Lithium ion battery state-of-charge estimation method and system based on improved NARX neural network

Abstract

The invention relates to the technical field of automobile battery detection, and discloses a lithium ion battery state-of-charge estimation method and system based on an improved NARX neural network, and the method comprises the steps: collecting dynamic parameters of a lithium ion battery during operation; performing standardization processing on the collected dynamic parameters, eliminating noise interference through filtering, and calibrating nonlinear characteristics of the battery to obtain a standardized data set; constructing a hybrid model in which the NARX neural network and the LSTM neural network are connected in series; inputting the standardized data set into a hybrid model for training to obtain a trained hybrid model; and inputting dynamic parameters acquired in real time into the trained hybrid model, and outputting a state-of-charge estimation value of the lithium ion battery. According to the lithium ion battery state-of-charge estimation method and system based on the improved NARX neural network, the NARX neural network and the LSTM neural network are combined in series, so that the stability and precision of lithium ion battery state-of-charge estimation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of automotive battery detection, and particularly to a method and system for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network. Background Art

[0002] Currently, there are many vehicle technologies using batteries on the market, and the ampere-hour integration method is mainly used for current sampling and estimation. However, the sampling accuracy and sampling interval of the ampere-hour integration method have certain uncertainties or variabilities because each sampling may be affected by various factors such as temperature changes, battery aging, different charge and discharge rates, and battery self-discharge. As a result, the sampling results are not exactly the same, which will affect the accuracy of the measurement results and thus cause errors.

[0003] The state of charge of a lithium-ion battery is an important parameter in the battery management system, which reflects the ratio between the remaining battery charge and the fully charged state. Accurately estimating the state of charge of a lithium-ion battery is crucial for battery management technology. However, due to the influence of factors such as external interference, temperature changes, and electromagnetic interference, accurately estimating the state of charge of a lithium-ion battery is a typical non-linear instability problem.

[0004] During the training process of the NARX neural network, gradient explosion and gradient disappearance may occur, resulting in the parameters being easily trapped in local optima, thus affecting the estimation accuracy of the state of charge of the lithium-ion battery. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention proposes a method and system for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network to solve the above technical problems.

[0006] In the first aspect, a method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network is provided, including:

[0007] Collect dynamic parameters during the operation of the lithium-ion battery, including voltage, current, and temperature;

[0008] Perform standardization processing on the collected dynamic parameters, eliminate noise interference through filtering, and calibrate the non-linear characteristics of the battery to obtain a standardized data set;

[0009] Construct a hybrid model combining a series connection of an NARX neural network and an LSTM neural network;

[0010] Input the standardized data set into the hybrid model for training to obtain a trained hybrid model;

[0011] Input the dynamically collected parameters in real time into the trained hybrid model, and output the estimated value of the state of charge of the lithium-ion battery.

[0012] Further, the construction of the hybrid model includes:

[0013] The NARX neural network receives the historical output sequence and the historical input sequence through a time window sliding mechanism, and generates intermediate features based on a non-linear activation function;

[0014] The LSTM neural network receives the intermediate features and the dynamic parameters at the current moment, updates the cell state through an input gate, a forget gate and an output gate mechanism, and outputs the state of charge prediction value.

[0015] Further, the acquisition of the dynamic data is based on the DST standard working condition and the WLTC standard working condition.

[0016] Further, the training process includes:

[0017] The battery health state is preliminarily estimated by the NARX neural network to obtain prediction features, where the prediction features are based on the non-linear mapping of the historical output sequence and the historical input sequence;

[0018] The prediction features are concatenated with the dynamic parameters at the current moment to obtain an enhanced input vector;

[0019] The enhanced input vector is input into the LSTM neural network, and the model is iteratively trained by using the backpropagation algorithm combined with the Adam optimizer until the error index converges to a preset threshold.

[0020] Further, the method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network further includes:

[0021] The grey relational analysis method is used to verify the correlation degree between the dynamic data during the operation of the lithium-ion battery and the battery health state, and the optimal feature combination with a correlation degree higher than a preset threshold is selected;

[0022] After the model outputs, the prediction result is de-normalized and converted into a percentage value of the actual battery health state.

