CN113033104B - Lithium battery state of charge estimation method based on graph convolution - Google Patents

Lithium battery state of charge estimation method based on graph convolution Download PDF

Info

Publication number
CN113033104B
CN113033104B CN202110346298.1A CN202110346298A CN113033104B CN 113033104 B CN113033104 B CN 113033104B CN 202110346298 A CN202110346298 A CN 202110346298A CN 113033104 B CN113033104 B CN 113033104B
Authority
CN
China
Prior art keywords
state
lithium battery
network
charge
voltage
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110346298.1A
Other languages
Chinese (zh)
Other versions
CN113033104A (en
Inventor
金心宇
武钿登
汪庆文
林祉谦
金昀程
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202110346298.1A priority Critical patent/CN113033104B/en
Publication of CN113033104A publication Critical patent/CN113033104A/en
Application granted granted Critical
Publication of CN113033104B publication Critical patent/CN113033104B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • General Health & Medical Sciences (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Computer Hardware Design (AREA)
  • Game Theory and Decision Science (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Geometry (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Secondary Cells (AREA)

Abstract

The invention discloses a lithium battery state of charge estimation method based on graph convolution, which comprises the following steps: acquiring an original data segment of a lithium battery from an actual vehicle, wherein the original data segment comprises current, voltage and temperature state parameters, then carrying out Min-Max Scaling standardization processing on the original data segment to obtain three-dimensional data of current, voltage and temperature states after standardization processing, and inputting the three-dimensional data into a trained lithium battery state of charge estimation network to obtain a state of charge estimation value of the lithium battery; the lithium battery state of charge estimation network comprises a forward generation cycle network DG, an adjacent matrix A and an input sequence xinA weight matrix W and a full connection layer. The method utilizes the characteristic that the graph convolution network extracts the relational features, performs feature fusion on the output of the network and the input data at the next moment to serve as the real input of the graph convolution network, is suitable for the characteristics of non-linearity and dynamic change of the charge state of the lithium battery, and greatly improves the prediction accuracy of the charge state of the lithium battery.

