CN117538783A - A lithium-ion battery state-of-charge estimation method based on time-domain fusion converter - Google Patents

Abstract

The invention relates to a lithium ion battery state of charge estimation method based on a time domain fusion converter, which comprises data acquisition and preprocessing; firstly, acquiring battery data in a period of time by utilizing a sliding window technology, then performing multi-step prediction by utilizing a prediction model, simultaneously establishing a correlation model between a battery state and voltage, current and temperature by utilizing a quantile regression method, then comparing a prediction result with actual data, and performing weighted average or statistical learning fusion processing to obtain a comprehensive battery state estimation result; constructing a TFT model; feature extraction and coding; estimating SOC; model training and performance evaluation, calculating an estimation error, and evaluating the accuracy and the robustness of an estimation method. The invention has the characteristics of high accuracy, robustness, real-time performance and expandability.

Description

A method for estimating state of charge of lithium-ion batteries based on time-domain fusion converter

Technical Field

The present invention belongs to the technical field of lithium-ion battery detection, and in particular relates to a lithium-ion battery charge state estimation method based on a time-domain fusion converter.

Background Art

Battery Energy Storage Systems (BESS) are becoming an essential component of distribution networks due to their ability to provide multiple grid services and adapt to the variability and intermittent behavior of renewable generation resources. However, in fixed electrochemical BESS, performance and capacity degradation is inevitable due to aging and poor operating modes. Therefore, the battery state of charge (SOC) is an important parameter that represents the remaining capacity of the battery and needs to be continuously and accurately tracked by the battery management system (BMS) to optimize battery performance and extend battery life. Even for mature BESS technologies, such as lithium-ion batteries, accurate and reliable SOC estimation under a wide range of operating conditions is a challenge and can only rely on indirect methods.

Depending on the battery type, the SOC estimation method is different and can be divided into three types: one is non-model-based or direct measurement, two is model-based method, and three is data-driven method.

Direct measurement includes a variety of measurement methods, such as open circuit voltage (OCV), ampere-hour integration (Ah), and internal resistance measurement. In some cases, the OCV method can provide good accuracy, but it faces severe challenges in non-equilibrium states. Since the electrolyte is evenly distributed inside the battery, it takes a long time for the battery to reach the equilibrium point, so it is difficult to measure OCV in real time. Another method to measure the state of charge of the battery is to use Ah integration, or coulomb counting, which integrates the battery current over time, but due to errors in the determination of initial parameter values, open-loop properties, charge and discharge hysteresis effects, and the sensitivity of the current sensor to noise and drift, this method also lacks accuracy.

The second model-based approach includes state observer and filter-based methods, depending on the underlying control principle. Although the state observer method is more complex to develop, it is suitable for situations with high model uncertainty. On the other hand, the filter-based method first applies a filter to denoise and then estimates the state of the system. Although the filter-based method has reliable real-time accuracy and practicality in the iterative process between the battery model and the closed-loop SOC estimation, its implementation for complex systems may be limited by the potential computational burden of the filter processing.

The three data-driven methods are based on machine learning algorithms and rely only on sensor measurements and operating data, rather than the electrochemical model of the battery. Considering the highly nonlinear interdependence between battery parameters, operating conditions, and SOC, machine learning algorithms can be customized to simulate the relationship between battery parameters, operating conditions, and SOC. There are several types of data-driven techniques, such as support vector regression (SVR) and neural networks (NN). Neural network technology has been widely used in the estimation of lithium-ion battery SOC. The neural network group is further subdivided into ordinary neural networks, deep neural networks, such as recurrent neural network-long short-term memory (RNN-LSTM) models and self-supervised transformer models. Deep neural networks are superior to ordinary neural networks in terms of evaluation indicators, but are more computationally demanding.

At present, the industry generally uses the ampere-hour integration method to estimate the state of charge. This method is simple and easy to implement, but it has strong limitations. The main reasons are: the factors considered in the ampere-hour integration method model are single, lacking feedback correction capabilities, and as the number of battery uses increases, the estimation accuracy drops significantly.

Summary of the invention

The purpose of the present invention is to overcome the above disadvantages and provide a lithium-ion battery state of charge estimation method based on a time domain fusion converter with high accuracy, robustness, real-time performance and scalability.

