Equalization method for lithium battery pack of underwater vehicle by using neuron network model

Abstract

The application relates to a method for balancing a lithium battery pack of an underwater vehicle by adopting a neuron network model and a system for realizing the balancing. The balancing strategy with a battery SOC as an index and the balancing strategy with a voltage as an index are fused with each other by the method for establishing the neuron network model. The neuron network model which is fully trained dynamically selects the balancing index and strategy according to the data in the battery operation, and the efficiency of the balancing strategy is effectively improved according to the neuron network model output. The balancing strategy with the battery SOC as the index or the balancing strategy with the voltage as the index is used to balance the lithium battery pack of the underwater vehicle, so that the inconsistency of the whole battery pack is eliminated, overcharging or overdischarging of the battery is avoided, the service life of the battery is prolonged, and the lithium battery pack of the underwater vehicle is prevented from being damaged due to self-ignition caused by thermal runaway of the battery, thereby avoiding great losses.

Description

Technical Field

[0001] This invention belongs to the field of energy management strategies for lithium battery packs in underwater vehicles, and relates to a method for balancing lithium battery packs in underwater vehicles using a neural network model. Background Technology

[0002] Underwater vehicles are a crucial platform for underwater transportation technology and have become a focus of attention worldwide. Lithium-ion batteries, with their high voltage, large capacity, low self-discharge, and long lifespan, have become the primary power source for underwater vehicles. However, it's impossible to make every battery exactly the same during manufacturing. Differences in quantity and the varying physical and chemical properties of individual batteries contribute to an overall imbalance in the battery pack. As batteries are used, this imbalance worsens, easily leading to overcharging and over-discharging of individual cells, thus reducing the overall lifespan of the battery pack. Therefore, accurate battery balancing is essential to improve battery pack lifespan and reliability.

[0003] Currently, in battery balancing variables, voltage is widely used as the balancing metric because it effectively reflects the battery's state of charge (SOC) and is relatively easy to obtain. However, voltage is affected by current, leading to inaccurate balancing metrics. Therefore, a balancing strategy based on the difference in SOC has been proposed to avoid the influence of current. However, during the initial stage of battery use when SOC is high, and during the later stage when SOC is low, battery voltage changes are particularly drastic. This means that even a small change in SOC can cause a significant change in battery voltage during these two stages. Therefore, when using SOC as the balancing variable, even if the SOCs of two battery cells are balanced, their voltages may still differ.

[0004] Therefore, simply using voltage as the balancing indicator leads to inaccurate balancing results as voltage is affected by current. Conversely, using a balancing strategy based solely on the difference in battery SOC (State of Charge) can result in voltage differences even when the SOC of two battery cells is balanced, especially in the early stages of battery use when the SOC is high or low. These voltage differences increase inconsistencies between batteries, leading to overcharging or over-discharging, severely reducing battery lifespan, and potentially causing thermal runaway and spontaneous combustion. Once thermal runaway occurs, it triggers a chain reaction, resulting in the complete combustion of reactants. Since underwater vehicles operate in enclosed environments with limited internal space, heat accumulates rapidly, increasing the risk of explosions, malfunctions, and even casualties. Therefore, lithium battery pack balancing is crucial for the safety of underwater vehicles. Summary of the Invention

[0005] Technical problems to be solved To overcome the shortcomings of existing technologies, this invention proposes a method for balancing lithium battery packs in underwater vehicles using a neural network model. The method identifies the current state of the battery's SOC and voltage values ​​through the neural network model. Between two balancing strategies—one that uses voltage as the balancing index and the other that uses the difference in battery SOC—the method selects the balancing strategy that best suits the current state of the battery to balance the lithium battery pack in the underwater vehicle, which can effectively eliminate inconsistencies.

[0006] Therefore, the present invention aims to provide an effective balancing strategy for lithium battery packs under different usage conditions and for addressing the specific causes of internal inconsistencies, thereby overcoming the shortcomings of existing balancing variables that are unsuitable for different battery usage conditions.

