further mine the degradation information in the charge-discharge cycle, the greater the difference of the characteristic characterization in the cycle is assumed to be, so that the charge-discharge cycle loss prior term is introduced in the loss in the pre-training stage. The constraint of the cycle loss prior term is utilized to train the encoder, so that the influence of different charge and discharge cycles on the battery performance can be considered, and the battery performance prediction precision is improved

Step 304, respectively inputting the multi-dimensional working condition information sequences of the marked battery in the charge and discharge cycles to the trained encoders to obtain the battery working condition characteristics of the charge and discharge cycles.

In this embodiment, the encoder and decoder are obtained through two-stage training, and the decoder training uses the training results of the encoder. After the encoder is trained, the encoder is used for carrying out characteristic re-characterization on the working condition information, so that the influence on model precision when abnormal signals or biased signals are input is reduced.

And 306, taking the battery working condition characteristics of the charge and discharge cycles as the input of the decoder, and adjusting the parameters of the decoder according to the predicted battery performance and the labeling information output by the decoder.

The training of the decoder is supervised learning. And acquiring a pre-trained labeling sample set, wherein the sample set comprises multi-dimensional working condition information of the battery in charge and discharge cycles and labeled battery performance. And (3) inputting multidimensional working condition information of the battery in the sample set in the past charge and discharge cycles to a trained encoder, and outputting battery working condition characteristics of the past charge and discharge cycles through the encoder. And taking the battery working condition characteristics of the charge and discharge cycles as the input of a decoder, and outputting predicted battery performance of the battery through the decoder. And adjusting parameters of the decoder according to the predicted battery performance and the difference of the labeling information to complete one iteration training.

Step 308, determining whether the iteration end condition is satisfied.

If yes, step 310 is executed, if not, step 306 is executed iteratively until the iteration end condition is satisfied. The iteration end condition may be that the iteration number reaches a preset iteration number, or that the decoder model is stable (for example, the relative change value of the loss function is smaller than a specified value).

Step 310, a trained decoder is obtained.

The machine model structure of the decoder can adopt BP network, LSTM or RNN.

In this embodiment, the training processes of the encoder and the decoder are decoupled, the encoder and the decoder are trained respectively through two-stage training, the encoder is used for carrying out characteristic re-characterization on the working condition information, the influence on the model precision when abnormal signals or biased signals are input is reduced, further, the trained encoder is used for obtaining the characteristic representation of the multi-dimensional working condition information, all the historical working condition characterization information is fused, the influence of the historical use condition on the battery capacity is more comprehensively considered, and the high-precision prediction on the battery capacity can be realized.

Considering that the above-described way of training the encoder and decoder requires two stages of training, which is relatively time consuming, in another embodiment, the way of training the encoder and decoder may also employ end-to-end training, thereby speeding up the model training.

For example, inputting a multi-dimensional working condition information sequence of the historical charge-discharge cycle of the marked battery into a trained encoder to obtain battery working condition characteristics of the historical charge-discharge cycle; taking the battery working condition characteristics of the battery in the past charge and discharge cycles as the input of a decoder, and outputting predicted battery performance through the decoder; and (3) adjusting parameters of the encoder and the decoder according to the predicted battery performance and the marking information, and iteratively returning to the step of inputting the multi-dimensional working condition information sequence of the marked battery in the history charge-discharge cycle to the trained encoder until the training ending condition is met, so as to obtain the trained encoder and decoder.

The embodiment adopts a mode of joint training of the encoder and the decoder, thereby improving the training speed of the encoder and the decoder.

In another embodiment, the manner in which the encoder is trained includes: constructing a positive sample pair and a negative sample pair based on multidimensional working condition information of a plurality of batteries in a past charge and discharge cycle; and comparing, expressing and learning the encoder based on the constraint of the charge-discharge cycle loss prior term by using the positive sample pair and the negative sample pair to obtain the trained encoder.

The comparison and representation learning (Contrastive Learning) is an unsupervised learning method, does not need manually labeled category label information, directly uses data itself as supervision information, learns the characteristic expression of sample data, and is applied to downstream tasks (the downstream tasks in the embodiment are the decoder for predicting the battery performance).