[0023] In a second aspect, there is provided a system for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network, based on the method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network described in any one of the foregoing, including:

[0024] A data acquisition module, configured to acquire dynamic parameters during the operation of the lithium-ion battery, including voltage, current and temperature;

[0025] A preprocessing module, configured to perform normalization processing on the acquired dynamic parameters, eliminate noise interference through filtering, and calibrate the non-linear characteristics of the battery at the same time to obtain a normalized data set;

[0026] A hybrid model module, configured to build a hybrid model combining a NARX neural network and an LSTM neural network in series;

[0027] A training module, configured to input the standardized data set into the hybrid model for training to obtain a trained hybrid model;

[0028] An output module, configured to input dynamically collected dynamic parameters into the trained hybrid model and output an estimated value of the state of charge of the lithium-ion battery.

[0029] Further, the hybrid model module includes:

[0030] An input layer, configured to receive the standardized voltage, current, temperature data, and the sequence of historical battery health states;

[0031] A hidden layer, including NARX neural network units and LSTM neural network units, configured for feature extraction and time-dependence modeling;

[0032] An output layer, mapped to an estimated value of the battery health state through a fully connected network and output an error confidence interval.

[0033] Further, it further includes a dynamic verification module, which calculates the estimation error of the battery health state under the current working condition in real time. If the error exceeds a preset threshold, the model parameters are incrementally updated through an online learning algorithm, and the updated model is stored in the local database.

[0034] In a third aspect, there is provided a system for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network, including a processor and a memory storing program instructions, where the processor is configured to execute the method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network as described in any one of the foregoing when running the program instructions.

[0035] The invention adopting the above technical solution has the following advantages:

[0036] By combining the NARX neural network and the LSTM neural network in series, the invention improves the stability and accuracy of the estimation of the state of charge of the lithium-ion battery. Description of the Drawings

[0037] In order to more clearly illustrate the specific embodiments of the present invention, the drawings required for use in the specific embodiments will be briefly introduced below. In all the drawings, the elements or parts are not necessarily drawn to scale.

[0038] Figure 1Flow chart of the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention;

[0039] Figure 2 Structure of the NARX neural network in the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention Figure 1 ;

[0040] Figure 3 Structure of the NARX neural network in the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention Figure 2 ;

[0041] Figure 4 Network architecture diagram of the time series prediction model in the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention;

[0042] Figure 5 Simulation diagram of the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention Figure 1 ;

[0043] Figure 6 Simulation diagram of the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention Figure 2 ;

[0044] Figure 7 Schematic diagram of the training result of the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention. Detailed implementation manners

[0045] The embodiments of the technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, so they are only examples and cannot be used to limit the protection scope of the present invention.

[0046] As Figures 1 to 7 shown, the lithium-ion battery state of charge estimation method based on the improved NARX neural network of the present invention includes:

[0047] Step S01, collecting dynamic parameters during the operation of the lithium-ion battery, including voltage, current, and temperature;

[0048] Step S02, performing standardization processing on the collected dynamic parameters, eliminating noise interference through filtering, and calibrating the non-linear characteristics of the battery to obtain a standardized data set;

[0049] Step S03, constructing a hybrid model combining a series connection of NARX neural network and LSTM neural network;

[0050] Step S04: Input the standardized data set into the hybrid model for training to obtain the trained hybrid model;

[0051] Step S05: Input the dynamically collected parameters in real time into the trained hybrid model and output the estimated value of the state of charge of the lithium-ion battery.

[0052] Specifically, the first step: We first collect the current, voltage, and SOC data measured by the battery under the DST working condition and the FUDS working condition;

[0053] The second step: Select the data set D=(X, y), where X represents the voltage V(k) and current at the sampling time k, and y(k) represents SOC(k);

[0054] The third step: Train the NRAX model y(t)=f(y(t-1)...,y(t-D y ),u(t-D u )...u(t-1))+e(t);

[0055] Among them, y (t) represents the output at the sampling time t, u (t) represents the external input at the sampling time t, D y and D u are the time series (maximum delay order) of the output vector and the input vector respectively. f(.) is the non-linear function to be fitted, and e(t) is the error term;

[0056] The fourth step: Use NRAX to complete the first SOC estimation and add the predicted SOC to the data set;

[0057] The fifth step: Divide the re-constituted data set into a training set and a validation set, and first complete the training of the LSTM model.

[0058] In this embodiment, the construction of the hybrid model includes:

[0059] The NARX neural network receives the historical output sequence and the historical input sequence through the time window sliding mechanism and generates intermediate features based on the non-linear activation function;

[0060] The LSTM neural network receives the intermediate features and the dynamic parameters at the current moment, updates the cell state through the input gate, forget gate, and output gate mechanisms, and outputs the predicted value of the state of charge.