[0007] Another objective of the invention is to provide a system for implementing the equalization strategy and a suitable equalization circuit.

[0008] Technical solution A method for balancing lithium battery packs in underwater vehicles using a neural network model includes the following steps: Step 1: The lithium battery pack was tested using HPPC experiments to obtain the SOC-OCV curve; Step 2: Construct a multi-layer neural network model. The first layer is the input layer with two neuron variables. The second and third layers are hidden layers. The first hidden layer has 10 neurons, and the second hidden layer has 15 neurons. The fourth layer is the output layer with two neuron variables. All neurons in each layer are fully connected. Each layer is set with a ReLU activation function and a CrossEntropyLoss loss function. The backpropagation optimizer of the neural network model uses Adam, which dynamically adjusts the learning rate of each parameter using the first and second moment estimates of the gradient. Step 3: Assign identification tags to the current SOC and voltage values ​​of the battery, where the tag for the balancing strategy effect based on battery SOC is 1, and the tag for the balancing strategy effect based on battery voltage is 2. Step 3.1: In the first hidden layer, the two input variables SOC and voltage value are fully connected with the weight parameter matrix W1 and the bias vector b1 to obtain 10 feature values ​​for each input value; The fully connected operation of the first hidden layer is: Where: a represents the feature of the first hidden layer. W1 is the first hidden layer 2. The weight parameter matrix is ​​10, and b1 is the first hidden layer 1. The bias vector is 10; f is the ReLU activation function, x is the variable input to the activation function f; n1 represents the number of variables in the second layer that are input from the first layer. Step 3.2: The output value of the first hidden layer is non-linearized by the ReLU function. In the second hidden layer, a fully connected operation is performed with the weight parameter matrix W2 and the bias vector b2 to obtain 15 feature values ​​for each input value. The fully connected operation of the second hidden layer is: Where: c represents the second hidden layer feature. W2 is the second hidden layer 10. The weight parameter matrix is ​​15, b2 is the bias vector of the second hidden layer; n2 represents the number of variables in the third layer that are input to the second layer. Step 3.3: The output value of the second hidden layer is non-linearized using the ReLU function, and then fully connected with the weight parameter matrix W3 and the bias vector b3 to obtain two probability values ​​representing the two equilibrium strategies output by the neural network model. ; The fully connected operation with the weight parameter matrix W3 and the bias vector b3 is as follows: in: The probability values ​​of the equilibrium strategy, where: k is the identification label in step 3, with a value of 1 or 2; W3 is the output layer 15. The weight parameter matrix of layer 2, and b3 is the output layer 1. 2. Bias vector; n3 represents the number of variables input from the third layer to the fourth layer; Step 4: Select the equilibrium strategy based on the probability values ​​of the two equilibrium strategies output by the neural network model. When the probability of the equalization strategy output by the neural network model using voltage as an indicator... The probability is greater than that of an equilibrium strategy based on battery SOC. When the voltage is high, a balancing strategy based on voltage is selected; otherwise, a balancing strategy based on battery SOC is selected.

[0009] Step 1 involves using HPPC experiments to test the lithium battery pack and obtaining the SOC-OCV curve as follows: S11. Charge the battery at a constant current and constant voltage in an environment of 25°C until the battery SOC is 100%; after an interval of 3 hours, discharge it at a constant current of 1C for 30 seconds. S12. After a 40-second interval, charge at a constant current rate of 1C for 30 seconds, let stand for 30 minutes, and finally discharge at a constant current rate of 1 / 3C for 18 minutes, and let stand for 3 hours. S13. Repeat S11 and S12 to obtain the SOC-OCV curve; S14. Using the least squares method, an 8th-order polynomial is fitted to the curve to obtain the relationship between OCV and SOC.