A method for estimating the state of charge of a lithium-ion battery based on a time domain fusion converter of the present invention comprises the following steps:

Step 1: Data acquisition and preprocessing

Collect the real-time data of the power battery voltage _Vk , current _Ik and temperature _Tk , and digitally process them by the data acquisition module. In the data preprocessing stage, perform noise filtering and outlier processing to ensure the accuracy and consistency of the data;

Step 2: Time domain fusion (also known as time scale fusion)

First, the sliding window technology is used to obtain battery data within a period of time, and then the prediction model is used to perform multi-step predictions. At the same time, the quantile regression method is used to establish a correlation model between the battery state and voltage, current, and temperature. Then, the prediction results are compared with the actual data, and weighted average or statistical learning fusion processing is performed to obtain a comprehensive battery state estimation result;

Step 3: TFT Model Construction

Based on the structure of the Transformer model, Temporal Fusion Transformers (TFT) are constructed. TFT consists of multiple modules, each of which is used to learn feature representations at different time scales. The input of the model includes the data after time domain fusion and other related information.

Step 4: Feature extraction and encoding

The TFT model uses a multi-layer attention mechanism and a self-attention mechanism to extract and encode features of the time-domain fused data. These features include the average current, voltage volatility, temperature change rate, etc. The encoded feature representation can better capture the dynamic characteristics of the battery system.

Step 5: SOC estimation

Using the feature representation learned by the TFT model, the state of charge of the lithium-ion battery is estimated through the fully connected layer and the output layer. The estimated results are the state of charge percentage and the remaining power.

Step 6: Model training and performance evaluation

The TFT model is trained and optimized using historical data. The SOC value estimated by TFT is compared with the actually measured SOC, and the estimation error is calculated. The root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the accuracy and robustness of the estimation method.

The above-mentioned method for estimating the state of charge of a lithium-ion battery based on a time-domain fusion converter, wherein: the data collection described in step 1 includes:

The current data, voltage data and battery surface temperature data of the lithium-ion battery are obtained, and the data are preprocessed. The characteristics of the input TFT are the normalized battery voltage V, current I and temperature T.

The above-mentioned lithium-ion battery state of charge estimation method based on time domain fusion converter, wherein: the time domain fusion described in step 2 is that there is a unique SOC actual value in a given time series data set, and each entity i is associated with a set of static covariates, namely, charging current I _in , lithium-ion battery internal resistance R _in and discharge depth χ _in at each time step t∈[0,T _i ], and the time-dependent input features are subdivided into two categories: one is the lithium-ion battery internal resistance R _in and discharge depth χ _in ; the other is the known input charging current I _in ;

Using quantile forecasting, the definition of the multi-step forecasting problem can be simplified into the following formula:

In the formula, is the quantile value q predicted at the τth step in the future at time t, f _q (.) is the prediction model, y _i,tk:t is the historical target variable, z _i,tk:t is the past observable variable, xi _,tk:t+τ is the a priori known time-varying variable in the future, and s _i is the static covariate.

The above-mentioned lithium-ion battery state of charge estimation method based on time domain fusion converter, wherein: the TFT model described in step 3 realizes the prediction of quantiles through the quantile loss function:

Where Ω is the training data domain containing M samples, W is the weight of TFT, Q is the output quantile set (Q = {0.1, 0.5, 0.9}), and L(Ω, W) is the quantile loss under the average single time series and average prediction point. and There will be almost one positive and one negative, so the formula can be converted to:

The fitting quantile is a target value of 0.9, and the above formula is substituted to obtain:

There are two situations:

like That is, if the model prediction is too small, the loss will increase more;

like That is, the model prediction is too large and the loss increase will be less.

Regularization is performed, which is expressed as:

The main components of TFT are:

(1) Gated Residual Network (GRN), which skips any unused components in the architecture and provides adaptive depth and network complexity to accommodate a wide range of datasets and scenarios;

(2) variable selection network (VSN), which selects relevant input variables at each time step;

(3) Static Covariate Encoder (SCE), which integrates static features into the network and adjusts the temporal dynamics by encoding the context vector;

(4) Temporal processing, which learns long-term and short-term temporal relationships from observed and known time-varying inputs. Sequence-to-sequence layers are used for local processing, while long-term dependencies are captured using a novel interpretable multi-head attention block;

(5) Prediction interval, which determines the range of possible target values at each prediction level through quantile prediction.