[0061] In this embodiment, the acquisition of dynamic data is based on the DST standard working condition and the WLTC standard working condition.

[0062] In this embodiment, the training process includes:

[0063] The initial estimation of the battery health state is carried out through a NARX neural network to obtain prediction features, where the prediction features are based on the non-linear mapping of the historical output sequence and the historical input sequence;

[0064] The prediction features are concatenated with the dynamic parameters at the current moment to obtain an enhanced input vector;

[0065] The enhanced input vector is input into the LSTM neural network, and the model is iteratively trained using the backpropagation algorithm combined with the Adam optimizer until the error index converges to a preset threshold.

[0066] In this embodiment, the method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network further includes:

[0067] The grey relational analysis method is used to verify the correlation between the dynamic data during the operation of the lithium-ion battery and the battery health state, and the optimal feature combination with a correlation higher than the preset threshold is selected;

[0068] After the model output, the prediction result is de-normalized and converted into a percentage value of the actual battery health state.

[0069] Specifically, for a standard NARX dynamic neural network, where y(t) is the output sequence at time t and x(t) is the input sequence at time t.

[0070] Mathematical structure expression: y(t) = f(y(t - 1) y(t - 2)... y(t - ny) x(t - 1) x(t - 2)... x(t - nx));

[0071] Among them, y(t - 1) is the output sequence at time t - 1, x(t - 1) is the input sequence at time t - 1, y(t - ny) is the historical input sequence, and x(t - nx) is the historical output sequence.

[0072] Since in the training of the NARX dynamic neural network, the expected output is fed back to the input end, which makes the prediction effect of the NARX neural network better, and the NARX dynamic neural network becomes a unidirectional neural network, simplifying the network structure.

[0073] Mathematical structure expression as formula y(t) = ^f(y(t - 1); x(t - 1)) = ^f(y(t - 1) y(t - 2)... y(t - ny) x(t - 1) x(t - 2)... x(t - nx));

[0074] Among them, y(t) is the predicted value of RxL, ^f is the non-linear function under this network. NARX has a good memory function and can be used to process time series. Extract the required current and voltage, BMS current and voltage, and the temperature during charging as the indirect health factors of lithium-ion degradation information to replace the traditional capacity and internal resistance direct health factors. Judge the accuracy of the proposed health factors through grey relational analysis, and establish the remaining life model of lithium-ion batteries based on dynamic neural networks. At the same time, the LSTM neural network technology will also be used.

[0075] Specifically, electric vehicles will face various different working conditions during actual driving, and there are also various uncertain factors. Therefore, the extraction of training features before building the prediction model is very crucial. While being able to better reflect the actual working state of the power battery, the data should also be easy to obtain. Measuring various parameters inside the battery such as internal resistance and electrolyte solution concentration changes often requires building precise models or using large instruments, and can only be obtained when static, which is not suitable for actual working conditions. However, parameters such as voltage, current, and temperature during the battery's working process can obtain relatively accurate values with the help of the battery management system, and these parameters can directly reflect the working state of the battery and play an important role in SOC estimation. Therefore, this paper selects voltage, current, and temperature as the input features of the model to enhance the applicability and practicality of the algorithm.

[0076] To verify the effectiveness of the proposed model, the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are selected as the model accuracy measurement criteria. The calculation formulas are as follows:

[0077]

[0078] Among them, M t and N t respectively represent the observed value of the battery SOC at time t and the output value predicted by the model. f is the number of prediction samples.

[0079] The smaller the values of RMSE, MAE, and MAPE, the higher the model accuracy and the better the applicability.

[0080] Then, a lithium-ion battery with a capacity of 2.5 Ah is used as the test object, and data of current, voltage, and SOC are collected at an ambient temperature of 20°C under different battery working conditions.

[0081] During the verification period, the battery will be fully discharged and then left standing for a period of time before use to avoid the influence of the non-linear characteristics inside the battery on the experimental results.

[0082] The experimental verification results show that under the DST standard working condition and the WLTC standard working condition, the root mean square errors of the lithium battery state of charge of this model reach 3.38×10^8.75×10^-4 respectively. Compared with the unimproved algorithm, the estimation accuracy is improved by 42.4% and 20.5% respectively in terms of the root mean square error.

[0083] The model combines the advantages of the NARX model and the LSTM model. It not only breaks the limitations of traditional models, but also retains the LSTM memory unit, has external input non-linearity, and can effectively prevent gradient explosion and gradient disappearance.