Because the comparative representation represents category label information that does not require manual labeling, it becomes particularly important to construct the positive and negative samples required for training. The comparison indicates that there are many ways to learn to construct positive and negative sample pairs, one way is data enhancement, specifically, for one anchor sample (anchor), positive sample pairs can be constructed by data enhancement, then the other samples in one training batch can be taken as negative sample pairs. The object of the contrast representation learning is to reduce the distance between the anchor sample and the positive sample and to enlarge the distance between the anchor sample and the negative sample, so that the encoder for feature extraction is trained.

After the encoder is obtained by using the comparison representation learning training, in order to improve the feature extraction capability of the encoder, a small amount of labeling samples can be used for supervised learning, the encoder obtained by using the comparison representation learning is adjusted, and the feature extraction capability of the encoder is improved.

Considering that the actual battery performance has close relation with charge and discharge cycles, in order to further mine the cycle inner decline information, the more the cycle number is increased, the more the health state of the battery is gradually declined, and the more the characteristic characterization difference in the cycle is; thereby introducing a charge-discharge cycle loss prior term in the loss of the pre-training stage.

In the embodiment, the encoder is trained by using the comparison representation learning, and training is performed based on the constraint of the charge-discharge cycle loss prior term in the comparison representation learning process, so that the influence of different charge-discharge cycles on the battery performance can be considered, and the battery performance prediction precision is improved.

Specifically, as shown in fig. 4, the positive and negative sample pairs are constructed based on the multidimensional operating condition information of the plurality of batteries in the past charge and discharge cycles, including:

step 402, a multi-dimensional working condition information sequence of a plurality of batteries in a past charge and discharge cycle is obtained.

Specifically, a sample set is obtained by collecting multidimensional working condition information sequences of a plurality of batteries in a past charge-discharge cycle. The sample set comprises multidimensional working condition information sequences of batteries with different numbers in the past charge and discharge cycles. For example, the multi-dimensional working condition information sequence of the battery A in the 1 st to 100 th charge and discharge cycles and the multi-dimensional working condition information sequence of the battery B in the 1 st to 200 th charge and discharge cycles are included.

And step 404, carrying out random masking processing on the multi-dimensional working condition information sequence in the previous charge and discharge cycle in a time sequence dimension to obtain a masked multi-dimensional working condition information sequence.

Specifically, the multi-dimensional working condition information sequence in the previous charge and discharge cycle is subjected to random masking in a time sequence dimension, the masking object can be a mapping feature projected by a layer of transducer, and the masking mode is an independent two-dimensional mask obeying Bernoulli distribution. The purpose of the mask is to do information perturbation such that subsequent positive sample pairs fit to maximize mutual information consistency.

Step 406, selecting the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window, wherein the window sizes of the first window and the second window are different.

Specifically, a first window is adopted to obtain a group of masked multi-dimensional working condition information sequences in a sample set, and a second window is adopted to obtain a group of masked multi-dimensional working condition information sequences in the sample set, wherein the sizes of the first window and the second window are different. For example, the first window is [ a ] ₁ ,b ₁ ]Any group of time windows is taken to be in [ a ] ₁ ,b ₁ ]Is a multi-dimensional working condition information sequence after masking, and is [ a ] in a second window ₂ ,b ₂ ]Any group of time windows is taken to be in [ a ] ₂ ,b ₂ ]Is a masked multi-dimensional operating mode information sequence.

Step 408, it is determined whether the first window and the second window overlap. If yes, go to step 410, if no, go to step 412.

And 410, taking the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window as positive sample pairs, and taking the multi-dimensional working condition information sequences on non-overlapping time sequences in the first window and the second window as negative sample pairs.

