Battery pack SOC estimation method based on multi-scale deep learning

Abstract

The invention relates to the technical field of lithium battery working condition management, in particular to a battery pack SOC estimation method based on multi-scale deep learning, and the method comprises the steps: collecting the data of a lithium battery, including the continuous voltage, the current signal, the instantaneous voltage, the instantaneous current, the battery temperature, the environment temperature and the internal resistance data of the lithium battery; performing windowing denoising processing on the continuous voltage and current signals to obtain current and voltage windowing processing signals; performing coding preprocessing on the current and voltage windowing processing signal to generate current and voltage windowing code stream data; constructing a multi-scale lightweight deep learning SOC prediction algorithm network which comprises a time sequence network module, a planarization connection layer, a time sequence feature presentation layer, a lightweight convolutional network module, an attribute feature presentation layer, a feature fusion module and a reasoning module; and sending the current and voltage windowing code stream data and the temperature data into a multi-scale lightweight deep learning SOC prediction algorithm network for SOC reasoning prediction.

Description

Technical field:

[0001] The present invention relates to the technical field of lithium battery operating condition management, and specifically to a battery pack SOC estimation method based on multi-scale deep learning, which can improve the accuracy of battery state of charge prediction, thereby improving battery management efficiency and extending the service life of equipment. Background technology:

[0002] With the implementation of various national new energy policies and the rise of the new energy industry, various new energy products have entered thousands of households, but the occurrence of various new energy product accidents has seriously affected people’s lives and property safety. The battery state of charge in the lithium battery battery management system plays a vital role in avoiding overcharging and over-discharging of lithium batteries. Accurately predicting the battery state of charge can effectively avoid overcharging and discharging of lithium batteries. Due to the extreme environment in which lithium batteries are located, the complexity of the environment and the diversity of interference factors make the battery state of charge (SOC) deep learning algorithm face severe challenges in the field of battery management. The goal of the present invention is to accurately predict the state of charge of lithium batteries and shorten the response time of the battery management system while ensuring the accuracy of SOC estimation.

[0003] At present, battery state of charge estimation mostly adopts the ampere-hour integration method and the extended Kalman filter algorithm to perform SOC estimation. The ampere-hour integration method integrates the current in a certain period of time to obtain the change in charge during this period and then calculates the SOC. The current value requires a high-precision current sensor, and its accuracy is affected by temperature and noise. As the integration accumulates, the error will become larger and larger. The extended Kalman filter method uses the equivalent circuit model of the battery to construct the state equation and observation method to infer the battery state of charge. Traditional methods often have the problems of long SOC estimation response time, inaccurate estimation, large amount of calculation, instability and high dependence on the battery model. Summary of the invention:

[0004] In view of the shortcomings and deficiencies in the prior art, the present invention proposes a battery pack SOC estimation method based on multi-scale deep learning, which combines a data preprocessing method and adopts a parallel processing architecture to simultaneously learn the characteristics of lithium battery time series data and attribute data, shortens the response time while ensuring the accuracy of SOC estimation, and improves the generalization ability of the algorithm.

[0005] The present invention is achieved by the following measures:

[0006] A battery pack SOC estimation method based on multi-scale deep learning, characterized by comprising the following steps:

[0007] Step 1: Collect lithium battery data, including continuous voltage, current signal, instantaneous voltage, instantaneous current, battery temperature, ambient temperature, and internal resistance data of the lithium battery, and perform window denoising processing on the continuous voltage and current signals to obtain current and voltage window processing signals; Step 2: Perform encoding preprocessing on the current and voltage window processing signals to generate current and voltage window code stream data;

[0008] Step 3: Construct a multi-scale lightweight deep learning SOC prediction algorithm network, including a time series network module, a flattened connection layer, a time series feature representation layer, a lightweight convolutional network module, an attribute feature representation layer, a feature fusion module, and an inference module;

[0009] Step 4: Send the current and voltage windowed code stream data and temperature data into the multi-scale lightweight deep learning SOC prediction algorithm network for SOC inference prediction.

[0010] In step 1 of the present invention, the timing data and attribute data at different times collected by the data acquisition module are used, and the timing data includes a voltage timing data sequence V=V i (i=1,2...,t) and the current timing data sequence I=I j (j=1,2...,t), where t represents the time dimension, and the attribute data includes the instantaneous voltage, instantaneous current, acquisition time, lithium battery internal temperature, ambient temperature, and internal resistance data at that moment.