[0084] The proposed algorithm has been verified under different experimental working conditions. The experimental results show that this model has high stability and estimation performance. Compared with the LSTM, NARX and LSSVM algorithms, the total error of the proposed model is controlled within 2%, further verifying the effectiveness and robustness of the algorithm.

[0085] In some other embodiments, there is provided a method for estimating the state of charge of a lithium-ion battery based on the improved NARX neural network according to any one of the foregoing, including:

[0086] A data acquisition module configured to acquire dynamic parameters during the operation of the lithium-ion battery, including voltage, current and temperature;

[0087] A preprocessing module configured to perform standardization processing on the acquired dynamic parameters, eliminate noise interference through filtering, and calibrate the non-linear characteristics of the battery to obtain a standardized data set;

[0088] A hybrid model module configured to construct a hybrid model combining a series connection of an NARX neural network and an LSTM neural network;

[0089] A training module configured to input the standardized data set into the hybrid model for training to obtain a trained hybrid model;

[0090] An output module configured to input the dynamically acquired real-time parameters into the trained hybrid model and output an estimated value of the state of charge of the lithium-ion battery.

[0091] In this embodiment, the hybrid model module includes:

[0092] An input layer for receiving the standardized voltage, current, temperature data and the sequence of the historical battery health state;

[0093] A hidden layer containing NARX neural network units and LSTM neural network units for feature extraction and time-dependent modeling;

[0094] An output layer mapped to an estimated value of the battery health state through a fully connected network and outputting an error confidence interval.

[0095] Specifically, an LSTM-based power battery SOC prediction model is established. The model mainly consists of an input layer, a hidden layer, and an output layer. Among them, X represents the input features, and Y represents the output features;

[0096] The input layer receives the external data of the battery and performs data preprocessing and dataset partitioning, etc., and converts the input data into a format that can be processed inside the neural network;

[0097] In the output layer, the trained LSTM model is used to calculate the predicted output data, and inverse normalization and other processing are performed on the predicted data.

[0098] In the following formula, tanh and σ represent activation functions. The input gate, forget gate, output gate, and cell state corresponding to time t are represented by i t , f t , o t and c t respectively, and the corresponding weight matrices are represented by w i , w f , w o and w c respectively. The biases of each gate state are represented by b i , b f , b o and b c respectively.

[0099] it = σ(w i h t-1 + w i x t + b i ) is determined by the sigmoid layer of the forget gate, and part of the information in the cell state is forgotten. The output data h t-1 from the previous layer and the input data x t at this moment are used as new inputs to obtain the output of the cell state at time t - 1.

[0100] f t = σ(w f h t-1 + w f x t + b f )c t = f t c t-1 + i t tanh(w c h t-1 + w c x t + b c ) is to incorporate the information

[0101] Stored in the cell state, mainly completed by the following steps. First, take the result i of the sigmoid layer in the input gate t as the information to be updated; then create a new vector c by the tanh layer t and add it to the cell state; finally, multiply f t with the old cell state c t-1 to complete the information forgetting function, and add the candidate information i t c t to update the cell state.

[0102] o t =σ(w o h t-1 +w o x t +b o ) is the sigmoid layer processing the cell state, selecting the part of the information to be output, and then processing it with tanh. Multiply the two parts of the information obtained to get the output value. From the above, the output signal of the LSTM at time t is h t =o t tanh(c t ).

[0103] In the time dimension direction, the hidden layers contained in the LSTM model are unfolded to form a deep learning network, automatically mining the internal relationship between the numerical change trend of the SOC and these influencing factors, and realizing the prediction of the power battery SOC. This device integrates NARX and LSTM to obtain a more accurate data.

[0104] In this embodiment, it further includes a dynamic verification module. The dynamic verification module calculates the estimation error of the battery health state under the current working condition in real time. If the error exceeds the preset threshold, the model parameters are incrementally updated through an online learning algorithm, and the updated model is stored in the local database.

[0105] The present invention uses the current temperature measured by transaction detection, and is dynamically detected relative to traditional measurement. This invention has a more accurate NARX and LSTM combined monitoring, which can reduce the data error to 2% error, and the measured data is more accurate.

[0106] Using matlab simulation analysis, considering both performance overhead and prediction effect, the current and voltage are selected as eigenvalue, and the LSTM layer has 50 hidden units to learn the time dependence relationship of the input data.

