A lithium ion battery soc, soe and soh joint estimation method based on a hybrid neural network

Abstract

The application belongs to the technical field of lithium ion battery and power automobile battery management, and discloses a lithium ion battery SOC, SOE and SOH joint estimation method based on a hybrid neural network. The established hybrid neural network takes battery voltage and current as input, and is composed of a customized gate control (CGC) layer and a task regression layer. The CGC layer includes a shared expert network, a specific task expert network and a task gate network, and is used for learning the coupling relationship among SOC, SOE and SOH. The task regression layer adopts three independent neural networks to respectively output SOC, SOE and SOH. A homoscedastic uncertainty model is introduced, and a joint loss function is constructed. After the hybrid neural network model is trained by using a data set, the model is used for simultaneous online estimation of lithium ion battery SOC, SOE and SOH under actual discharge conditions.

Description

Technical Field

[0001] This invention relates to the fields of lithium-ion battery technology and electric vehicle battery management technology, specifically to a method for jointly estimating the SOC, SOE, and SOH of lithium-ion batteries based on a hybrid neural network. Background Technology

[0002] Lithium-ion batteries have become the primary power source for electric vehicles due to their high energy density, high output power, and long charge-discharge life. The state of charge (SOC), state of energy (SOE), and state of health (SOH) parameters of lithium-ion batteries serve as the basis for the decision-making and control of the battery management system. Accurate estimation of these parameters is crucial for the efficient and safe operation and extended lifespan of lithium-ion batteries. However, since SOC, SOE, and SOH are internal characteristics of the battery, they cannot be directly measured and can only be indirectly estimated through measurable parameters such as current and voltage.

[0003] Benefiting from the development of artificial intelligence, machine learning-based state estimation methods for lithium-ion batteries have gradually become a research hotspot in the field. Currently, there are abundant single-state estimation methods for lithium-ion batteries. There is also a large body of research on joint estimation methods for two states, most of which are based on equivalent circuit models, often requiring extensive experiments to determine the parameters of these models. However, during the lifecycle of a lithium-ion battery, there are coupling relationships between SOC, SOE, and SOH. Due to the involvement of multiple states and time scales, the physical relationships between SOC, SOE, and SOH are difficult to represent directly. Specifically, SOC is closely related to current, while SOE is directly related to voltage in addition to current. Furthermore, SOC and SOE change rapidly over time, representing small-scale quantities, while SOH changes slowly with increasing battery charge-discharge cycles, exhibiting large-scale characteristics. Under different SOH conditions, SOC and SOE exhibit different variation patterns. These complex coupling relationships make the development of joint estimation methods for SOC, SOE, and SOH challenging. Summary of the Invention

[0004] To address the problem that existing methods for jointly estimating SOC, SOE, and SOH of lithium-ion batteries suffer from difficulties in directly characterizing the coupling relationships among the three state parameters, and that SOC and SOE change rapidly with discharge while SOH changes slowly with charge-discharge cycles, exhibiting different time-scale characteristics, this invention provides a method for jointly estimating SOC, SOE, and SOH of lithium-ion batteries based on a hybrid neural network. This method takes the voltage and current time series of the lithium-ion battery as input, learns the coupling relationships among the three state parameters through a hybrid neural network composed of a custom gated (CGC) layer and a task regression layer, and employs a task regression network corresponding to different state change characteristics, combined with a joint loss function based on homoscedastic uncertainty, to achieve simultaneous online estimation of SOC, SOE, and SOH of lithium-ion batteries under actual discharge conditions. The method includes the following steps:

[0005] Step 1: Obtain the lithium-ion battery test dataset, which includes the battery's rated capacity and voltage, current, and time data during the test.

[0006] Step 2: Calculate the SOC, SOE, and SOH of the lithium-ion battery based on the lithium-ion battery test dataset to obtain a dataset containing voltage, current, SOC, SOE, and SOH.

[0007] Step 3: Construct a hybrid neural network model, which includes an input layer, a custom gating (CGC) layer, a task regression layer, and three output layers that output SOC, SOE, and SOH respectively. The input layer uses the battery's voltage and current as input data. The CGC layer includes 12 shared expert networks, 3 task-specific expert networks, and 3 task gating networks, used to learn the coupling relationship between SOC, SOE, and SOH of the lithium-ion battery. Each expert network consists of a convolutional neural network and a self-attention mechanism connected in series, and each task gating network consists of a fully connected layer and a Softmax layer. The task regression layer includes three independent regression networks. The SOC regression network and the SOE regression network each consist of a long short-term memory neural network and two fully connected layers, while the SOH regression network consists of four convolutional neural networks and two fully connected layers.