[0011] The time series data preprocessing module in the multi-scale lightweight deep learning SOC prediction algorithm network in step 3 of the present invention is used to perform embedding encoding preprocessing on the current and voltage time series window data sequences to complete the shallow representation of the voltage and current data. The time series data network module is composed of two bidirectional long short-term memory layers and a fully connected layer. The processed shallow representation data is sent to the time series data network module to extract the time series features into a time series feature matrix. The bidirectional long short-term memory network in the bidirectional long short-term memory layer is composed of two long short-term memory networks in opposite directions. The long short-term memory network is composed of the input data x at time t. t , cell state C t , temporary cell state C t ', hidden layer state h t 、Forget gate t 、Memory Gate t , output gate O t By forgetting the information in the cell state and memorizing new information, useful information for subsequent moment calculations can be transmitted, useless information is discarded, and the hidden state h is synchronously output at each moment t , where forgetting, memory, and output are determined by the hidden state h at the previous moment t-1 , and the current input xt The calculated forget gate, memory gate, and output gate are controlled, where:

[0012] The forget gate is used to selectively forget feature information. The forget gate calculation formula is: f t =sigmoid(W f ·[x t ;h t-1 ]+b f ), where W f is the forget gate weight, b f is the forget gate bias, [x t ;h t-1 ] is the concatenation of the input at time t and the hidden state at time t-1. The sigmoid activation function is used to map the forget gate data to the interval [0,1]. The memory gate is used to pass the long-term memory information downward. The memory gate calculation formula is: t =sigmoid(W i ·[x t ;h t-1 ]+b i ), W i is the memory gate weight, b i is the memory gate bias, and calculates the current cell state: C t =f t *C t-1 +i t *C t ', C t '=tanh(W c ·[x t ;h t-1 ]+b c ), the output gate fuses the long-term memory with the short-term memory and outputs it. The calculation formula of the output gate is: t =sigmoid(W o ·[x t ;h t-1 ]+b o ), h t =tanh(C t )*O t , where W o is the output gate weight parameter, b o is the output gate bias, h t Represents the final output data at time t.

[0013] The time series data feature representation layer in step 3 of the present invention is composed of a flattening layer and a fully connected layer, and the time series feature matrix extracted by the time series data network module is reconstructed into a dimensionality-reduced feature vector.

[0014] The lightweight convolutional network module described in step 3 of the present invention is composed of a one-dimensional convolutional module and a convolutional residual layer, which is used to extract attribute data features such as instantaneous current, instantaneous voltage, ambient temperature, and internal resistance; the feature fusion module is composed of a fully connected layer and a DropOut layer, which is used to prevent overfitting of the algorithm network.

[0015] The present invention extracts time-frequency information from measurable data such as voltage and current inside the lithium battery, constructs a multi-scale lightweight deep learning algorithm, and synchronously extracts the timing information of the voltage and current of the lithium battery at different times, as well as the attribute information of the internal resistance, instantaneous current, instantaneous voltage, temperature, and ambient temperature of the lithium battery to calculate and estimate the SOC of the lithium battery. Specifically, the PyTorch framework is used to construct a multi-scale lightweight deep learning algorithm model, and the bidirectional long short-term memory network branch is used to fully extract the current and voltage timing characteristics. At the same time, a lightweight convolutional neural network is used to analyze the influence of the lithium battery's internal resistance, instantaneous current, instantaneous voltage, battery temperature, and ambient temperature on the SOC at the current moment, thereby enhancing the algorithm's anchoring and generalization capabilities, improving the SOC calculation accuracy while shortening the calculation time, and improving the real-time capability of the lithium battery system. Description of the drawings:

[0016] Attached Figure 1 It is a schematic diagram of the multi-scale lightweight deep learning network structure in the present invention. Specific implementation method:

[0017] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0018] Embodiment 1:

[0019] This example combines the time series data preprocessing method and the multi-scale lightweight deep learning algorithm to estimate the state of charge of the lithium battery. By extracting the time series features of the internal voltage time series data sequence and the current time series data sequence of the lithium battery, the instantaneous voltage, instantaneous current, acquisition time, lithium battery internal temperature, ambient temperature, and internal resistance data are simultaneously mined to complete the prediction of the lithium battery state of charge, enhance the algorithm's anchoring and generalization capabilities, improve the SOC calculation accuracy while shortening the calculation time, and improve the real-time performance of the lithium battery system;

[0020] This example provides a SOC prediction algorithm based on multi-scale lightweight deep learning. The overall implementation process steps are as follows:

[0021] In the first step, the lithium battery data acquisition module collects continuous voltage and current signals, and simultaneously collects instantaneous voltage, instantaneous current, battery temperature, ambient temperature, and internal resistance data, and performs window denoising processing on the continuous voltage and current signals to obtain current and voltage window processing signals;

[0022] The second step is to perform encoding preprocessing on the current and voltage window processing signals to generate current and voltage window code stream data;

[0023] The third step is to build a multi-scale lightweight deep learning SOC prediction algorithm network structure, construct a timing network module, a flattened connection layer, a timing feature representation layer, a lightweight convolutional network module, an attribute feature representation layer, a feature fusion module and an inference module to form a multi-scale lightweight deep learning algorithm network, and fully explore the data correlation characteristics of lithium batteries;

[0024] The fourth step is to send the current and voltage windowed code stream data and attribute data such as temperature into the multi-scale lightweight deep learning SOC prediction algorithm network for SOC inference prediction;

[0025] In the first step of this example, the timing data and attribute data at different times collected by the data acquisition module are used. The timing data includes the voltage timing data sequence V=V i (i=1,2...,t) and the current timing data sequence I=I j (j=1,2...,t), where t represents the time dimension, and the attribute data includes the instantaneous voltage, instantaneous current, acquisition time, lithium battery internal temperature, ambient temperature, and internal resistance data at that moment;

[0026] The multi-scale lightweight deep learning algorithm network constructed in this example has the following structure: Figure 1 As shown, the multi-scale lightweight deep learning algorithm network includes a time series data preprocessing module, a time series network module, a flattened connection layer, a time series feature representation layer, a lightweight convolutional network module, an attribute feature representation layer, a feature fusion module and an inference module;

[0027] The time series data preprocessing module performs embedding coding preprocessing on the current and voltage time series window data sequences to complete the shallow representation of the voltage and current data;

[0028] The time series data network module is composed of two bidirectional long short-term memory layers and a fully connected layer. The processed shallow representation data is sent to the time series data network module, and the time series features are extracted into a time series feature matrix through the module;

[0029] The bidirectional LSTM network is composed of two LSTM networks in opposite directions. The LSTM network is composed of the input data x at time t. t , cell state C t , temporary cell state C t ', hidden layer state h t 、Forget gate t 、Memory Gate t , output gate O tBy forgetting the information in the cell state and memorizing new information, useful information for subsequent moment calculations can be transmitted, useless information is discarded, and the hidden state h is synchronously output at each moment t , where forgetting, memory, and output are determined by the hidden state h at the previous moment t-1 , and the current input x t The calculated forget gate, memory gate, and output gate are controlled, where the forget gate is used to selectively forget feature information. The forget gate calculation formula is: t =sigmoid(W f ·[x t ;h t-1 ]+b f ), where W f is the forget gate weight, b f is the forget gate bias, [x t ;h t-1 ] is the concatenated data of the input at time t and the hidden layer state at time t-1, and the sigmoid activation function is used to map the forget gate data to the interval range of [0,1];

[0030] The memory gate is used to pass long-term memory information downward. The memory gate calculation formula is: i t =sigmoid(W i ·[x t ;h t-1 ]+b i ), W i is the memory gate weight, b i is the memory gate bias. Calculate the current cell state: C t =f t *C t-1 +i t *C t ', C t '=tanh(W c ·[x t ;h t-1 ]+b c ) The output gate merges the long-term memory with the short-term memory and outputs it. The calculation formula of the output gate is: t =sigmoid(W o ·[x t ;h t-1 ]+b o ), h t =tanh(C t )*O t , where W o is the output gate weight parameter, b o is the output gate bias, h tRepresents the final output data at time t; the time series data feature representation layer is composed of a flattening layer and a fully connected layer, and the time series feature matrix extracted by the time series data network module is reconstructed into a reduced-dimensional feature vector;

[0031] The lightweight convolutional network module is composed of a one-dimensional convolutional module and a convolutional residual layer, which is used to extract attribute data features such as instantaneous current, instantaneous voltage, ambient temperature, and internal resistance. The feature fusion module is composed of a fully connected layer and a DropOut layer, which is used to prevent overfitting of the algorithm network.

[0032] The multi-scale lightweight deep learning SOC prediction algorithm implemented by the present invention is used to perform a fusion verification test on the measurable current, voltage, temperature, and internal resistance data inside the lithium battery, and good verification results are achieved in the data sets with ambient temperatures of 20°C and 25°C. The multi-scale lightweight deep learning SOC prediction algorithm proposed in the present invention can fully analyze the impact of lithium battery internal resistance, current, voltage, battery temperature, and ambient temperature data on SOC.