[0107] LSTM has a total of 5 layers for a part of data simulation verification results as Figure 5 shown, and the detailed internal training data results as Figure 7 shown. Through Figure 5The actually measured changes and fluctuations in current and voltage Figure 6 By comparing with the NARX algorithm improved by the LSTM algorithm, it can be seen that the error of the model in data is less than 2%, indicating that this algorithm has good accuracy and meets the design requirements.

[0108] Figure 7 Some of the internal training results are as follows:

[0109] Epoch [100 / 100], Loss: 0.000006 Training completed, Average relative error: 0.19%.

[0110] In some other embodiments, a lithium-ion battery state of charge estimation system based on an improved NARX neural network is provided, including a processor and a memory storing program instructions. The processor is configured to execute the lithium-ion battery state of charge estimation method based on the improved NARX neural network as described in any one of the foregoing when running the program instructions.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. A method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network, characterized in that, Including: Collect dynamic parameters during the operation of a lithium-ion battery, including voltage, current, and temperature; Perform standardization processing on the collected dynamic parameters, eliminate noise interference through filtering, and calibrate the non-linear characteristics of the battery to obtain a standardized data set; Construct a hybrid model combining a NARX neural network and an LSTM neural network in series; Input the standardized data set into the hybrid model for training to obtain a trained hybrid model; Input the dynamically collected parameters in real time into the trained hybrid model and output the estimated value of the state of charge of the lithium-ion battery.

2. The method according to claim 1, wherein The construction of the hybrid model includes: The NARX neural network receives the historical output sequence and the historical input sequence through a time window sliding mechanism and generates intermediate features based on a non-linear activation function; The LSTM neural network receives the intermediate features and the dynamic parameters at the current moment, updates the cell state through the input gate, forget gate, and output gate mechanisms, and outputs the predicted value of the state of charge.

3. The method according to claim 1, characterized in that, The collection of the dynamic data is based on the DST standard working condition and the WLTC standard working condition.

4. The method according to claim 1, wherein The training process includes: Preliminarily estimate the state of health of the battery through the NARX neural network to obtain prediction features, where the prediction features are based on the non-linear mapping of the historical output sequence and the historical input sequence; Perform feature splicing on the prediction features and the dynamic parameters at the current moment to obtain an enhanced input vector; Input the enhanced input vector into the LSTM neural network, and use the backpropagation algorithm combined with the Adam optimizer to iteratively train the model until the error index converges to a preset threshold.

5. The method according to claim 1, wherein The method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network further includes: Verify the correlation between the dynamic data during the operation of the lithium-ion battery and the state of health of the battery through the grey relational analysis method, and screen out the optimal feature combination with a correlation higher than the preset threshold; After the model output, perform inverse standardization processing on the prediction result and convert it into a percentage value of the actual state of health of the battery.

6. A state of charge estimation system for a lithium-ion battery based on an improved NARX neural network, characterized in that, The method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network according to any one of claims 1 to 5 includes: A data acquisition module configured to collect dynamic parameters during the operation of a lithium-ion battery, including voltage, current, and temperature; A preprocessing module configured to perform standardization processing on the collected dynamic parameters, eliminate noise interference through filtering, and calibrate the non-linear characteristics of the battery to obtain a standardized data set; A hybrid model module configured to construct a hybrid model combining a NARX neural network and an LSTM neural network in series; A training module configured to input the standardized data set into the hybrid model for training to obtain a trained hybrid model; An output module configured to input the dynamically collected parameters in real time into the trained hybrid model and output the estimated value of the state of charge of the lithium-ion battery.

7. The lithium-ion battery state of charge estimation system based on the improved NARX neural network according to claim 6, characterized in that, The hybrid model module includes: An input layer for receiving the standardized voltage, current, temperature data, and the sequence of the historical state of health of the battery; A hidden layer containing NARX neural network units and LSTM neural network units for feature extraction and time-dependent modeling; The output layer maps to the estimated value of the battery health state through a fully connected network and outputs the error confidence interval.

8. The state of charge estimation system for lithium-ion batteries based on the improved NARX neural network according to claim 6, characterized in that, It also includes a dynamic verification module. The dynamic verification module calculates the estimation error of the battery health state under the current working condition in real time. If the error exceeds the preset threshold, the model parameters are incrementally updated through an online learning algorithm, and the updated model is stored in the local database.

9. A lithium-ion battery state of charge estimation system based on an improved NARX neural network, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute the method for estimating the state of charge of a lithium-ion battery based on an improved NARX neural network according to any one of claims 1 to 5 when running the program instructions.

Images