Wherein if the first window and the second window overlap, it is indicated that they overlap in time sequence, and in theory, the characteristic representations of the two should be identical, the first window may be defined asThe masked multi-dimensional working condition information sequence in the second window and the masked multi-dimensional working condition information sequence in the second window are used as positive sample pairs. And the non-overlapped parts are different, and the multi-dimensional working condition information sequences on the non-overlapped time sequences in the first window and the second window are used as negative sample pairs. For example, for a sequence existing in two timing steps [ a ] ₁ ,b ₁ ],[a ₂ ,b ₂ ]Wherein 0 < a ₁ ≤a ₂ ≤b ₁ ≤b ₂ Then the overlap of the two sequences is selected [ a ] ₂ ,b ₁ ]As a cutting portion, a first window [ a ] ₁ ,b ₁ ]Inside masked multi-dimensional working condition information sequence and second window [ a ] ₂ ,b ₂ ]The masked multi-dimensional working condition information sequence in the inner is taken as a positive sample pair, and a non-overlapped part [ a ₁ ,a ₂ ]And [ b ] ₁ ,b ₂ ]The masked multidimensional working condition information sequence in the sample is taken as a negative sample pair.

Step 412, taking the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window as negative sample pairs.

Specifically, if the first window and the second window do not overlap, the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window may be used as a negative sample pair.

The positive sample pair and the negative sample pair are constructed by adopting the time sequence mask and cutting modes, and the time sequence relation between the positive sample pair and the negative sample pair is utilized by overlapping the time sequence, so that the encoder can be assisted to learn the characteristics irrelevant to the position better and have better robustness on the time sequence class.

In another embodiment, the positive and negative pairs of samples are constructed based on multi-dimensional operating condition information of a plurality of batteries in a past charge-discharge cycle, as shown in fig. 5, comprising:

step 502, taking a multidimensional working condition information sequence of a plurality of batteries in a past charge and discharge cycle as training data.

And 504, taking a multi-dimensional working condition information sequence of one charge-discharge cycle from training data of the current training batch as an anchor sample, performing data enhancement processing on the anchor sample to obtain a positive sample, and obtaining a positive sample pair according to the anchor sample and the positive sample.

Specifically, batch-wise (batch) trains the encoder with training data, e.g., 10 training data may be employed per training. And taking a multi-dimensional working condition information sequence of one charge-discharge cycle as an anchor sample (anchor) in training data of the current training batch, and carrying out data enhancement processing on the anchor sample to obtain a positive sample. The data enhancement processing may be performed on the anchor samples in various manners, and the essence of the data enhancement processing is to disturb the data, for example, the anchor samples may be cut, the anchor samples may be masked, and the like.

When the encoder is trained, the training target is to pull the distance between the anchor sample and the positive sample, so that the purpose of constructing the positive sample pair is to enable the encoder to learn the maximized mutual information consistency of one sample under different treatments by carrying out disturbance treatment on the anchor sample, thereby improving the characteristic characterization accuracy of the encoder.

Step 506, taking a negative sample of the multi-dimensional working condition information sequence of one charge-discharge cycle of the non-anchor sample from the training data of the current training batch, and obtaining a negative sample pair according to the anchor sample and the negative sample.

Specifically, the multidimensional working condition information sequence of one charge-discharge cycle of the anchor sample in the training data of the current training batch is regarded as data of different categories from the anchor sample, the multidimensional working condition information sequence can be used as a negative sample, and a negative sample pair is obtained according to the anchor sample and the negative sample. In performing the training of the encoder, the training goal is to increase the anchor sample to negative sample distance.

In this embodiment, positive samples are constructed by performing data enhancement processing on the multidimensional working condition information sequence, and samples of different types of anchor samples are used as negative samples, so that a positive sample pair and a negative sample pair can be constructed as well, and the encoder can learn the maximized mutual information consistency of one sample under different processing, thereby improving the feature characterization accuracy of the encoder.

In another embodiment, the comparison representation learning is performed on the encoder based on the constraint of the charge-discharge cycle loss prior term by using the positive sample pair and the negative sample pair to obtain the trained encoder, and the method comprises the following steps: inputting the positive sample pair and the negative sample pair into an encoder respectively; calculating time sequence contrast loss and sample contrast loss according to the output of the encoder, wherein the sample contrast loss takes a charge-discharge cycle loss prior term as a constraint, and the charge-discharge cycle loss prior term amplifies the sample weight of a negative sample with larger phase difference with the positive sample cycle number, and reduces the sample weight of the negative sample with smaller phase difference with the positive sample cycle number; and adjusting the encoder based on the constraint of the time sequence comparison loss and the sample comparison loss and performing iteration training until the iteration condition is met, so that the trained encoder is obtained.