[0010] The balancing strategy steps using voltage as an indicator in step 4 are as follows: S21. Collect the voltage difference between two individual battery cells; S22. Amplify the collected voltage by at least 100 times; S23. When the voltage between two individual battery cells is inconsistent, activate the equalization circuit; S24. After the equalization between individual battery cells is completed, the equalization between battery packs is started by treating the entire battery pack as a single battery cell until the inconsistency of the entire battery pack is eliminated.

[0011] The balancing strategy steps in step 4, which use battery SOC as an indicator, are as follows: S31. Based on the relationship between SOC and voltage, the voltage difference between two individual battery cells is converted into the difference in SOC; S32. The ΔSOC′ generated by the battery during the equalization time is calculated using the ampere-hour integration method; S33. If ΔSOC = ΔSOC′, the two battery cells are considered to be in a balanced state. If they are not in a balanced state, the equalization circuit is activated. When the difference ΔSOC between the two individual battery cells is not equal to the ΔSOC′ generated by the battery during the equalization time, the equalization process continues. S34. After the equalization between individual battery cells is completed, the entire battery pack is treated as a single battery cell and the equalization between the packs is performed until the inconsistency of the entire battery pack is eliminated.

[0012] The data output by the first hidden layer and the second hidden layer are processed using the Dropout operation. Each hidden layer is set with a Dropout value of 0.5, which allows 50% of the neurons in each hidden layer to be randomly selected not to participate in the update of the network weight parameter matrix.

[0013] The weight parameter matrices W1, W2, and W3, as well as the bias vectors b1, b2, and b3, are randomly initialized by the system and then continuously updated by the Adam optimizer.

[0014] The neural network model is trained using 70% of the battery's SOC, voltage data, and tag data as training data, following the data flow in step 3.

[0015] A balancing system for implementing the method of balancing lithium battery packs of underwater vehicles using a neural network model includes a CUK circuit connecting two individual battery cells and a flyback transformer circuit connecting two battery packs.

[0016] The balancing circuit between individual battery cells uses a CUK balancing circuit, and the balancing circuit between battery packs adopts a flyback transformer balancing topology circuit.

[0017] An application of the method for balancing lithium battery packs of underwater vehicles using a neural network model involves inputting the battery's operating state at any given time into the neural network model, determining the balancing strategy for the current battery operating state based on the output of the neural network model, and then selecting the corresponding balancing strategy to activate the balancing circuit and perform balancing processing on the battery.

[0018] Beneficial effects This invention proposes a method for balancing lithium battery packs in underwater vehicles using a neural network model, along with a system for achieving this balancing. By establishing a neural network model, balancing strategies based on battery SOC (State of Charge) and voltage can be integrated. A fully trained neural network model can dynamically select balancing indicators and strategies based on battery operating data, effectively improving the efficiency of the balancing strategy. Balancing the lithium battery pack of the underwater vehicle using either a battery SOC-based or voltage-based strategy eliminates inconsistencies within the entire battery pack, preventing overcharging or over-discharging, extending battery life, and avoiding significant losses due to spontaneous combustion caused by battery thermal runaway.

[0019] Specifically, the following points are mentioned: 1. This invention provides an balancing method combining two variables. Through an experimentally trained neural network model, it determines the current battery usage and provides the most suitable balancing strategy for the current battery state. Compared to balancing strategies that use only one balancing variable during battery operation, this algorithm fully considers the characteristics of each stage of battery operation and performs corresponding balancing control for different usage conditions, greatly improving energy utilization and balancing effect. This enables underwater electric power equipment to operate reliably when handling complex tasks with multiple operating conditions. 2. Compared with the existing method of calculating the remaining power using simple ampere-hour integration, this invention directly converts the SOC by voltage, making the balancing strategy more accurate and faster; 3. By establishing a neural network model, balancing strategies based on battery SOC and those based on voltage can be integrated. A fully trained neural network model can dynamically select balancing indices based on battery operating data, effectively improving the efficiency of the balancing strategy. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the present invention; Figure 2This is the SOC-OCV curve of the battery obtained from the example experiment; Figure 3 It is a neural network model structure; Figure 4 This is a system diagram of an implementation example of the equilibrium strategy; Figure 5 It is a balanced topology circuit. Detailed Implementation

[0021] The present invention will now be further described in conjunction with the embodiments and accompanying drawings: Figure 4 This is a system diagram of the balancing strategy in this embodiment. The input of the BP neural network model receives the current SOC and voltage values ​​of the battery, along with the corresponding tags. The output of the BP neural network model is connected to the balancing system of the lithium battery pack. In the balancing system, Cuk circuits connect two individual battery cells, and flyback transformer circuits connect two battery packs.