Description

Lithium battery state of charge estimation method based on graph convolution
Technical Field
The invention relates to the technical field of lithium batteries and neural networks, in particular to a lithium battery state of charge estimation method based on graph convolution.
Background
The research of the lithium battery state-of-charge prediction algorithm with high precision, high prediction speed and high reliability has important significance for the electric automobile. According to relevant research at home and abroad, the existing large amount of prediction research on the state of charge is found to be based on a battery equivalent model or from the perspective of an internal electrochemical mechanism of a battery, so that a large amount of battery operation data in practical engineering application is not effectively utilized, and the prediction precision and the generalization capability are hardly substantially improved. Because the lithium ion battery is still an electrochemical device in nature, and is characterized by time variation and nonlinearity, the chemical reaction process in the lithium ion battery is very complex, and simple experiments cannot be performed, therefore, a prediction method is needed, which can predict the state of charge of the lithium battery by combining a battery operation mechanism and a large amount of process data such as measurement and operation generated in the engineering practical application of the lithium ion battery.
Disclosure of Invention
The invention aims to solve the technical problem that prediction accuracy and generalization capability are insufficient in lithium battery state-of-charge estimation, and provides a lithium battery state-of-charge estimation method based on graph convolution.
In order to solve the technical problem, the invention provides a lithium battery state of charge estimation method based on graph convolution, which comprises the following steps: acquiring an original data segment of a lithium battery from an actual vehicle, wherein the original data segment comprises current, voltage and temperature state parameters, then carrying out Min-Max Scaling standardization processing on the original data segment to obtain three-dimensional data of current, voltage and temperature states after standardization processing, and inputting the three-dimensional data into a trained lithium battery state of charge estimation network to obtain a state of charge estimation value of the lithium battery;
the lithium battery state of charge estimation network comprises a forward generation cycle network DG, an adjacent matrix A and an input sequence xinA weight matrix W and a full connection layer, inputting the input sequences of the current time and the first k-1 times into a forward generation circulating network DG, building an adjacent matrix A by the output of the forward generation circulating network DG, and then inputting the input sequence x of the current time into a forward generation circulating network DGtAnd the output x of the previous momentout t-1Obtaining x after feature fusionin(ii) a Fused vector xinInner product with adjacent matrix A and weight matrix W to obtain output x at current momentout tThen will output x at the current momentout tObtaining a predicted value y of the lithium battery state of charge estimation network through one layer of full connection layer outputtWherein:
xin＝xt+xout t-1，xtexpressed as the input sequence of the current time, xout t-1Is the output of the last moment in time,
xout t＝A·xin·W，
yt＝w·xout t+ b, where w and b both represent learning parameters.
The improvement of the lithium battery state of charge estimation method based on graph convolution is as follows:
the forward generation cycle model DG is built on the basis of a gated cyclic neural network GRU, and a vector (x) is inputt,xt-1,xt-2,…,xt-k+1) Generating a cyclic network DG in the forward direction, wherein k represents the size of the time window to be selected; firstly, the k vectors are simultaneously input into a forward generation cyclic network DG, x composed of k different gate control cyclic neural networks GRUt-k+1With randomly initialized hidden vectors via gated recurrent neural network G1Calculating and outputting a hidden vector h1，G1The output hidden vector is then compared with the input vector xt-kAre sent into a gated cyclic neural network G together2Gated recurrent neural network G2The output hidden vector is then compared with the input vector xt-k-1Recurrent neural network G via gating3Obtaining the final output hidden vector h3Iteratively calculating backward according to the above until the cyclic neural network G is gatedk-1Calculated hidden vector hk-1And the input vector xtAre sent into G togetherkObtaining the final hidden vector hk。
The lithium battery state of charge estimation method based on graph convolution is further improved as follows:
the final hidden vector h output by the forward generation loop model DGkPerforming cross multiplication operation, and constructing a graph edge node network at the current moment through a Dropout layer with the fire rate of 0.3: an adjacency matrix a of size 3 × 3, expressed as:
A＝Dropout(ht T·ht) Wherein h ist＝hkGenerating a hidden vector h obtained by the cyclic model DG at the t step for the forward directiont TIs htThe transposing of (1).
The lithium battery state of charge estimation method based on graph convolution is further improved as follows:
the training process of the lithium battery state of charge estimation network comprises the following steps:
1) designing a loss function:
Figure BDA0003000811010000021
wherein the content of the first and second substances,
Figure BDA0003000811010000022
indicating the marked true state of charge value, ytRepresenting the predicted value of the state of charge of the lithium battery at the current moment output by the state of charge estimation network, wherein N represents the number of data point sets for training;
2) acquiring current, voltage and temperature state parameters of the lithium battery from an actual vehicle as an original data segment during training, marking by a high-low pressure standing inverse method to obtain the real state of charge of the lithium battery, using the real state of charge as a comparison label for training and verifying the state of charge value, and carrying out standardization processing on the original data segment to obtain standardized current, voltage and temperature data, wherein the standardized current, voltage and temperature data are divided into a training set, a verification set and a test set according to proportions of 80%, 10% and 10%;
gradient updating is carried out by using a random gradient descent optimization method, so that a loss function J (theta) is optimized, and 120 times of alternate training are carried out to finally complete model convergence; and verifying by using the verification set and the test set to obtain a trained pt model file, thereby obtaining a trained lithium battery state of charge estimation network.
The lithium battery state of charge estimation method based on graph convolution is further improved as follows:
the process of the high-low pressure standing reverse-pushing method is as follows:
1) inputting unmarked original data segments, and searching the maximum and minimum voltage standing points of the battery;