The above-mentioned lithium-ion battery state of charge estimation method based on time domain fusion converter, wherein: the feature extraction and encoding described in step 4, first, the gated residual network accepts a main input a and an optional context vector c, and obtains:

GRN _ω (a,c)=LayerNorm(a+GLU _ω (η ₁ )) (6)

η ₁ =W _1,ω η ₂ +b _1,ω (7)

η ₂ =ELU(W _2,ω a+W _3,ω c+b _2,ω ) (8)

Where ELU is the exponential linear unit activation function, is the middle layer, LayerNorm is the standard layer normalization, ω is the exponent representing the weight, when W _2,ω a+W _3,ω c+b _2,ω ＞＞0, ELU activation will act as an identity function; when W _2,ω a+W _3,ω c+b _2,ω ＜＜0, ELU activation will produce a constant output, resulting in linear layer behavior; a component gating layer based on gated linear units (GLUs) is used to provide the flexibility to suppress the unwanted parts of the architecture for any given dataset; let As input, GLU can be expressed as:

GLU _ω (γ)＝σ(W _4,ω γ+b _4,ω )⊙(W _5,ω γ+b _5,ω ) (9)

Where σ(.) is the sigmoid activation function, is the weight, is the bias, ⊙ is the element-wise Hadamard product, dmodel is the hidden state size, and GLU allows TFT to control the contribution of GRN to the original input. In the absence of a context vector, GRN simply treats the context input as zero, c = 0 in (5); during training, dropout is applied before the gating layer and the normalization layer, η ₁ in (3);

Entity embeddings are used as feature representations for categorical variables. Each input variable is converted into a (dmodel)-dimensional vector that matches the dimension of subsequent layers for skip connections. Independent variable selection networks are used for all static, past, and future inputs. represents the transformed input of the j-th variable at time t, is the flattened vector of all past inputs at time T. The variable selection weights are generated by passing the GRN input Ξt and the external context vector _cs , followed by a Softmax layer:

In the formula, is the vector of variable selection weights, c _s is obtained from the static covariate encoder;

At each time step, an additional layer of nonlinear processing is added Input into its own GRN, represented as:

In the formula, is the processed eigenvector of variable j. Note that each variable has its own The weights are shared across all time steps t. Then, the processed features are weighted by their variable selection weights and merged, expressed as:

In the formula, is the jth element of the vector v _χt ;

A separate GRN encoder is used to generate four different context vectors _cs , _ce , _cc , and _ch . These context vectors are connected to various positions of the temporal fusion decoder, including:

(1) Time variable selection (c _s );

(2) Local processing of temporal features (c _c , c _h );

(3) enriching temporal features (c _e ) with static information;

The attention mechanism is based on the key and query The relationship between The scale is as follows:

Attention(Q,K,V)＝A(Q,K)V (13)

Where A(.) is a normalization function. A common choice is the scaled dot product attention:

In order to improve the learning ability of the standard attention mechanism, multi-head attention is proposed, using different heads for different representation subspaces, expressed as:

In the formula, are the weights of the key, query, and value associated with the header, and is the linear combination of the outputs from all heads H _h connected;

Modify the multi-head attention so that each head shares the value and all heads are summed, expressed as:

In the formula, is the value weight shared by all heads, Used for the final linear mapping.

The above-mentioned method for estimating the state of charge of a lithium-ion battery based on a time-domain fusion converter, wherein: the SOC estimation described in step 5 uses the available SOC value to estimate the capacity value of the battery during charging and discharging, expressed as:

E _dk ＝∫I _k V _k dt (20)

Where C _k is the battery capacity of the kth cycle, _Edk is the discharge energy of the same cycle, SOC _kmax and SOC _kmin are the maximum and minimum SOC values of the cycle respectively;

In the constructed TFT model, the battery voltage V, current I and temperature T are taken as future inputs, the charging current I _in is the static covariate input, and the lithium-ion battery internal resistance R _in and the discharge depth χ _in , which are measured at each step, are taken as unknown inputs and finally input to SOC.

The above-mentioned lithium-ion battery state of charge estimation method based on time domain fusion converter, wherein: the root mean square error (RMSE) and mean absolute error (MAE) described in step 6 are expressed as:

Where N is the total number of training samples, SOC _k is the SOC estimated by the model at time step k, is the actual SOC value at time step k.