[0022] The balancing circuit between individual battery cells uses a CUK balancing circuit, and the balancing circuit between battery packs adopts a flyback transformer balancing topology circuit. See Figure 5 Energy between adjacent battery cells can be transferred via capacitor storage. A PWM wave is then used to control the MOSFETs to achieve equalization. After individual cell equalization is complete, equalization between battery packs begins. The inter-pack equalization circuit is based on a flyback transformer equalization topology.

[0023] In this embodiment, the BP neuron network model is first trained. The training steps are as follows: Step 1: Obtained from HPPC experiments Figure 2 The SOC-OCV curve shows that in the middle range, the voltage changes relatively smoothly with SOC, while at high and low SOC, even small changes in SOC result in large changes in voltage. Therefore... Figure 2 The SOC-OCV curve reflects different operating states of the battery, and depending on the different operating states of the battery, it is more appropriate to choose voltage or battery SOC for balancing. The process of using HPPC experiments to test lithium battery packs and obtain the SOC-OCV curves is as follows: S11. Charge the battery at a constant current and constant voltage in an environment of 25°C until the battery SOC is 100%; after an interval of 3 hours, discharge it at a constant current of 1C for 30 seconds. S12. After a 40-second interval, charge at a constant current rate of 1C for 30 seconds, let stand for 30 minutes, and finally discharge at a constant current rate of 1 / 3C for 18 minutes, and let stand for 3 hours. S13. Repeat steps S11 and S12 to obtain the SOC-OCV curve (see below). Figure 2 ; S14. By fitting the curve with an 8th-order polynomial using the least squares method, the relationship between OCV and SOC is obtained.

[0024] Step 2: Construct a multi-layer neural network model: The neural network model is determined to have two inputs and two outputs. The input variables are the current SOC and voltage values ​​of the battery. The output results are either an balancing strategy using voltage as the balancing variable or an balancing strategy using battery SOC as the balancing variable. There are two hidden layers in the middle. The first hidden layer has 10 neurons to extract data features as meticulously as possible. The second hidden layer has 15 neurons to further extract more important feature information. Finally, two output results are obtained through a fully connected layer. Due to limited experimental data, to prevent overfitting of the neural network model and improve its generalization ability, a Dropout operation is used after the two hidden layers. The Dropout value is set to 0.5 for each hidden layer, so that 50% of the neurons in each hidden layer are randomly selected not to participate in the update of the network weight parameter matrix.

[0025] Step 3: Read the battery's SOC, voltage data, and labels. Labels indicating good performance of the balancing strategy based on battery SOC are assigned as '1', and labels indicating good performance of the balancing strategy based on battery voltage are assigned as '2'. Convert these to tensor data types that the neural network can recognize. Use 70% of the data as training data and 30% as test data. Since the battery SOC and voltage values ​​are of different orders of magnitude, to prevent the neural network model from biasing its weights towards larger values, the samples need to be standardized to ensure the data is symmetrical about the origin.