In this embodiment, an encoder is described by taking a transducer as an example, and the encoder is formed by stacking a plurality of decoder modules in a transducer architecture, wherein a specific structure of a single decoder module is shown in fig. 6.

The transform decoder consists of two layers of sub-layers with regular terms and residual connections, and the formula is:

sub_Layer_Output＝LayerNorm(x+(SubLayer(x)))

wherein, the sub-layer of the first layer is composed of multi-head attention, and the expression is as follows:

MultiHead(Q,K,V)＝Concat(head ₁ ,...,head _h )W ^O

head _i ＝Attention(QW _i ^Q ,KW _i ^K ,VW _i ^V )

the attention used here takes the form of dot product with normalization, the function of which is as follows:

the second layer subnetwork is a forward propagation network calculated by bits, and the expression is as follows:

FFN(x)＝max(0,xW ₁ +b ₁ )W ₂ +b ₂

specifically, the trained loss function of the encoder includes two, a timing contrast loss and a sample contrast loss, respectively.

In order to learn the characteristics of multi-time dimension resolution, adopting structured contrast loss combining time sequence contrast loss and sample contrast loss, downsampling positive and negative examples in the time dimension by using a maximum pooling layer (max pool), respectively calculating and updating the sample contrast loss and the time sequence contrast loss of a level lower than the time sequence resolution, and finally weighting the loss functions of different time sequence resolution levels to obtain the final structured contrast loss.

To further mine the intra-cycle degradation information, it is assumed that as the number of cycles increases, the state of health of the battery gradually degrades, the greater the variability of the intra-cycle feature characterization; thereby introducing a cycle loss prior term g (delta C) in the case loss of the pre-training stage, amplifying the softmax sample weight of a large negative sample with a larger cycle number than the positive sample, and reducing the sample weight of the negative sample with a smaller cycle number than the positive sample.

The update formulas of the comparison loss of the time sequence level, the case comparison loss and the total level comparison loss are as follows:

wherein r is _i,t Is the feature vector extracted by the feature extraction network at the t time step of the r sample,is the timing hierarchy contrast penalty.

g(c _i,j ) Is a loss correction term for introducing loop information for samples i and j, g (c) _i,j ) W and b are adjustable parameters, control the distribution of correction terms,is the sample contrast loss.

Wherein the goal of the sample contrast loss is to pull the distance between positive pairs of samples closer, increasing the distance between negative pairs of samples.

In this embodiment, the encoder is trained, so that the characteristic encoder can be used to perform characteristic re-characterization on the working condition information.

In one embodiment, a battery performance prediction method will be described taking battery performance as an example of battery capacity. The implementation of the method comprises two stages:

The first stage is training of the model.

Wherein the training phase of the model may comprise the steps of:

the first step: and collecting multidimensional working condition information sequences of a plurality of batteries in the past charge and discharge cycle. For example, the sensor is used for collecting working condition information such as charge and discharge voltage, current, temperature and the like of the battery.

And a second step of: the positive pair of samples is constructed by randomly sampling two overlapping sub-sequences, and the negative pair of samples is constructed by randomly sampling two non-overlapping sub-sequences. Specifically, a multidimensional working condition information sequence of a plurality of batteries in a past charge and discharge cycle is obtained; carrying out random masking processing on the multi-dimensional working condition information sequence in the previous charge and discharge cycle in a time sequence dimension to obtain a masked multi-dimensional working condition information sequence; the masked multi-dimensional working condition information sequence in a group of first window and the masked multi-dimensional working condition information sequence in a group of second window are taken, and the window sizes of the first window and the second window are different; and if the first window and the second window are overlapped, taking the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window as positive sample pairs, and taking the multi-dimensional working condition information sequences in the non-overlapped time sequences in the first window and the second window as negative sample pairs.

And a third step of: the feature extraction network is trained by constraining whether the feature is a positive and negative pair of samples using a transformer feature encoder to extract the feature. Specifically, a positive sample pair and a negative sample pair are input to an encoder, respectively; calculating time sequence contrast loss and sample contrast loss according to the output of the encoder, wherein the sample contrast loss takes a charge-discharge cycle loss prior term as a constraint, and the charge-discharge cycle loss prior term amplifies the sample weight of a negative sample with larger phase difference with the positive sample cycle number, and reduces the sample weight of the negative sample with smaller phase difference with the positive sample cycle number; and adjusting the encoder based on the constraint of the time sequence comparison loss and the sample comparison loss and performing iteration training until the iteration condition is met, so that the trained encoder is obtained.