Claims

1. A battery pack SOC estimation method based on multi-scale deep learning, characterized in that: The following steps are involved: Step 1: Collect lithium battery data, including the lithium battery's continuous voltage, current signal, instantaneous voltage, instantaneous current, battery temperature, ambient temperature, and internal resistance data, and perform window denoising processing on the continuous voltage and current signals to obtain current and voltage window processing signals; Step 2: Encode and pre-process the current and voltage window processing signals to generate current and voltage window code stream data; Step 3: Construct a multi-scale lightweight deep learning SOC prediction algorithm network, including a time series network module, a flattened connection layer, a time series feature representation layer, a lightweight convolutional network module, an attribute feature representation layer, a feature fusion module, and an inference module; Step 4: Send the current and voltage windowed code stream data and temperature data into the multi-scale lightweight deep learning SOC prediction algorithm network for SOC inference prediction.

2. The battery pack SOC estimation method based on multi-scale deep learning according to claim 1, characterized in that: In step 1, the timing data and attribute data at different times collected by the data acquisition module are used. The timing data includes the voltage timing data sequence V=V i (i=1,2...,t) and the current timing data sequence I=I j (j=1,2...,t), where t represents the time dimension, and the attribute data includes the instantaneous voltage, instantaneous current, acquisition time, lithium battery internal temperature, ambient temperature, and internal resistance data at that moment.

3. The battery pack SOC estimation method based on multi-scale deep learning according to claim 1, characterized in that: The time series data preprocessing module in the multi-scale lightweight deep learning SOC prediction algorithm network described in step 3 is used to perform embedding encoding preprocessing on the current and voltage time series window data sequences to complete the shallow representation of the voltage and current data. The time series data network module is composed of two bidirectional long short-term memory layers and a fully connected layer. The processed shallow representation data is sent to the time series data network module to extract the time series features into a time series feature matrix. The bidirectional long short-term memory network in the bidirectional long short-term memory layer is composed of two long short-term memory networks in opposite directions. The long short-term memory network is composed of the input data x at time t. t , cell state C t , temporary cell state C' t , hidden layer state h t 、Forget gate t 、Memory Gate t , output gate O t By forgetting the information in the cell state and memorizing new information, useful information for subsequent moment calculations can be transmitted, useless information is discarded, and the hidden state h is synchronously output at each moment t , where forgetting, memory, and output are determined by the hidden state h at the previous moment t-1 , and the current input x t The calculated forget gate, memory gate, and output gate are controlled, where: The forget gate is used to selectively forget feature information. The forget gate calculation formula is: f t =sigmoid(W f ·[x t ;h t-1 ]+b f ), where W f is the forget gate weight, b f is the forget gate bias, [x t ;h t-1 ] is the concatenated data of the input at time t and the hidden state at time t-1. The sigmoid activation function is used to map the forget gate data to the interval [0,1]. The memory gate is used to pass the long-term memory information downward. The memory gate calculation formula is: i t =sigmoid(W i ·[x t ;h t-1 ]+b i ), W i is the memory gate weight, b i is the memory gate bias, and calculates the current cell state: C t =f t *C t-1 +i t *C t ', C t '=tanh(W c ·[x t ;h t-1 ]+b c ), the output gate fuses the long-term memory with the short-term memory and outputs it. The calculation formula of the output gate is: O t =sigmoid(W o ·[x t ;h t-1 ]+b o ), h t =tanh(C t )*O t , where W o is the output gate weight parameter, b o is the output gate bias, h t Represents the final output data at time t.

4. The battery pack SOC estimation method based on multi-scale deep learning according to claim 1, characterized in that: The time series data feature representation layer in step 3 is composed of a flattening layer and a fully connected layer, and the time series feature matrix extracted by the time series data network module is reconstructed into a dimensionality-reduced feature vector.

5. The battery pack SOC estimation method based on multi-scale deep learning according to claim 1, characterized in that: The lightweight convolutional network module in step 3 is composed of a one-dimensional convolutional module and a convolutional residual layer, which is used to extract attribute data features such as instantaneous current, instantaneous voltage, ambient temperature, and internal resistance; the feature fusion module is composed of a fully connected layer and a DropOut layer, which is used to prevent overfitting of the algorithm network.

Images