[0026] Step 3.1: In the first hidden layer, the input SOC, voltage value, weight parameter matrix W1, and bias vector b1 are subjected to a fully connected operation to obtain 10 feature values ​​for each input value; The fully connected operation of the first hidden layer is: Where: a represents the feature of the first hidden layer. ; x is the input value of the neural network model, W1 is the first hidden layer 2 The weight parameter matrix is ​​10, and b1 is the first hidden layer 1. The bias vector is 10, f is the ReLU activation function; n1 represents the number of variables in the second layer as input to the first layer; Step 3.2: The output value of the first hidden layer is non-linearized by the ReLU function. In the second hidden layer, a fully connected operation is performed with the weight parameter matrix W2 and the bias vector b2 to obtain 15 feature values ​​for each input value. The fully connected operation of the second hidden layer is: Where: c represents the second hidden layer feature. W2 is the second hidden layer 10. The weight parameter matrix is ​​15, b2 is the bias vector of the second hidden layer; n2 represents the number of variables in the third layer that are input to the second layer. Step 3.3: The output value of the second hidden layer is nonlinearized using the ReLU function and then fully connected with the weight parameter matrix W3 and the bias vector b3 to obtain the probability values ​​of the two equilibrium strategies output by the neural network model. ; The fully connected operation with the weight parameter matrix W3 and the bias vector b3 is as follows: in: The probability values ​​of the equilibrium strategy, where: k is the identification label in step 3, with a value of 1 or 2; W3 is the output layer 15. The weight parameter matrix of layer 2, and b3 is the output layer 1. 2. Bias vector.

[0027] W1, W2, and W3 are weight parameter matrices: The W1 is 2 A 10-weight parameter matrix is ​​initially obtained by random initialization by the system and then continuously updated by the Adam optimizer. The W2 is 10 The weight parameter matrix consists of 15 parts. The initial matrix is ​​randomly initialized by the system and then continuously updated by the Adam optimizer. The W3 is 15 2. Weight parameter matrix. The initial matrix is ​​randomly initialized by the system and then continuously updated by the Adam optimizer.

[0028] b1, b2, and b3 are bias vectors: b1 is 1 10 bias vector parameters, the initial vector is randomly initialized by the system and then continuously updated by the Adam optimizer; b2 is 1 15 bias vector parameters, the initial vector is randomly initialized by the system and then continuously updated by the Adam optimizer; b3 is 1 2. Bias vector parameters: The initial vector is randomly initialized by the system and then continuously updated by the Adam optimizer.

[0029] Step 4: Forward Propagation of the Neural Network Model: Calculate the model loss using the loss function based on the model's output value and labels. Then, perform backpropagation to obtain the gradient value of each layer. Calculate the exponential moving average of the gradient, taking into account the gradient momentum from previous time steps. Next, calculate the exponential moving average of the squared gradients and apply a weighted average to the squared gradients. Correct the bias in the gradient mean to reduce its impact on the early stages of training. Update the weight parameter matrix of each layer, which is the initial learning rate multiplied by the ratio of the square root of the gradient mean to the gradient variance. After each training round, test the model on the test dataset and save the loss value. Do not perform backpropagation during testing. After multiple training rounds, select the model with the lowest loss value on the test dataset.

[0030] The trained BP neural network model is obtained through steps 1 to 4. This neural network model can select an appropriate balancing strategy based on the real-time state of the battery. The process is as follows: Step 1) Input the battery's working state at any time into the trained neural network model. The neural network model will provide the most suitable balancing strategy for the current battery working state. When the neural network model outputs the probability of an equilibration strategy using voltage as an indicator... The probability is greater than that of an equilibrium strategy based on battery SOC. If the voltage is the balancing strategy, select the balancing strategy based on the battery SOC and execute step 2; otherwise, select the balancing strategy based on the battery SOC and execute step 3.

[0031] Step 2): The steps of the voltage-based balancing strategy are as follows: S21. Collect the voltage difference between the two individual battery cells; S22. Amplify the collected voltage by at least 100 times; S23. Compare the voltage between two individual battery cells. If the voltage between the two individual battery cells is inconsistent, the equalization circuit is activated. S24. After the equalization between individual battery cells is completed, the entire battery pack is treated as a single battery cell and the equalization between the packs is started until the inconsistency of the entire battery pack is eliminated.