Fourth step: and extracting the representation of the historical battery working condition information by using the trained encoder, and specifically, respectively inputting the multi-dimensional working condition information sequences of the marked battery in the past charge and discharge cycles into the trained encoder to obtain the battery working condition characteristics of the past charge and discharge cycles.

Fifth step: the decoder is trained. The method comprises the steps of carrying out iteration, taking battery working condition characteristics of a battery in a charge and discharge cycle of the past time as input of a decoder, and adjusting parameters of the decoder according to predicted battery performance and marking information output by the decoder until iteration ending conditions are met, so that the trained decoder is obtained.

The second stage: and predicting the battery capacity of the battery to be tested by using the trained model.

Specifically, a multidimensional working condition information sequence of the battery to be tested in a past charge and discharge cycle is obtained; respectively inputting multidimensional working condition information sequences of charge and discharge cycles in the past into a pre-trained encoder, and outputting battery working condition characteristics of the charge and discharge cycles in the past through the encoder; and (3) merging the working condition characteristics of the battery in the past charge-discharge cycle, inputting the battery into a trained decoder, and outputting the predicted battery capacity of the battery to be tested through the decoder.

According to the battery performance prediction method, the characteristic extraction is carried out on the collected working condition information through the upstream comparison representation learning method, so that the influence on model prediction results when input data are biased is reduced. By means of the downstream decoder, all historical working condition representation information is fused, influences of historical use conditions on battery capacity are comprehensively considered, and high-precision prediction of the battery capacity can be achieved. Meanwhile, different charge and discharge cycle numbers are used as model constraint information, so that characteristic information of the battery in different use stages is utilized more effectively.

It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.

Based on the same inventive concept, the embodiment of the application also provides a battery performance prediction device for realizing the above-mentioned battery performance prediction method. The implementation of the solution provided by the device is similar to the implementation described in the above method, so the specific limitations in the embodiments of the device for predicting battery performance provided below may be referred to above as limitations of the method for predicting battery performance, and will not be repeated here.

In one embodiment, as shown in fig. 7, there is provided a battery performance prediction apparatus including:

the data acquisition module 702 is configured to acquire a multidimensional working condition information sequence of the battery under test in a past charge and discharge cycle.

The encoding module 704 is configured to input the multidimensional working condition information sequences of the previous charge and discharge cycles into the pre-trained encoders respectively, and output the battery working condition characteristics of the previous charge and discharge cycles through the encoders.

The decoding module 706 is configured to fuse the battery operating condition characteristics of the previous charge and discharge cycles, input the battery operating condition characteristics to a trained decoder, and output the predicted battery performance of the battery under test through the decoder.

According to the battery performance prediction device, the multi-dimensional working condition information sequence in the previous charging cycle is input to the pre-trained encoder to obtain the battery working condition characteristics of the previous charging and discharging cycle, and then the battery performance is predicted through the encoder, on one hand, the pre-trained encoder is used for carrying out characteristic characterization on the multi-dimensional working condition information of the previous charging and discharging cycle, so that the influence on model accuracy when abnormal signals or biased signals are input is reduced, on the other hand, the battery performance is obtained based on the multi-dimensional working condition information sequence in the previous charging cycle, the multi-dimensional working condition information is fused, the influence on the battery caused by the existing use condition can be fully considered, and the accuracy of battery performance prediction is improved.

In another embodiment, the decoding module is used for splicing the battery working condition characteristics of the charge and discharge cycles for the past to obtain battery characteristics; the battery characteristics are input to a trained decoder, through which the predicted battery performance is output.

In another embodiment, the encoding module is further configured to train the encoder based on the constraint of the charge-discharge cycle loss prior term by using the positive sample pair and the negative sample pair, to obtain a trained encoder.

The decoding module is also used for respectively inputting the multi-dimensional working condition information sequences of the marked battery in the charge and discharge cycles to the trained encoder to obtain the battery working condition characteristics of the marked battery in the charge and discharge cycles.