Compared with the existing methods, the present invention has obvious beneficial effects. From the above technical solutions, it can be seen that the present invention adopts multi-step prediction and sets quantile regression. Since the weight is 9:1, during training, the model will tend to predict larger numbers more and more, so that the Loss decreases faster, and the entire fitting hyperplane of the model will move upward, so that the 90th percentile value of the target variable can be well fitted. Therefore, it has the following advantages:

(1) High accuracy: Considering that every time a lithium-ion battery is charged and discharged, and in different time periods of a single charge, the internal characteristics of the battery will undergo subtle changes, such as internal resistance and discharge depth, which will affect the service life of the lithium-ion battery by affecting the internal characteristics of the lithium-ion battery, resulting in subtle changes in the results of each estimation of the lithium-ion battery, thereby affecting the estimation accuracy of the lithium-ion battery. Therefore, the present invention adopts time-sharing and partitioning to estimate the lithium-ion battery and sets quantile estimation, thereby improving the estimation accuracy;

(2) Robustness: This method has good adaptability and robustness to different types of lithium-ion batteries and working conditions, and can reliably estimate the state of charge in a variety of application scenarios.

(3) Real-time performance: Due to the use of an efficient TFT network structure, this method can achieve fast state of charge estimation in applications with high real-time requirements.

(4) Scalability: This method can be integrated with other battery management algorithms and systems to provide a more complete battery management solution.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the present invention;

Figure 2 TFT model structure diagram;

FIG3 is a graph showing the SOC estimation results of a lithium-ion battery based on TFT in a test example;

FIG4 is a graph showing the SOC estimation error results of a TFT-based lithium-ion battery in an experimental example.

DETAILED DESCRIPTION

Example:

As shown in FIG1 , a method for estimating the state of charge of a lithium-ion battery based on a time domain fusion converter includes the following steps:

Step 1: Data acquisition and preprocessing

Real-time data such as power battery voltage _Vk , current _Ik and temperature _Tk are collected and digitized by the data acquisition module. In the data preprocessing stage, noise filtering and outlier processing are performed to ensure the accuracy and consistency of the data.

Specifically,

The three characteristic data of current, voltage and battery surface temperature in the battery sample data P are Z-Score standardized, and the conversion function is:

Where μ is the mean of the sample data (current, voltage, and battery surface temperature, respectively), and σ is the standard deviation of the sample data.

The original two-dimensional feature data format of the battery sample data is constructed into a three-dimensional data format for use as the input of the network. The three-dimensional data format is [time_step, input_size, batch_size], where time_step represents the time step. For example, if the data of the first 6 moments are used to predict the fourth moment, then time_step = 6. Input_size represents the input feature dimension. In this embodiment, input_size = 6, which represents three dimensions: current, voltage, battery surface temperature, internal resistance, charging current, and discharge depth. Batch_size represents the number of samples, that is, the number of samples put into the network.

Step 2: Time domain fusion (also known as time scale fusion)

First, the sliding window technology is used to obtain battery data within a period of time, and then the prediction model is used for multi-step prediction. At the same time, the quantile regression method is used to establish the correlation model between the battery state and the voltage, current, and temperature. Next, the prediction results are compared with the actual data, and the weighted average or statistical learning fusion processing is performed to obtain a comprehensive battery state estimation result. Time scale fusion mainly performs multi-step prediction, sets quantile regression, and fusion data processing.

Specifically, the time scale fusion is that there is a unique SOC actual value in a given time series data set, and each entity i is associated with a set of static covariates, namely, the charging current I _in , the lithium-ion battery internal resistance R _in and the discharge depth χ in, at each time step t∈[0,T _i ]. The time-dependent input features are subdivided into two categories: one is the lithium-ion battery internal resistance R _in and the discharge depth χ _in , which can only be measured at each step and are unknown in advance; the other is the known input charging current I _in, which can be determined in advance (can be directly measured).

The definition of the quantile forecast for the multi-step forecast problem can be simplified into the following formula:

In the formula, is the quantile value q predicted at the τth step in the future at time t, f _q (.) is the prediction model, y _i,tk:t is the historical target variable, z _i,tk:t is the past observable, xi _,tk:t+τ is the a priori known time-varying variable in the future, and s _i is the static covariate.