[0032] Step 3) The balancing strategy based on battery SOC is as follows: S31. Based on the relationship between SOC and voltage, the voltage difference between two individual battery cells is converted into the difference in SOC; S32. The ΔSOC′ generated by the battery during the equalization time is calculated using the ampere-hour integration method; S33. If ΔSOC = ΔSOC′, then the two battery cells are considered to have reached equilibrium. If they are not equal, then the balancing process begins. When the difference ΔSOC between the two individual battery cells is not equal to the ΔSOC′ generated by the battery during the balancing time, the balancing process continues. S34. After the equalization between individual battery cells is completed, the entire battery pack is treated as a single battery cell and the equalization between the packs is performed until the inconsistency of the entire battery pack is eliminated.

[0033] This invention integrates balancing strategies based on battery SOC and voltage using a neural network model. The fully trained neural network model dynamically selects the balancing index based on battery operating data, choosing appropriate balancing strategies for different battery operating states, effectively improving the efficiency of the balancing strategy. It promptly eliminates inconsistencies between battery cells during operation, slows down battery capacity decay, extends battery life, and significantly increases the range of the underwater vehicle. Furthermore, it avoids thermal runaway faults caused by overcharging and discharging, effectively protecting the lithium battery pack and the underwater vehicle itself.

Claims

1. A method for balancing lithium battery packs in underwater vehicles using a neural network model, characterized in that... Includes the following steps: Step 1: The lithium battery pack was tested using HPPC experiments to obtain the SOC-OCV curve; Step 2: Construct a multi-layer neural network model. The first layer is the input layer with two neuron variables. The second and third layers are hidden layers, with 10 neurons in the first hidden layer and 15 neurons in the second hidden layer. The fourth layer is the output layer with two neuron variables. All neurons in each layer are fully connected. Each layer uses the ReLU activation function and the CrossEntropyLoss loss function. The backpropagation optimizer for the neural network model uses Adam, dynamically adjusting the learning rate of each parameter using the first and second moment estimates of the gradient. Step 3: Assign identification tags to the current SOC and voltage values ​​of the battery, where the tag for the balancing strategy effect based on battery SOC is 1, and the tag for the balancing strategy effect based on battery voltage is 2. Step 3.1: In the first hidden layer, the two input variables SOC and voltage value are fully connected with the weight parameter matrix W1 and the bias vector b1 to obtain 10 feature values ​​for each input value; The fully connected operation of the first hidden layer is: Where: a represents the feature of the first hidden layer. W1 is the first hidden layer 2 The weight parameter matrix is ​​10, where b1 is the first hidden layer 1. The bias vector is 10; f is the ReLU activation function, x is the variable input to the activation function f; n1 represents the number of variables in the second layer that are input from the first layer. Step 3.2: The output value of the first hidden layer is non-linearized by the ReLU function. In the second hidden layer, a fully connected operation is performed with the weight parameter matrix W2 and the bias vector b2 to obtain 15 feature values ​​for each input value. The fully connected operation of the second hidden layer is: Where: c represents the second hidden layer feature. W2 is the second hidden layer 10. The weight parameter matrix is ​​15, b2 is the bias vector of the second hidden layer; n2 represents the number of variables in the third layer that are input to the second layer. Step 3.3: The output value of the second hidden layer is non-linearized using the ReLU function, and then fully connected with the weight parameter matrix W3 and the bias vector b3 to obtain two probability values ​​representing the two equilibrium strategies output by the neural network model. ; The fully connected operation with the weight parameter matrix W3 and the bias vector b3 is as follows: in: The probability values ​​of the equilibrium strategy, where: k is the identification label in step 3, with a value of 1 or 2; W3 is the output layer 15. The weight parameter matrix of layer 2, and b3 is the output layer 1.

2. Bias vector; n3 represents the number of variables input from the third layer to the fourth layer; Step 4: Select the equalization strategy based on the probability values ​​of the two equalization strategies output by the neural network model: When the probability of the equalization strategy output by the neural network model using voltage as an indicator is... The probability is greater than that of an equilibrium strategy based on battery SOC. When the voltage is high, a balancing strategy based on voltage is selected; otherwise, a balancing strategy based on battery SOC is selected.