2) if the maximum voltage and the minimum voltage standing points are not included, entering the step 4); if the maximum voltage standing point and the minimum voltage standing point exist, the SOC value of the battery corresponding to the standing point is found according to the OCV curve of the open-circuit voltage of the battery, and then the state points at the highest voltage moment and the lowest voltage moment are reversely pushed for one hour to the left and the right respectively by utilizing an ampere-hour integration method to obtain the actual SOC condition of the corresponding range time point;
3) the unmarked original data segment finishes marking the state of charge value in the corresponding range time point through the step 2), then judges whether the original data segment has unmarked points outside the range, if the unmarked points do not exist, the marking is successful, and the process is ended; if the unmarked point exists, entering the step 4);
4) intercepting the data segment with the unmarked state of charge value point, adding the left data segment with the unmarked state of charge value point into the low-voltage discharge, adding the right side of the left data segment with the unmarked state of charge value point into the high-voltage charging working condition, so that the left side generates a minimum voltage standing point and the right side generates a maximum voltage standing point, then inputting the working condition into the battery simulation platform to simulate the working condition to generate real operation data and extracting, and returning to the step 2) to begin the next marking.
The invention has the following beneficial effects:
1. according to the invention, a forward generation cycle model DG is built by utilizing a gated cyclic neural network GRU, and the current, voltage and temperature states of the lithium battery at the current moment are only associated with the state values at the first k-1 moments, but not all the state values at the previous moment, so that the distance of information flow is more flexible while the calculation burden is reduced;
2. according to the method, a forward generation cycle model DG is utilized to construct an adjacent matrix, information at the front k-1 moment of the lithium battery is converted into a graph edge node network for graph convolution network state relation learning, real-time data generated by the operation of the lithium battery are continuously corrected, and the generalization capability of the model for predicting the charge state is improved;
3. the method utilizes the characteristic that the graph convolution network extracts the relational features, performs feature fusion on the output of the network and the input data at the next moment to serve as the real input of the graph convolution network, is suitable for the characteristics of non-linearity and dynamic change of the charge state of the lithium battery, and greatly improves the prediction accuracy of the charge state of the lithium battery.
Drawings
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.
FIG. 1 is a schematic diagram of a state of charge estimation network for a lithium battery according to the present invention;
FIG. 2 is a schematic flow chart of labeling the state of charge of a lithium battery by a high-low pressure standing reverse pushing method;
fig. 3 is a schematic structural comparison diagram of the gated recurrent neural network GRU and the forward generative cycle model DG.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto:
embodiment 1, a lithium battery state of charge estimation method based on graph convolution, includes the following steps:
step 1, obtaining current, voltage and temperature state parameters of a lithium battery from an actual vehicle as original data segments, inputting the original data segments into an upper computer, marking the actual charge state value of the lithium battery by a high-low pressure standing reverse pushing method, and carrying out Min-Max Scaling standardization processing on the original data segments to obtain three-dimensional data after standardization processing;
step 1.1, marking the real charge state of the lithium battery by a high-low pressure standing reverse pushing method
The high-low voltage standing inverse method aims to reduce accumulated errors caused by a current sensor, correct the charge state label value of a specific time point through charging and discharging curves near the highest and lowest standing points for multiple times, and improve the accuracy of data annotation, and as shown in fig. 2, the specific flow steps are as follows:
1) inputting unmarked original data segments, and searching the maximum and minimum voltage standing points of the battery;
2) if the maximum voltage and the minimum voltage standing points are not included, entering the step 4); if the maximum voltage standing point and the minimum voltage standing point exist, the SOC value of the battery state of charge corresponding to the standing points is found according to an Open Circuit Voltage (OCV) curve of the battery, and then an ampere-hour integration method is utilized to reversely push the state points of the highest voltage moment and the lowest voltage moment to the left and the right respectively for one hour to obtain the actual state of charge condition of a certain range of time points;
3) the unmarked original data segment finishes marking the state of charge value in a certain range of time point through the step 2), then judges whether the original data segment has unmarked points outside the range, if the unmarked points do not exist, the marking is successful, and the process is ended; if the unmarked point exists, entering the step 4);
4) intercepting a data segment with an unmarked state of charge value point, adding a low-voltage discharge condition to the left side of the data segment without the marked state of charge value point, adding a high-voltage charging condition to the right side of the data segment without the marked state of charge value point, so that a minimum voltage standing point is generated on the left side, a maximum voltage standing point is generated on the right side, then inputting the working condition into a battery simulation platform to simulate the working condition to generate real operation data and extracting, and returning to the step 2) to start the next marking;
the Open Circuit Voltage (OCV) curve of the battery is the prior art, and is obtained by carrying out working condition experiment on the corresponding type battery of an automobile in advance: discharging the battery in a full-charge state, and standing for 30 minutes to record voltage values at two ends of the battery at the moment when the charge state is reduced by 5% every time until the electric quantity of the battery is discharged, and fitting after a corresponding point set is obtained to obtain a battery discharge open-circuit voltage curve; similarly, a battery in a zero-charge state is charged, the charging operation procedure is completely the same as the discharging operation procedure, a battery charging open-circuit voltage curve is obtained, and finally the average calculation is carried out on the two curves, so that a final battery open-circuit voltage (OCV) curve is obtained;
the battery simulation platform is the prior art, the obtained current, voltage and temperature parameter data are led into the battery simulation platform, the working state of the automobile battery under the same working condition can be truly simulated, the phenomenon that data is re-accessed by using an actual automobile is avoided, on one hand, the danger caused by operating the actual automobile by a worker is avoided, and on the other hand, the experiment efficiency is improved;