[0008] Step 4: Construct the joint loss function of the hybrid neural network model based on homoscedastic uncertainty, using the mean squared errors of SOC, SOE, and SOH estimates as the losses for the corresponding tasks, and setting the homoscedastic uncertainty parameters corresponding to SOC, SOE, and SOH as learnable parameters to determine the relative weights of the losses for the three tasks; train the hybrid neural network model using the dataset and the joint loss function.

[0009] Step 5: Input the voltage and current time series of the lithium-ion battery to be estimated into the trained hybrid neural network model, and obtain the SOC, SOE and SOH estimation results of the lithium-ion battery to be estimated.

[0010] This invention offers the following advantages: It learns the coupling relationships between SOC, SOE, and SOH by setting up a shared expert network, a task-specific expert network, and a task-gated network within the same model; it ensures that the task regression structures correspond to the rapid changes in SOC and SOE and the slow changes in SOH by setting separate long short-term memory neural network regression structures for SOC and SOE and a convolutional neural network regression structure for SOH; and it determines the relative weights of the three task losses using learnable parameters through a joint loss function based on homoscedastic uncertainty. Therefore, even using only battery voltage and current as input, a single hybrid neural network can simultaneously output the estimation results for SOC, SOE, and SOH. Attached Figure Description

[0011] Figure 1 This is a flowchart of a method for jointly estimating SOC, SOE, and SOH of a lithium-ion battery based on a hybrid neural network, according to an embodiment of the present invention.

[0012] Figure 2 This is a structural diagram of the joint estimation model for SOC, SOE, and SOH of a lithium-ion battery according to an embodiment of the present invention.

[0013] Figure 3 This is the SOC estimation result of the lithium battery in the embodiments of the present invention;

[0014] Figure 4 This is the SOE estimation result of the lithium battery in the embodiment of the present invention;

[0015] Figure 5 This is the SOH estimation result of the lithium battery in the embodiments of the present invention; Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, a joint estimation method for SOC, SOE, and SOH of lithium-ion batteries based on a hybrid neural network specifically includes the following steps:

[0018] Step 1: Obtain the lithium-ion battery test dataset, which includes the battery's rated capacity and voltage, current, and time data during the test. In this embodiment, the data source is the urban road driving cycle UDDS test data of three lithium-ion batteries with a rated capacity of 4.85Ah, numbered W8, W9, and W10, provided by the Stanford Energy Control Laboratory at Stanford University. W8 and W9 batteries serve as reference batteries, and W10 serves as the battery whose state parameters are to be estimated. In this experiment, W8, W9, and W10 batteries were subjected to constant current charging at C / 4, C / 2, and 1C rates, respectively. When the battery voltage reached 4V, constant voltage charging was immediately initiated until the current dropped below 50mA. The C / 4 constant current and constant voltage charging were then designed to bring the battery voltage to 4.2V, corresponding to 100% SOC. During the discharge phase, the battery was discharged from 100% SOC to 80% SOC using a constant current of C / 4. Then, UDDS discharge was performed to discharge the battery from 80% SOC to 20% SOC.

[0019] Step 2: Calculate the SOC, SOE, and SOH of the lithium-ion battery based on the lithium-ion battery test dataset to obtain the complete dataset. In this embodiment, the SOC of the lithium-ion battery is calculated using the ampere-hour integration method:

[0020]

[0021] in, This represents the initial state of charge (SOC) value of a lithium-ion battery. Let be the state of charge value at time t. It is a lithium-ion battery. This represents the maximum usable capacity of the lithium-ion battery during its current cycle.

[0022] In this embodiment, the SOE of the lithium-ion battery is calculated using the following formula:

[0023]

[0024] in, This represents the initial state of energy (SGE) of a lithium-ion battery. The energy state value at time t. and These are lithium-ion batteries and voltage, respectively. This represents the maximum usable energy of a lithium-ion battery during its current cycle.

[0025] In this embodiment, the SOH of the lithium-ion battery is calculated using the following formula:

[0026]

[0027] in, This represents the maximum usable capacity of the lithium-ion battery during its current cycle. This refers to the rated capacity of the lithium-ion battery.

[0028] Step 3: Construct a hybrid neural network model, which includes an input terminal, a custom gated (CGC) layer, a task regression layer, and three output terminals that output SOC, SOE, and SOH respectively. In this embodiment, the input terminal uses the battery's voltage and current as input data, and the three output terminals output the battery's SOC, SOE, and SOH respectively. Figure 2 As shown, the model consists of a Custom Gated (CGC) layer and a task regression layer. The CGC layer comprises 12 shared expert networks, 3 task-specific expert networks, and 3 task-gated networks, used to learn the coupling relationships between the three state parameters of a lithium-ion battery. Each expert network has the same structure, consisting of a convolutional neural network and a self-attention mechanism cascaded together. The gating network consists of a fully connected layer and a softmax layer. The task regression layer utilizes three independent neural networks to regress the output of the CGC layer to SOC, SOE, and SOH. The regression networks for SOC and SOE have the same structure, consisting of a long short-term memory neural network and two fully connected layers. The regression network for SOH consists of four convolutional neural networks and two fully connected layers.