2. The method for balancing lithium battery packs of underwater vehicles using a neural network model according to claim 1, characterized in that: Step 1 involves using HPPC experiments to test the lithium battery pack and obtaining the SOC-OCV curve as follows: S11. Charge the battery at a constant current and constant voltage in an environment of 25°C until the battery SOC is 100%; after an interval of 3 hours, discharge it at a constant current of 1C for 30 seconds. S12. After a 40-second interval, charge at a constant current rate of 1C for 30 seconds, let stand for 30 minutes, and finally discharge at a constant current rate of 1 / 3C for 18 minutes, and let stand for 3 hours. S13. Repeat S11 and S12 to obtain the SOC-OCV curve; S14. Using the least squares method, an 8th-order polynomial is fitted to the curve to obtain the relationship between OCV and SOC.

3. The method for balancing lithium battery packs of underwater vehicles using a neural network model according to claim 1, characterized in that: The balancing strategy steps using voltage as an indicator in step 4 are as follows: S21. Collect the voltage difference between two individual battery cells; S22. Amplify the collected voltage by at least 100 times; S23. When the voltage between two individual battery cells is inconsistent, activate the equalization circuit; S24. After the equalization between individual battery cells is completed, the equalization between battery packs is started by treating the entire battery pack as a single battery cell until the inconsistency of the entire battery pack is eliminated.

4. The method for balancing lithium battery packs of underwater vehicles using a neural network model according to claim 1, characterized in that: The balancing strategy steps in step 4, which use battery SOC as an indicator, are as follows: S31. Based on the relationship between SOC and voltage, the voltage difference between two individual battery cells is converted into the difference in SOC; S32. The ΔSOC′ generated by the battery during the equalization time is calculated using the ampere-hour integration method; S33. If ΔSOC = ΔSOC′, the two battery cells are considered to be in a balanced state. If they are not in a balanced state, the equalization circuit is activated. When the SOC difference ΔSOC between the two individual battery cells is not equal to the ΔSOC′ generated by the battery during the equalization time, the equalization process continues. S34. After the equalization between individual battery cells is completed, the entire battery pack is treated as a single battery cell and the equalization between the packs is performed until the inconsistency of the entire battery pack is eliminated.

5. The method for balancing lithium battery packs of underwater vehicles using a neural network model according to claim 1, characterized in that: The data output by the first hidden layer and the second hidden layer are processed using the Dropout operation. Each hidden layer is set with a Dropout value of 0.5, which allows 50% of the neurons in each hidden layer to be randomly selected not to participate in the update of the network weight parameter matrix.

6. The method for balancing lithium battery packs of underwater vehicles using a neural network model according to claim 1, characterized in that: The weight parameter matrices W1, W2, and W3, as well as the bias vectors b1, b2, and b3, are randomly initialized by the system and then continuously updated by the Adam optimizer.

7. The method for balancing lithium battery packs of underwater vehicles using a neural network model according to claim 1, characterized in that: The neural network model is trained using 70% of the battery's SOC, voltage data, and tag data, following the data flow in step 3.

8. A balancing system for implementing the method of balancing lithium battery packs of underwater vehicles using a neural network model as described in any one of claims 1 to 7, characterized in that: A CUK circuit is connected between two individual battery cells, and a flyback transformer circuit is connected between the two battery packs.

9. The equalization system according to claim 8, characterized in that: The balancing circuit between individual battery cells uses a CUK balancing circuit, and the balancing circuit between battery packs adopts a flyback transformer balancing topology circuit.

10. An application of the method for balancing lithium battery packs of underwater vehicles using a neural network model as described in any one of claims 1 to 7, characterized in that: The battery's operating state at any given time is input into the neural network model. Based on the output of the neural network model, an equalization strategy for the current battery operating state is given. Then, the corresponding equalization strategy is selected to activate the equalization circuit and perform equalization processing on the battery.

Images