finally obtaining a marked real charge state value by a high-low pressure standing reverse-pushing method;
step 1.2, respectively carrying out standardization processing on three dimensions of voltage, current and temperature in an original data section, aiming at eliminating the difference of data in different dimension characteristic sizes and increasing the precision and accuracy of a processing algorithm in the next stage, wherein the standardization processing adopts a Min-Max Scaling standardization method to carry out linear transformation on the original data so that the result is mapped to [0,1 ]]The range of (1), the implementation of the equal ratio scaling of the original data, X is the original data, Xmax、XminRespectively representing the maximum and minimum values of the data, XrThe normalized data values are represented by the following formula:
Figure BDA0003000811010000051
and finally obtaining three-dimensional data of the current, voltage and temperature states after standardization processing.
Step 2, constructing an adjacent matrix A
And constructing a forward generation cycle model DG by using the gated cycle neural network GRU, and then constructing an adjacency matrix A by using the forward generation cycle model DG.
Step 2.1, building a gated recurrent neural network GRU
Gated recurrent neural network GRU is a special type of recurrent neural network that is able to learn the potential long-term dependence of temporal information by using gating cells that can control the flow of information and alleviate the problem of gradient disappearance from information accumulation, which has two types of gates: reset gate rtAnd an update gate ztHidden state h of gated recurrent neural network GRUtCan be calculated as:
Figure BDA0003000811010000052
wherein h istImplicit in the representation of the current timeHidden state, ht-1Indicating the hidden state at the last moment in time,
Figure BDA0003000811010000053
is a candidate state calculated using new input information and is by element-by-element multiplication;
updating the door ztDetermining how much new information to update, ztIs calculated as follows:
zt＝σ(Wzxt+Uzht-1) Wherein x istAn input vector, W, representing time tzAnd UzWeight vectors representing a calculation process are initialized randomly during model initialization, and are continuously corrected along with training of a data set and updating of model parameters, an optimal value is reached after training is finished, sigma represents a sigmoid activation function, and a calculation expression of the function is as follows:
Figure BDA0003000811010000054
the activation function may map any input value to [0,1 ]]Interval, output as a probability to achieve the effect of gating selection;
candidate states
Figure BDA0003000811010000055
The calculation is as follows:
Figure BDA0003000811010000056
w and U are weight vectors when calculating candidate states, tanh is an activation function, and the specific calculation expression is as follows:
Figure BDA0003000811010000061
the activation function maps arbitrary input values to [ -1,1 ] values]The interval enables the output average value of the function to be 0, so that the training efficiency is improved, and the parameters are easy to reach the optimal quality;
reset gate rtThe amount of reservation of the previous information stream can be controlledSimilar to the update gate, the reset gate is represented as:
rt＝σ(Wrxt+Urht-1) Wherein W isrAnd UrRepresenting the calculation of weight vectors in the reset gate equation;
the representation of time t depends on all previous input vectors, and therefore the state of the t-th step is represented by the following equation:
ht＝GRU(xt,xt-1,xt-2,…,x1) Wherein x ist,xt-1,xt-2,…,x1Respectively representing input vectors x at time ttInput vector x at time t-1t-1Up to the input vector x at the starting instant 11(ii) a In order to simplify the above calculation formula, the calculation process of the gated recurrent neural network GRU is represented by GRU; it is noted that the gated recurrent neural network GRU does not compute all input vectors before time t at one time, but computes a loop nest; as shown in fig. 3(a), the hidden state at the time t-1 and the input vector at the current time t are calculated by the GRU to obtain the hidden state at the time t, and are retained as the calculation basis at the time t +1, and are continuously nested in a backward loop, the hidden state value at each time is more or less retained and transmitted all the time, the more the state value at the time point closer to the current time t is retained, and the less the state value at the time point farther from the current time t is retained;
step 2.2, building a forward generation cycle model DG by using a gated cyclic neural network GRU
For a gated recurrent neural network GRU, obtaining each state value needs to trace back to an initial value of the state; the lithium battery usually needs to work for a long time, if the state at the latest moment is calculated every time, all the past state values are taken into consideration, the calculation load is increased, and the distance of information flow is limited, so that a forward generation cycle model DG is built by using a GRU, the state at each moment is only related to the first k-1 state values, but not all the previous state values, and k is a hyper-parameter set through experiments (the k value affects the prediction accuracy, and the better prediction accuracy can be achieved by generally setting the k value to be 300 in the experiments); from step S2.1, the state at step t is briefly expressed as:
ht＝GRU(xt,xt-1,xt-2,…,xt-k+1)
wherein k represents the size of a time window to be selected, GRU is a gated cyclic neural network, the gated cyclic neural network GRU is compared with a forward generation cyclic model DG, as shown in FIG. 3, the value of k is set to 3 for clear representation, and for the sake of no loss of generality, it is assumed that 5 times are input and detailed solutions are respectively performed;
for gated recurrent neural networks GRU, as in FIG. 3(a), the input vector x for 5 time instants1、x2、x3、x4、x5Respectively input the corresponding gated cyclic neural networks G0Although the 5 time inputs are passed through the 5 gated recurrent neural networks G corresponding to each other0(indicated by the rectangles with transverse lines in FIG. 3), but they are the same G0Network, all parameters shared, five gated recurrent neural networks G identified in FIG. 3(a)0Is the same gate-controlled recurrent neural network GRU, when inputting vector x1Input via gated recurrent neural network G0Obtaining a hidden state h1Hidden state h1Then, the linear process is processed by linear variation (indicated by the rectangle with oblique lines in FIG. 3) to obtain the output y1At this time, the hidden state h1Will be retained with the input vector x2Together via a gated recurrent neural network G0Obtaining a hidden state h2(ii) a The following operation is similar to the above, but always via the same gated recurrent neural network G0The network repeatedly nests the cyclic calculation of the formula, finally obtain h5Is composed of x5,x4,x3,x2,x1All information of five inputs, therefore there may be h5＝GRU(x5,x4,x3,…,x1)；