Step 3: TFT Model Construction

Based on the structure of the Transformer model, Temporal Fusion Transformers (TFT) is constructed. TFT consists of multiple modules, each of which is used to learn feature representations at different time scales. The input of the model includes the data after time domain fusion and other related information. The model structure is shown in Figure 2.

Specifically, the TFT model realizes quantile prediction by designing a quantile loss function:

Where Ω is the training data domain containing M samples, W is the weight of TFT, Q is the output quantile set (Q = {0.1, 0.5, 0.9}), and L(Ω, W) is the quantile loss under the average single time series and average prediction point. and There will be almost one positive and one negative, so the formula can be converted to:

The fitting quantile is a target value of 0.9, and the above formula is substituted to obtain:

There are two situations:

like That is, if the model prediction is too small, the loss will increase more;

like That is, the model prediction is too large and the loss increase will be less.

In order to avoid the problem of inconsistent prediction dimensions at different prediction points, regularization processing is performed, which is expressed as:

The TFT model is designed to use canonical components to efficiently build feature representations for each input type (i.e., static, known, observed inputs) to achieve high predictive performance on a wide range of problems. The main components of TFT are:

(1) Gated Residual Network (GRN), which skips any unused components in the architecture and provides adaptive depth and network complexity to accommodate a wide range of datasets and scenarios;

(2) variable selection network (VSN), which selects relevant input variables at each time step;

(3) Static Covariate Encoder (SCE), which integrates static features into the network and adjusts the temporal dynamics by encoding the context vector;

(4) Temporal processing, which learns long-term and short-term temporal relationships from observed and known time-varying inputs. Sequence-to-sequence layers are used for local processing, while long-term dependencies are captured using a novel interpretable multi-head attention block;

(5) Prediction interval, which determines the range of possible target values at each prediction level through quantile prediction.

Step 4: Model training, feature extraction and encoding

The TFT model uses a multi-layer attention mechanism and a self-attention mechanism to extract and encode features of the time-domain fused data. These features include the average current, voltage volatility, temperature change rate, etc. The encoded feature representation can better capture the dynamic characteristics of the battery system.

Specifically, the feature extraction and encoding is first a gated residual network accepting a main input a and an optional context vector c to obtain:

GRN _ω (a,c)=LayerNorm(a+GLU _ω (η ₁ )) (7)

η ₁ =W _1,ω η ₂ +b _1,ω (8)

η ₂ =ELU(W _2,ω a+W _3,ω c+b _2,ω ) (9)

Where ELU is the exponential linear unit activation function, is the middle layer, LayerNorm is the standard layer normalization, and ω is the exponent representing the weight. When W _2,ω a+W _3,ω c+b _2,ω ＞＞0, the ELU activation will act as an identity function; when W _2,ω a+W _3,ω c+b _2,ω ＜＜0, the ELU activation will produce a constant output, resulting in linear layer behavior. A component gating layer based on gated linear units (GLUs) is used to provide the flexibility to suppress unwanted parts of the architecture for any given dataset. Let As input, GLU can be expressed as:

GLU _ω (γ)＝σ(W _4,ω γ+b _4,ω )⊙(W _5,ω γ+b _5,ω ) (10)

Where σ(.) is the sigmoid activation function, is the weight, is the bias, ⊙ is the element-wise Hadamard product, and dmodel is the hidden state size (common in TFT). GLU allows TFT to control how much the GRN contributes to the original input - if necessary, this layer may be skipped entirely, as GLU outputs are likely to be close to 0 to suppress nonlinear contributions. In the absence of a context vector, the GRN simply treats the context input as zero, i.e. c = 0 in (5). During training, dropout is applied before the gating layer and the normalization layer, i.e. η ₁ in (3).

We use entity embeddings as feature representations for categorical variables and linear transformations for continuous variables — transforming each input variable into a (dmodel)-dimensional vector that matches the dimensions of subsequent layers for skip connections. Separate variable selection networks are used for all static, past, and future inputs. represents the transformed input of the j-th variable at time t, is the flattened vector of all past inputs at time T. The variable selection weights are generated by passing the GRN input Ξt and the external context vector _cs , followed by a Softmax layer:

In the formula, is the vector of variable selection weights, and _cs is obtained from the static covariate encoder.