And the training module is used for iteratively executing the steps of taking the battery working condition characteristics of the battery in the past charge and discharge cycle as the input of the decoder and adjusting the parameters of the decoder according to the predicted battery performance and the labeling information output by the decoder until the iteration ending condition is met, so as to obtain the trained decoder.

In another embodiment, the encoding module includes:

the sample construction module is used for constructing a positive sample pair and a negative sample pair based on multidimensional working condition information of a plurality of batteries in a past charge and discharge cycle;

And the training module is used for comparing, representing and learning the encoder based on the constraint of the charge-discharge cycle loss prior term by utilizing the positive sample pair and the negative sample pair to obtain a trained encoder.

In another embodiment, the sample construction module is configured to obtain a multi-dimensional working condition information sequence of the plurality of batteries in a past charge and discharge cycle; carrying out random masking processing on the multi-dimensional working condition information sequence in the previous charge and discharge cycle in a time sequence dimension to obtain a masked multi-dimensional working condition information sequence; the masked multi-dimensional working condition information sequence in a group of first window and the masked multi-dimensional working condition information sequence in a group of second window are taken, and the window sizes of the first window and the second window are different; and if the first window and the second window are overlapped, taking the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window as positive sample pairs, and taking the multi-dimensional working condition information sequences in the non-overlapped time sequences in the first window and the second window as negative sample pairs.

In another embodiment, the sample construction module is configured to take a multi-dimensional working condition information sequence of the plurality of batteries in a past charge and discharge cycle as training data; taking a multi-dimensional working condition information sequence of one charge-discharge cycle as an anchor sample in training data of a current training batch, performing data enhancement processing on the anchor sample to obtain a positive sample, and obtaining a positive sample pair according to the anchor sample and the positive sample; and taking a negative sample of the multi-dimensional working condition information sequence of one charge-discharge cycle of the non-anchor sample from the training data of the current training batch, and obtaining a negative sample pair according to the anchor sample and the negative sample.

In another embodiment, the training module is configured to input the positive and negative pairs of samples into the encoder, respectively; calculating time sequence contrast loss and sample contrast loss according to the output of the encoder, wherein the sample contrast loss takes a charge-discharge cycle loss prior term as a constraint, and the charge-discharge cycle loss prior term amplifies the sample weight of a negative sample with larger phase difference with the positive sample cycle number, and reduces the sample weight of the negative sample with smaller phase difference with the positive sample cycle number; and adjusting the encoder based on the constraint of the time sequence comparison loss and the sample comparison loss and performing iteration training until the iteration condition is met, so that the trained encoder is obtained.

Each of the modules in the above-described battery performance prediction apparatus may be implemented in whole or in part by software, hardware, and combinations thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a server, and the internal structure of which may be as shown in fig. 8. The computer device includes a processor, a memory, an Input/Output interface (I/O) and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is for storing a list of target users. The input/output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a battery performance prediction method.

It will be appreciated by those skilled in the art that the structure shown in FIG. 8 is merely a block diagram of some of the structures associated with the present inventive arrangements and is not limiting of the computer device to which the present inventive arrangements may be applied, and that a particular computer device may include more or fewer components than shown, or may combine some of the components, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor implementing the steps of the battery performance prediction method of each of the above embodiments when the computer program is executed.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which when executed by a processor, implements the steps of the battery performance prediction method of each of the above embodiments.

In one embodiment, a computer program product is provided comprising a computer program which, when executed by a processor, implements the steps of the battery performance prediction method of the above embodiments.

It should be noted that, the user information (including but not limited to user equipment information, user personal information, etc.) and the data (including but not limited to data for analysis, stored data, presented data, etc.) related to the present application are information and data authorized by the user or sufficiently authorized by each party, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related country and region.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magnetic random access Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (Phase Change Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), and the like. The databases referred to in the embodiments provided herein may include at least one of a relational database and a non-relational database. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processor referred to in the embodiments provided in the present application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, or the like, but is not limited thereto.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The foregoing examples illustrate only a few embodiments of the application and are described in detail herein without thereby limiting the scope of the application. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of the application should be assessed as that of the appended claims.