A forward generation cycle model DG is built on the basis of a gated cyclic neural network GRU, and a vector (x) is inputt,xt-1,xt-2,…,xt-k+1) A cyclic network DG is generated in the forward direction, where k represents the time window size to be selected. Firstly, the k vectors are simultaneously input into a forward generation cyclic network DG, x composed of k different gate control cyclic neural networks GRUt-k+1With randomly initialized hidden vectors via gated recurrent neural network G1Calculating and outputting a hidden vector h1，G1The output hidden vector is then compared with the input vector xt-kAre sent into a gated cyclic neural network G together2Gated recurrent neural network G2The output hidden vector is then compared with the input vector xt-k-1Recurrent neural network G via gating3Obtaining the final output hidden vector h3Iteratively calculating backward according to the above until the cyclic neural network G is gatedk-1Calculated hidden vector hk-1And the input vector xtAre sent into G togetherkObtaining the final hidden vector hkAnd hide the vector hkWill be the only output h of the forward generated cyclic network DG at the current instant ttI.e. ht＝hk；
For the forward generation loop model DG, as shown in fig. 3(b), the k value is set to 3, and the forward generation loop network is represented by DG; different from the gated cyclic neural network GRU, the forward generation cyclic model DG shown in fig. 3(b) shares parameters with 3 different gated cyclic neural networks GRU, and performs cyclic nesting processing on inputs, the forward generation cyclic network DG identified by 5 virtual frames is the same forward generation cyclic network, and the forward generation cyclic network DG includes three gated cyclic neural networks GRU, which are G respectively1,G2,G3That is, the forward generation loop model DG needs to learn three times as many parameters as the gated recurrent neural network GRU, and three input vectors need to be input to each forward generation loop model DG (the gated recurrent neural network G of fig. 3 (a)) (see0Only one) is needed, and the three input vectors are sequentially calculated through 3 gated recurrent neural networks GRU to obtain one output; for example, the calculation process of the middle 3 rd virtual box in fig. 3(b) is: first x1,x2,x3Three input vectors are simultaneously input to generate a cyclic network DG in a forward direction, x1With randomly initialized hidden vectors (learned in post-training) via gated recurrent neural network G1Computing and outputting a hidden vector, G1The output hidden vector is then compared with the input vector x2Combined together and fed into a gated recurrent neural network G2Gated recurrent neural network G2The output hidden vector is then compared with the input vector x3Via gated recurrent neural network G3Obtaining the final output hidden vector h3Then, through linear variation, y is obtained3Therefore, can have h3＝GRU(x3,x2,x1) P in FIG. 31,p2The hidden vector h is the initial vector when the forward generation cyclic network DG network starts to calculate, is obtained by learning in the later training of the model, and is the biggest difference from the gated cyclic neural network GRU in that the hidden vector h is obtained at any momenttRelating to input only for the first k-1 moments, e.g. h5＝GRU(x5,x4,x3) Therefore, the built forward generation cycle model DG reduces the number of correlations between the state and the previous state value at any moment, so that the distance of the information flow is not longer and more flexible, and the obtained state information has more representation significance;
step 2.3, constructing an adjacent matrix A by forward generation of a cyclic model DG
In order to make the following graph convolution network learn and use the information, the hidden state at any time t includes information of the first k-1 state values, and the graph edge node network at the current time point, i.e., the adjacency matrix a with a size of 3 × 3, is constructed by performing cross product operation on the output of the cyclic model DG in step 2.2 and passing through the Dropout layer with the misfire rate of 0.3, and is expressed as:
A＝Dropout(ht T·ht)；
wherein h istFor the forward generation of the hidden vector, h, of the cyclic model DG obtained in step 2.2 at step tt TIs htTransposing;
step 3, building a lithium battery state of charge estimation network
The lithium battery state of charge estimation network is based on a graph volumeImprovement of product network, including forward generation of cyclic network DG, adjacent matrix A, input matrix (i.e. input sequence x)in) Weight matrix W and full connection layer, as shown in fig. 1; adding a forward generation circulating network DG on the basis of the original adjacency matrix A and the weight matrix W, wherein the output of the forward generation circulating network DG is used for building the adjacency matrix A of the current time point, and the output of the graph convolution network and the input data of the next moment are used for carrying out feature fusion to be used as the real input of the graph convolution network, so that the nonlinear capability of model prediction is improved; inputting the current time and the k-1 previous time into a sequence (x)t-4,xt-3,xt-2,xt-1,xt) (the input total dimension is k × 3, and the dimension at each time is 1 × 3) inputting the input forward generated cyclic network DG to obtain a 1 × 3 output matrix, then performing cross multiplication operation to obtain a 3 × 3 matrix, and building an adjacent matrix a after passing through a Dropout layer with a fire rate of 0.3; then inputting the sequence x at the current momentinInner product with adjacent matrix A and weight matrix W to obtain output x at current momentout tSetting the output at the previous time as xout t-1Then the input sequence x at the current timeinThe calculation is as follows:
xin＝xt+xout t-1wherein x istRepresenting the input sequence at the present moment, corresponding to the three-dimensional data of the current, voltage and temperature states after the normalization process acquired in step 1.2;
obtaining input sequence x of latest moment by feature fusioninFor improving the nonlinear capability of generating the cyclic model, and then further representing the model output x at the current momentout t：
xout t＝A·xin·W
The weight matrix W is a parameter matrix capable of learning and training, when a model is initialized, the value of the weight matrix W is initialized randomly, the weight values are continuously corrected along with the training of data and the updating of model parameters, and the optimal value is reached after the training is finished;
finally, at the current momentModel output xout tObtaining the output charge state predicted value y of the lithium battery charge state estimation network through a full connection layertIt can be expressed as:
yt＝w·xout t+ b, where w and b both represent learning parameters;
the characteristic of extracting the relation characteristic by using the graph convolution network is utilized, the output of the network and the input data at the next moment are subjected to characteristic fusion to serve as the real input of the graph convolution network, the characteristics of nonlinearity and dynamic change of the lithium battery state of charge are adapted, the final lithium battery state of charge value is predicted by utilizing the collected lithium battery current, voltage and temperature state parameters, and the prediction accuracy of the lithium battery state of charge is greatly improved;