At each time step, an additional layer of nonlinear processing is added Input into its own GRN, represented as:

In the formula, is the processed eigenvector of variable j. Note that each variable has its own The weights are shared across all time steps t. Then, the processed features are weighted by their variable selection weights and merged, expressed as:

In the formula, is the j-th element of the vector v _χt .

Four different context vectors _cs , _ce , _cc , and _ch are generated using separate GRN encoders. These context vectors are connected to various locations in the temporal fusion decoder, where static variables play an important role in the processing. Specifically, this includes:

(1) Time variable selection (c _s );

(2) Local processing of temporal features (c _c , c _h );

(3) Enriching temporal features ( _ce ) with static information.

For example, taking ζ as the output of the static variable selection network, the context of temporal variable selection will be as follows to encode.

A self-attention mechanism is adopted to learn long-term relationships between different time steps. The multi-head attention in the transformer-based architecture is modified to enhance interpretability. In general, the attention mechanism is based on the key and query The relationship between The scale is as follows:

Attention(Q,K,V)＝A(Q,K)V (14)

Where A(.) is a normalization function. A common choice is the scaled dot product attention:

In order to improve the learning ability of the standard attention mechanism, multi-head attention is proposed, using different heads for different representation subspaces, expressed as:

In the formula, are the weights of the key, query, and value associated with the header, and is the linear combination of the outputs from all heads H _h connected together.

Considering that each head uses different values, the attention weight alone cannot indicate the importance of a particular feature. Therefore, the multi-head attention is modified so that each head shares the value and all heads are summed, expressed as:

In the formula, is the value weight shared by all heads, For the final linear mapping, it can be seen from Equation (15) that each head can learn different temporal patterns while focusing on a common set of input features, which can be interpreted as a simple collection of attention weights into the merge matrix Compared with A(Q,K) in formula (11), The representation capacity is increased in an efficient manner.

Step 5: SOC estimation

The feature representation learned by the TFT model is used to estimate the state of charge of the lithium-ion battery through the fully connected layer and the output layer. The estimated results are the state of charge percentage and the remaining power.

Specifically, the SOC estimation uses the available SOC value to estimate the capacity value of the battery during charging and discharging, which is expressed as:

E _dk ＝∫I _k V _k dt (21)

Where C _k is the battery capacity of the kth cycle, _Edk is the discharge energy of the same cycle, SOC _kmax and SOC _kmin are the maximum and minimum SOC values of the cycle, respectively.

In the constructed TFT model, the battery voltage V, current I and temperature T are taken as future inputs, the charging current I _in is input as a static covariate, and the lithium-ion battery internal resistance R _in and discharge depth χ _in , which can only be measured at each step, are taken as unknown inputs and finally input into SOC.

Step 6: Model training and performance evaluation

In order to improve the accuracy of the estimation, the TFT model is trained and optimized using historical data. The SOC value estimated by TFT is compared with the actually measured SOC, and the estimation error is calculated. The root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the accuracy and robustness of the estimation method.

Specifically, the root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the model performance, expressed as:

Where N is the total number of training samples, SOC _k is the SOC estimated by the model at time step k, is the actual SOC value at time step k.

Test example

For a lithium-ion battery, the nominal voltage is 3.7V, the charging cut-off voltage is 4.2V, the rated capacity is 3000mAh, and the charging depth is 20%, that is, the remaining power is 600mAh. The voltage, current, and internal resistance of the lithium-ion battery at different surface temperatures are measured, see Table 1.

Table 1 Voltage, current, internal resistance at different temperatures

Surface temperature(℃) Current (mA) Voltage (V) Discharge depth (%) Internal resistance(mΩ) 25℃ 80 3.432 20 8.013 35℃ 79 3.369 20 8.256 45℃ 75 3.306 20 8.333

As can be seen from the table, the internal resistance of the battery at different temperatures is subtly different, and the higher the temperature, the higher the internal resistance, and the battery usage time is also different.

These variables are input into the constructed TFT model, and finally the SOC value is output, as shown in Figure 3. The estimated error is shown in Figure 4. It can be seen that the result obtained by inputting the battery data in time and segment is more accurate.