[0029] Step 4: Construct the joint loss function of the hybrid neural network model based on homoscedastic uncertainty. The mean squared errors of the SOC, SOE, and SOH estimates are used as the losses for the corresponding tasks, and the homoscedastic uncertainty parameters corresponding to SOC, SOE, and SOH are set as learnable parameters to determine the relative weights of the losses for the three tasks. In this embodiment, the loss function expression is as follows:

[0030]

[0031] in, , , These represent the mean squared error losses for estimating state parameters SOC, SOE, and SOH, respectively. , , These are learnable parameters. Coefficient terms. These are relative weighting coefficients. This can be considered a regularization term. The hybrid neural network model is trained using the dataset and the joint loss function. In this embodiment, the dataset is normalized, and data samples are generated using a sliding window method with a window length of 400 and a sliding step size of 100. The voltage, current, and corresponding battery SOC, SOE, and SOH after sliding window processing are used as the training dataset. The dataset is divided into a training set and a validation set in a 7:3 ratio, and the neural network model is trained using the Adam optimizer.

[0032] Step 5: Input the voltage and current time series of the lithium-ion battery to be estimated into the trained hybrid neural network model, and simultaneously obtain the estimated SOC, SOE, and SOH of the lithium-ion battery. In this embodiment, the trained model is used to estimate the SOC, SOE, and SOH of the W10 battery for all discharge cycles. The estimated SOC and SOE results when SOH is 1 and 0.93 are as follows: Figure 3 , Figure 4 As shown. The estimated SOH results and experimental results for all discharge cycles are as follows. Figure 5 As shown in Table 1, the root mean square error (RMSE), mean absolute error (MAE), and maximum absolute error (MAX) of the estimated SOC, SOE, and SOH in all the estimated discharge data demonstrate the feasibility and effectiveness of this method for simultaneous estimation of SOC, SOE, and SOH in lithium-ion batteries.

[0033] Table 1. Estimation errors (%) of SOC, SOE, and SOH for lithium-ion batteries

[0034] SOC 0.365 0.289 2.498 SOE 0.327 0.259 2.288 SOH 0.893 0.653 2.553

Claims

1. A method for jointly estimating SOC, SOE, and SOH of lithium-ion batteries based on a hybrid neural network, characterized in that, Includes the following steps: Step 1: Obtain the lithium-ion battery test dataset, which includes the battery's rated capacity and voltage, current, and time data during the test. Step 2: Calculate the state of charge (SOC), state of energy (SOE), and state of health (SOH) of the lithium-ion battery based on the lithium-ion battery test dataset to obtain a dataset containing voltage, current, SOC, SOE, and SOH. Step 3: Construct a hybrid neural network model, which includes an input terminal, a custom gating layer, a task regression layer, and three output terminals that output SOC, SOE, and SOH respectively. The input terminal uses the voltage and current sequence of the lithium-ion battery as input data, and the custom gating layer is used to learn the coupling relationship between the SOC, SOE, and SOH of the lithium-ion battery. The task regression layer uses three independent neural networks to regress the output of the custom gating layer to SOC, SOE, and SOH respectively. The regression neural networks for SOC and SOE each include a long short-term memory neural network and a fully connected layer, while the regression neural network for SOH includes a convolutional neural network and a fully connected layer. Step 4: Construct a joint estimation loss function for SOC, SOE, and SOH based on homoscedasticity uncertainty, and train the hybrid neural network model using the dataset containing voltage, current, SOC, SOE, and SOH and the joint estimation loss function; Step 5: Input the voltage and current sequences of the lithium-ion battery to be estimated into the trained hybrid neural network model, and obtain the SOC, SOE and SOH estimation results of the lithium-ion battery to be estimated.

2. The method for joint estimation of SOC, SOE, and SOH of lithium-ion batteries based on hybrid neural networks according to claim 1, characterized in that: The customized gating layer includes 12 shared expert networks, 3 task-specific expert networks, and 3 task gating networks; Each of the shared expert networks and each of the task-specific expert networks consists of a convolutional neural network and a self-attention mechanism connected in series. Each of the task-gated networks consists of a fully connected layer and a Softmax layer; The regression networks used for SOC and SOE have the same structure, consisting of a long short-term memory neural network and two fully connected layers. The regression network used for SOH consists of four convolutional neural networks and two fully connected layers.

Images