step 4, training and testing of lithium battery state of charge estimation network
Step 4.1, designing a loss function, performing loss calculation on a characteristic value level on the output of the lithium battery state of charge estimation network, and using a mean square error function, wherein an objective function of the loss function is shown in the following formula:
Figure BDA0003000811010000091
wherein the content of the first and second substances,
Figure BDA0003000811010000092
indicating the marked true state of charge value, ytThe predicted value of the lithium battery state of charge estimation network output is represented as the output of the step 3, and N represents the number of the data point sets for training;
step 4.2, training
The original data does not know the real state of charge value of the lithium battery, the real state of charge of the lithium battery is obtained by labeling through a high-low pressure standing reverse pushing method in the step 1.1 and is used as a comparison label during model training and verification, namely the standard of the state of charge value of the model training and verification, meanwhile, 86400 pieces of data (24 hours, 1 piece is obtained in 1 second) which are processed in a standardized way are obtained through the step 1.2 and are used as a training data set, the data set is divided into a training set, a verification set and a test set according to the proportion of 80%, 10% and 10%, wherein the training set is used for model training, and the verification set and the test set are used for final model evaluation;
performing gradient update using a stochastic gradient descent optimization method, thereby optimizing an overall objective function J (θ); because the forward generation cycle model DG and the graph convolution network have different structures and parameter numbers, different learning rates are respectively set, wherein the learning rate of the parameters of the forward generation cycle model DG is 3e-5, the learning rate of the parameters of the graph convolution network is 5e-5, no activation function is used, the size of batch is 8, model convergence is finally completed by alternate training, and accurate prediction of the battery charge state under a complex working condition is realized;
and (4) training the network for 120 times, and verifying by using a verification set to obtain a trained pt model file, thereby obtaining the trained lithium battery state of charge estimation network.
Step 4.3, testing
Evaluation is carried out in a test stage by using indexes of Mean Absolute Error (MAE), Mean Square Error (MSE) and R Square (RS), wherein the average absolute error: the index is an expected value of absolute error loss, and is an average value of the whole sample number obtained after the absolute value of the difference between the predicted value and the true value of the sample is summed, so that the situation that the error is offset positively and negatively can be effectively avoided, and the average absolute error of N samples can be represented by the following formula:
Figure BDA0003000811010000093
Figure BDA0003000811010000094
is the true value of the sample, ytIs the predicted value of the model.
Mean square error: the index is an expectation for square error, the value of the index is inconsistent with the dimension of a target variable, the average value of the whole number of samples is obtained after the square sum of the difference between the predicted value and the true value of the samples is based on, and the mean square error of N samples can be represented by the following formula:
Figure BDA0003000811010000101
Figure BDA0003000811010000102
is the true value of the sample, ytIs the predicted value of the model.
R square: also known as the coefficient of determinism or goodness of fit, reflects the degree of fit between the predicted value and the true value, the closer to 1, the better the model fits, and the R-squared of N samples can be represented by the following formula:
Figure BDA0003000811010000103
Figure BDA0003000811010000104
is the true value of the sample, ytIs the predicted value of the model, y0Is the mean of the real samples.
Loading a trained pt model file, and predicting the state of charge of the lithium battery by taking a test set as input, wherein the result proves that the method has obvious advantages; the error range of the lithium battery state-of-charge value predicted by the model and the marked real state-of-charge value is controlled within 3.5 percent (the average sum of the error indexes of the three is within 5 percent), and the error between the estimated value and the true value is within 5 percent, so that the method has the condition of application in the industrial field.
Step 5, estimating the state of charge of the lithium battery
The method comprises the steps of obtaining an original data section of the lithium battery from an actual vehicle, inputting the original data section of the lithium battery into an upper computer, wherein the original data section comprises current, voltage and temperature state parameters, then carrying out Min-Max Scaling standardization processing on the original data section in the upper computer through step 1.2 to obtain three-dimensional data of the current, voltage and temperature states after standardization processing, and inputting the three-dimensional data into a trained lithium battery state of charge estimation network to obtain a state of charge estimation value of the lithium battery.
Experiment 1:
all data sets used for training in example 1 were used as data sets for experiment 1, and a Tensorflow framework was used to build a model, and the GPU graphics card used NVIDIA 1050 Ti. Dividing a data set into a training set, a verification set and a test set according to the proportion of 80%, 10% and 10%, and simultaneously building four neural networks such as an extended Kalman, a BP neural network, a classical GRU and the method of the invention as comparison tests; each neural network for comparison is used for training all training sets 120 times, after each training is completed, a verification set is used for verifying an evaluation model, the data set of experiment 1 totally comprises 8640 pieces of data, the first 3800 pieces of data in the data set are respectively evaluated by Mean Absolute Error (MAE), Mean Square Error (MSE) and R Square (RS) indexes, and the results are shown in the following table:
Figure BDA0003000811010000105
Figure BDA0003000811010000111
error results of the three evaluation indexes MSE, MAE and R square are shown in detail in the table, wherein the front 3800 data are subjected to subsection test evaluation, the front 10%, 30%, 50%, 70% and 90% of data are separately tested, and the unit of the obtained error result is a percentage (%) value; according to the test results in the table, the method has obvious advantages on three evaluation indexes compared with the existing extended Kalman, BP neural network and classical GRU model based on the data set of the lithium battery of the electric vehicle obtained by the real vehicle.
Finally, it is also noted that the above-mentioned lists merely illustrate a few specific embodiments of the invention. It is obvious that the invention is not limited to the above embodiments, but that many variations are possible. All modifications which can be derived or suggested by a person skilled in the art from the disclosure of the present invention are to be considered within the scope of the invention.