As described above, TFT can better realize a more accurate estimation of the state of charge of the lithium-ion battery. This is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims (7)

1. A method for estimating the state of charge of a lithium-ion battery based on a time domain fusion converter, comprising the following steps: Step 1: Data acquisition and preprocessing Collect the real-time data of the power battery voltage _Vk , current _Ik and temperature _Tk , and digitally process them by the data acquisition module. In the data preprocessing stage, perform noise filtering and outlier processing to ensure the accuracy and consistency of the data; Step 2: Time domain fusion (also known as time scale fusion) First, the sliding window technology is used to obtain battery data within a period of time, and then the prediction model is used to perform multi-step predictions. At the same time, the quantile regression method is used to establish a correlation model between the battery state and voltage, current, and temperature. Then, the prediction results are compared with the actual data, and weighted average or statistical learning fusion processing is performed to obtain a comprehensive battery state estimation result; Step 3: TFT Model Construction Based on the structure of the Transformer model, Temporal Fusion Transformers (TFT) are constructed. TFT consists of multiple modules, each of which is used to learn feature representations at different time scales. The input of the model includes the data after time domain fusion and other related information. Step 4: Feature extraction and encoding The TFT model uses a multi-layer attention mechanism and a self-attention mechanism to extract and encode features of the time-domain fused data. These features include the average current, voltage volatility, temperature change rate, etc. The encoded feature representation can better capture the dynamic characteristics of the battery system. Step 5: SOC estimation Using the feature representation learned by the TFT model, the state of charge of the lithium-ion battery is estimated through the fully connected layer and the output layer. The estimated results are the state of charge percentage and the remaining power. Step 6: Model training and performance evaluation The TFT model is trained and optimized using historical data. The SOC value estimated by TFT is compared with the actual measured SOC, and the estimation error is calculated. The root mean square error and mean absolute error are used to evaluate the accuracy and robustness of the estimation method. 2. A lithium-ion battery state of charge estimation method based on a time domain fusion converter as claimed in claim 1, wherein: the data collection described in step 1 includes: The current data, voltage data and battery surface temperature data of the lithium-ion battery are obtained, and the data are preprocessed. The characteristics of the input TFT are the normalized battery voltage V, current I and temperature T. 3. A lithium-ion battery state of charge estimation method based on a time-domain fusion converter as claimed in claim 1, wherein: the time-domain fusion described in step 2 is that there is a unique SOC actual value in a given time series data set, and each entity i is associated with a set of static covariates, namely, charging current I _in , lithium-ion battery internal resistance R _in and discharge depth χ _in at each time step t∈[0,T _i ], and the time-dependent input features are subdivided into two categories: one is the lithium-ion battery internal resistance R _in and discharge depth χ _in ; the other is the known input charging current I _in ; Using quantile forecasting, the definition of the multi-step forecasting problem can be simplified into the following formula: In the formula, is the quantile value q predicted at the τth step in the future at time t, f _q (.) is the prediction model, y _i,tk:t is the historical target variable, z _i,tk:t is the past observable variable, xi _,tk:t+τ is the a priori known time-varying variable in the future, and s _i is the static covariate. 4. A lithium-ion battery state of charge estimation method based on a time domain fusion converter as claimed in claim 1, wherein: the TFT model in step 3 realizes the prediction of quantiles through a quantile loss function: Where Ω is the training data domain containing M samples, W is the weight of TFT, Q is the output quantile set (Q = {0.1, 0.5, 0.9}), and L(Ω, W) is the quantile loss under the average single time series and average prediction point. and There will be almost one positive and one negative, so the formula can be converted to: The fitting quantile is a target value of 0.9, and the above formula is substituted to obtain: There are two situations: like That is, if the model prediction is too small, the loss will increase more; like That is, if the model prediction is too large, the increase in Loss will be less; Regularization is performed, which is expressed as: The main components of TFT are: (1) Gated Residual Network (GRN), which skips any unused components in the architecture and provides adaptive depth and network complexity to accommodate a wide range of datasets and scenarios; (2) variable selection network (VSN), which selects relevant input variables at each time step; (3) Static Covariate Encoder (SCE), which integrates static features into the network and adjusts the temporal dynamics by encoding the context vector; (4) Temporal processing, which learns long-term and short-term temporal relationships from observed and known time-varying inputs. Sequence-to-sequence layers are used for local processing, while long-term dependencies are captured using a novel interpretable multi-head attention block; (5) Prediction interval, which determines the range of possible target values at each prediction level through quantile prediction. 