Claims

1. A battery performance prediction method, comprising:

2. The method of claim 1, wherein the fusing the battery operating characteristics of the charge and discharge cycles and inputting the fused battery operating characteristics to a trained decoder, outputting predicted battery performance of the battery under test via the decoder, comprises:

the battery working condition characteristics of the charge and discharge cycles are spliced for the time to obtain battery characteristics;

and inputting the battery characteristics into a trained decoder, and outputting predicted battery performance through the decoder.

3. The method of claim 1, wherein the manner in which the encoder and decoder are trained comprises:

training the encoder based on the constraint of the charge-discharge cycle loss prior term by using the positive sample pair and the negative sample pair to obtain a trained encoder;

respectively inputting the multi-dimensional working condition information sequences of the marked battery in the charge-discharge cycle of the calendar time into a trained encoder to obtain the working condition characteristics of the battery in the charge-discharge cycle of the calendar time;

and iteratively executing the step of taking the battery working condition characteristics of the previous charge and discharge cycle of the battery as the input of the decoder and adjusting the parameters of the decoder according to the predicted battery performance and the labeling information output by the decoder until the iteration ending condition is met, thereby obtaining the trained decoder.

4. A method according to claim 1 or 3, wherein the manner in which the encoder is trained comprises:

constructing a positive sample pair and a negative sample pair based on multidimensional working condition information of a plurality of batteries in a past charge and discharge cycle;

and comparing, expressing and learning the encoder based on the constraint of the charge-discharge cycle loss prior term by using the positive sample pair and the negative sample pair to obtain a trained encoder.

5. The method of claim 4, wherein constructing positive and negative pairs of samples based on multi-dimensional operating condition information of a plurality of batteries over a past charge-discharge cycle comprises:

acquiring multidimensional working condition information sequences of a plurality of batteries in a past charge and discharge cycle;

carrying out random masking processing on the multi-dimensional working condition information sequence in the previous charge and discharge cycle in a time sequence dimension to obtain a masked multi-dimensional working condition information sequence;

the method comprises the steps of respectively taking a masked multi-dimensional working condition information sequence in a group of first window and a masked multi-dimensional working condition information sequence in a second window, wherein the window sizes of the first window and the second window are different;

and if the first window and the second window are overlapped, taking the masked multi-dimensional working condition information sequence in the first window and the masked multi-dimensional working condition information sequence in the second window as positive sample pairs, and taking the non-overlapped multi-dimensional working condition information sequence in the first window and the non-overlapped working condition information sequence in the second window as negative sample pairs.

6. The method of claim 4, wherein constructing positive and negative pairs of samples based on multi-dimensional operating condition information of a plurality of batteries over a past charge-discharge cycle comprises:

taking a multidimensional working condition information sequence of a plurality of batteries in a past charge and discharge cycle as training data;

taking a multi-dimensional working condition information sequence of one charge-discharge cycle as an anchor sample in training data of a current training batch, performing data enhancement processing on the anchor sample to obtain a positive sample, and obtaining a positive sample pair according to the anchor sample and the positive sample;

and taking a negative sample of the multi-dimensional working condition information sequence of one charge-discharge cycle of the anchor sample from training data of the current training batch, and obtaining a negative sample pair according to the anchor sample and the negative sample.

7. The method of claim 4, wherein using the positive and negative pairs of samples to learn the comparative representation of the encoder based on constraints of a charge-discharge cycle loss prior term results in a trained encoder, comprising:

inputting the positive sample pair and the negative sample pair into the encoder, respectively;

calculating time sequence contrast loss and sample contrast loss according to the output of the encoder, wherein the sample contrast loss takes a charge-discharge cycle loss prior term as a constraint, the charge-discharge cycle loss prior term amplifies the sample weight of a negative sample with larger phase difference with the positive sample cycle number, and reduces the sample weight of a negative sample with smaller phase difference with the positive sample cycle number;

And adjusting the encoder based on the constraint of the time sequence comparison loss and the sample comparison loss, and performing iteration training until the iteration condition is met, so as to obtain a trained encoder.

8. A battery performance prediction apparatus, characterized in that the apparatus comprises:

9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 7 when the computer program is executed.

10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 7.