Claims (3)

1. A lithium battery state of charge estimation method based on graph convolution is characterized by comprising the following steps:
acquiring an original data segment of a lithium battery from an actual vehicle, wherein the original data segment comprises current, voltage and temperature state parameters, then carrying out Min-Max Scaling standardization processing on the original data segment to obtain three-dimensional data of current, voltage and temperature states after standardization processing, and inputting the three-dimensional data into a trained lithium battery state of charge estimation network to obtain a state of charge estimation value of the lithium battery;
the lithium battery state of charge estimation network comprises a forward generation cycle network DG, an adjacent matrix A and an input sequence xinA weight matrix W and a full connection layer, inputting the input sequences of the current time and the first k-1 times into a forward generation circulating network DG, building an adjacent matrix A by the output of the forward generation circulating network DG, and then inputting the input sequence x of the current time into a forward generation circulating network DGtAnd the output x of the previous momentout t-1Obtaining x after feature fusionin(ii) a Fused vector xinInner product with adjacent matrix A and weight matrix W to obtain output x at current momentout tThen outputs x of the current time are outputtedout tObtaining a predicted value y of the lithium battery state of charge estimation network through one layer of full connection layer outputtWherein:
xin＝xt+xout t-1，xtexpressed as the input sequence of the current time, xout t-1Is the output of the last moment in time,
xout t＝A·xin·W，
yt＝w·xout t+ b, where w and b both represent learning parameters;
the forward generation cyclic network DG is built based on a gated cyclic neural network GRU, and a vector (x) is inputt,xt-1,xt-2,…,xt-k+1) Generating a cyclic network DG over the forward direction, wherein k represents the size of the time window to be selected; the k vectors are input into a gated recurrent neural network GRU composed of k different gate control unitsForward generation of a cyclic network DG, xt-k+1With randomly initialized hidden vectors via gated recurrent neural network G1Calculating and outputting a hidden vector h1，G1The output hidden vector is then compared with the input vector xt-kAre sent into a gated cyclic neural network G together2Gated recurrent neural network G2The output hidden vector is then compared with the input vector xt-k-1Recurrent neural network G via gating3Obtaining the final output hidden vector h3Iteratively calculating backward according to the above until the cyclic neural network G is gatedk-1Calculated hidden vector hk-1And the input vector xtAre sent into G togetherkObtaining the final hidden vector hk。
2. The lithium battery state-of-charge estimation method based on graph convolution according to claim 1, characterized in that:
the final hidden vector h output by the forward generated cyclic network DGkPerforming cross multiplication operation, and constructing a graph edge node network at the current moment through a Dropout layer with the fire rate of 0.3: an adjacency matrix a of size 3 × 3, expressed as:
A＝Dropout(ht T·ht) Wherein h ist＝hkGenerating a hidden vector h obtained in the t step for the forward generation of the cyclic network DGt TIs htThe transposing of (1).
3. The lithium battery state-of-charge estimation method based on graph convolution according to claim 2, characterized in that:
the training process of the lithium battery state of charge estimation network comprises the following steps:
1) designing a loss function:
Figure FDA0003555372340000021
wherein the content of the first and second substances,
Figure FDA0003555372340000022
indicating the marked true state of charge value, ytRepresenting the predicted value of the state of charge of the lithium battery at the current moment output by the state of charge estimation network, wherein N represents the number of data point sets for training;
2) acquiring current, voltage and temperature state parameters of the lithium battery from an actual vehicle as an original data segment during training, marking by a high-low pressure standing reverse method to obtain the real state of charge of the lithium battery, using the real state of charge as a comparison label for training and verifying the state of charge value, and carrying out standardization processing on the original data segment to obtain standardized current, voltage and temperature data which are divided into a training set, a verification set and a test set according to the proportion of 80%, 10% and 10%;
gradient updating is carried out by using a random gradient descent optimization method, so that a loss function J (theta) is optimized, and model convergence is finally completed through alternate training for 120 times; using a verification set and a test set to verify to obtain a trained pt model file, thereby obtaining a trained lithium battery state of charge estimation network;
the process of the high-low pressure standing reverse-pushing method is as follows:
1) inputting unmarked original data segments, and searching the maximum and minimum voltage standing points of the battery;
2) if the maximum voltage and the minimum voltage standing points are not included, entering the step 4); if the maximum voltage standing point and the minimum voltage standing point exist, the SOC value of the battery corresponding to the standing point is found according to the OCV curve of the open-circuit voltage of the battery, and then the state points at the highest voltage moment and the lowest voltage moment are reversely pushed for one hour to the left and the right respectively by utilizing an ampere-hour integration method to obtain the actual SOC condition of the corresponding range time point;
3) completing the marking of the state of charge value in the corresponding range time point by the unmarked original data segment in the step 2), judging whether the original data segment has unmarked points outside the range, if not, marking successfully, and ending the process; if the unmarked point exists, entering the step 4);
4) intercepting a data segment with an unmarked state of charge value point, adding a left data segment with the unmarked state of charge value point into a low-voltage discharge state, adding a high-voltage charging working condition on the right side of the left data segment with the unmarked state of charge value point, generating a minimum voltage standing point on the left side and a maximum voltage standing point on the right side, inputting the working condition into a battery simulation platform to simulate the working condition to generate real operation data, extracting, and returning to the step 2) to begin the next marking.
CN202110346298.1A 2021-03-31 2021-03-31 Lithium battery state of charge estimation method based on graph convolution Active CN113033104B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110346298.1A CN113033104B (en) 2021-03-31 2021-03-31 Lithium battery state of charge estimation method based on graph convolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110346298.1A CN113033104B (en) 2021-03-31 2021-03-31 Lithium battery state of charge estimation method based on graph convolution