5. A lithium-ion battery state of charge estimation method based on a time domain fusion converter as claimed in claim 1, wherein: the feature extraction and encoding described in step 4 is firstly that the gated residual network accepts a main input a and an optional context vector c to obtain: GRN _ω (a,c)=LayerNorm(a+GLU _ω (η ₁ )) (6) η ₁ =W _1,ω η ₂ +b _1,ω (7) η ₂ =ELU(W _2,ω a+W _3,ω c+b _2,ω ) (8) Where ELU is the exponential linear unit activation function, is the middle layer, LayerNorm is the standard layer normalization, ω is the exponent representing the weight, when W _2,ω a+W _3,ω c+b _2,ω ＞＞0, ELU activation will act as an identity function; when W _2,ω a+W _3,ω c+b _2,ω ＜＜0, ELU activation will produce a constant output, resulting in linear layer behavior; a component gating layer based on gated linear units (GLUs) is used to provide the flexibility to suppress the unwanted parts of the architecture for any given dataset; let As input, GLU can be expressed as: GLU _ω (γ)＝σ(W _4,ω γ+b _4,ω )⊙(W _5,ω γ+b _5,ω ) (9) Where σ(.) is the sigmoid activation function, is the weight, is the bias, ⊙ is the element-wise Hadamard product, dmodel is the hidden state size, and GLU allows TFT to control the contribution of GRN to the original input. In the absence of a context vector, GRN simply treats the context input as zero, c = 0 in (5); during training, dropout is applied before the gating layer and the normalization layer, η ₁ in (3); Entity embeddings are used as feature representations for categorical variables. Each input variable is converted into a (dmodel)-dimensional vector that matches the dimension of subsequent layers for skip connections. Independent variable selection networks are used for all static, past, and future inputs. represents the transformed input of the j-th variable at time t, is a flattened vector of all past inputs at time T, and the variable selection weights are generated by passing the GRN input Ξt and the external context vector _cs , followed by a Softmax layer: In the formula, is the vector of variable selection weights, c _s is obtained from the static covariate encoder; At each time step, an additional layer of nonlinear processing is added Input into its own GRN, represented as: In the formula, is the processed eigenvector of variable j. Note that each variable has its own The weights are shared across all time steps t, and then the processed features are weighted by their variable selection weights and merged, expressed as: In the formula, is the jth element of the vector v _χt ; A separate GRN encoder is used to generate four different context vectors _cs , _ce , _cc , and _ch . These context vectors are connected to various positions of the temporal fusion decoder, including: (1) Time variable selection (c _s ); (2) Local processing of temporal features (c _c , c _h ); (3) enriching temporal features (c _e ) with static information; The attention mechanism is based on the key and query The relationship between The scale is as follows: Attention(Q,K,V)＝A(Q,K)V (13) Where A(.) is a normalization function, a common choice is the scaled dot-product attention: In order to improve the learning ability of the standard attention mechanism, multi-head attention is proposed, using different heads for different representation subspaces, expressed as: In the formula, are the weights of the key, query, and value associated with the header, and is the linear combination of the outputs from all heads H _h connected; Modify the multi-head attention so that each head shares the value and all heads are summed, expressed as: In the formula, is the value weight shared by all heads, Used for the final linear mapping. 6. A lithium-ion battery state of charge estimation method based on a time domain fusion converter as claimed in claim 1, wherein: the SOC estimation described in step 5 uses the available SOC value to estimate the capacity value of the battery during charging and discharging, expressed as: E _dk ＝∫I _k V _k dt (20) Where C _k is the battery capacity of the kth cycle, _Edk is the discharge energy of the same cycle, SOC _kmax and SOC _kmin are the maximum and minimum SOC values of the cycle respectively; In the constructed TFT model, the battery voltage V, current I and temperature T are taken as future inputs, the charging current I _in is the static covariate input, and the lithium-ion battery internal resistance R _in and the discharge depth χ _in , which are measured at each step, are taken as unknown inputs and finally input to SOC. 7. A lithium-ion battery state of charge estimation method based on a time domain fusion converter as claimed in claim 1, wherein: the root mean square error (RMSE) and mean absolute error (MAE) described in step 6 are expressed as: Where N is the total number of training samples, SOC _k is the SOC estimated by the model at time step k, is the actual SOC value at time step k.