Publications (2)

Publication Number Publication Date
CN113033104A CN113033104A (en) 2021-06-25
CN113033104B true CN113033104B (en) 2022-05-24

Family

ID=76453147

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110346298.1A Active CN113033104B (en) 2021-03-31 2021-03-31 Lithium battery state of charge estimation method based on graph convolution

Country Status (1)

Country Link
CN (1) CN113033104B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113933725B (en) * 2021-09-08 2023-09-12 深圳大学 Method for determining state of charge of power battery based on data driving
CN115061050B (en) * 2022-08-19 2022-11-01 南通博鎏机电科技有限公司 Intelligent monitoring method for battery electric quantity

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110674922A (en) * 2019-08-14 2020-01-10 西北工业大学 Network representation obtaining method based on deep learning
EP3620990A1 (en) * 2018-09-05 2020-03-11 Siemens Aktiengesellschaft Capturing network dynamics using dynamic graph representation learning
CN112067998A (en) * 2020-09-10 2020-12-11 昆明理工大学 Lithium ion battery state of charge estimation method based on deep neural network
EP3786653A1 (en) * 2018-04-23 2021-03-03 Samsung SDI Co., Ltd. Battery state estimation method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11637331B2 (en) * 2017-11-20 2023-04-25 The Trustees Of Columbia University In The City Of New York Neural-network state-of-charge and state of health estimation

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3786653A1 (en) * 2018-04-23 2021-03-03 Samsung SDI Co., Ltd. Battery state estimation method
EP3620990A1 (en) * 2018-09-05 2020-03-11 Siemens Aktiengesellschaft Capturing network dynamics using dynamic graph representation learning
CN110674922A (en) * 2019-08-14 2020-01-10 西北工业大学 Network representation obtaining method based on deep learning
CN112067998A (en) * 2020-09-10 2020-12-11 昆明理工大学 Lithium ion battery state of charge estimation method based on deep neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
State-of-Charge Estimation for Lithium-Ion Batteries Based on a Nonlinear Fractional Model;Baojin Wang等;《IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY》;20170101;全文 *
基于ECTLBO-ELM模型的荷电状态估计;周勇等;《广西大学学报(自然科学版)》;20210225;全文 *

Also Published As

Publication number Publication date
CN113033104A (en) 2021-06-25

Similar Documents

Publication Publication Date Title
US20220114421A1 (en) Method for estimating the capacity of lithium battery based on convolution long-short-term memory neural network
Yang et al. State-of-charge estimation of lithium-ion batteries based on gated recurrent neural network
CN113419187B (en) Lithium ion battery health estimation method
CN110398697B (en) Lithium ion health state estimation method based on charging process
CN110888059B (en) Charge state estimation algorithm based on improved random forest combined volume Kalman
CN113805064B (en) Lithium ion battery pack health state prediction method based on deep learning
Mamo et al. Long short-term memory with attention mechanism for state of charge estimation of lithium-ion batteries
CN111443293A (en) Lithium battery state of health (SOH) estimation method based on data driving
CN113033104B (en) Lithium battery state of charge estimation method based on graph convolution
CN111537884B (en) Method and device for acquiring service life data of power battery, computer equipment and medium
Ma et al. Robust state of charge estimation for Li-ion batteries based on cubature kalman filter with generalized maximum correntropy criterion
CN111783361A (en) Numerical weather forecast mixed data assimilation method based on triple multi-layer perceptron
CN114397577A (en) New energy automobile lithium battery health state assessment method based on ASTUKF-GRA-LSTM model
CN113406503A (en) Lithium battery SOH online estimation method based on deep neural network
CN116298936A (en) Intelligent lithium ion battery health state prediction method in incomplete voltage range
Schmitt et al. Effective estimation of battery state-of-health by virtual experiments via transfer-and meta-learning
Li et al. State of charge estimation of lithium-ion batteries based on PSO-TCN-Attention neural network
Schmitt et al. State-of-health estimation by virtual experiments using recurrent decoder–encoder based lithium-ion digital battery twins trained on unstructured battery data
Hatherall et al. Remaining discharge energy estimation for lithium-ion batteries using pattern recognition and power prediction
Wang et al. Research on SOC estimation method of hybrid electric vehicles battery based on the grey wolf optimized particle filter
CN116840724A (en) Lithium ion battery health condition estimation and residual service life prediction method
Schorsch et al. SOC and SOH estimation for Li-ion battery based on an equivalent hydraulic model. Part II: SOH power fade estimation
Li et al. State of charge estimation of lithium-ion batteries using improved BP neural network and filtering techniques
Kuang et al. State-of-charge estimation hybrid method for lithium-ion batteries using BiGRU and AM co-modified Seq2Seq network and H-infinity filter
Jiao et al. A methodology for state of health estimation of battery using short-time working condition aging data

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant