A dynamic equalization control method for a new energy vehicle battery pack

Abstract

The application discloses a kind of new energy automobile battery pack dynamic equalization control method, comprising the following steps: acquisition battery pack during operation Multidimensional data;Carry out noise suppression, obtain Multidimensional denoising data;Estimate the state parameter of each battery by using the improved Kalman filtering algorithm based on adaptive forgetting factor;Construct multidimensional equalization priority evaluation model, determine high priority battery;Based on the hybrid algorithm of deep reinforcement learning and model predictive control, generate the dynamic equalization strategy of each high priority battery;Based on bidirectional step-down-boost converter topology, using adaptive duty cycle regulation algorithm according to dynamic equalization strategy Equalization control.This application improves the equalization control comprehensiveness and reliability of new energy automobile battery pack by combining improved Kalman filtering algorithm, multidimensional equalization priority evaluation model, hybrid algorithm based on deep reinforcement learning and model predictive control, adaptive duty cycle regulation algorithm.

Description

Technical Field

[0001] This application relates to the field of battery pack equalization control technology, and in particular to a dynamic equalization control method for new energy vehicle battery packs. Background Technology

[0002] Battery equalization control is a key technology for solving the cell consistency problem. Its core objective is to maintain consistency in the state of charge (SOC) and voltage of each cell through energy transfer or replenishment, thereby maximizing the battery pack's capacity, extending its lifespan, and ensuring safe operation. Currently, mainstream equalization control methods include passive and active equalization. Passive equalization uses resistors to dissipate excess energy from high-capacity cells, resulting in a simple structure but high energy loss and low equalization efficiency. Active equalization uses converters to achieve energy transfer between cells, resulting in low energy loss. However, traditional active equalization often relies on single voltage or SOC data and employs fixed algorithms and equalization strategies, making it difficult to adapt to the complex operating conditions of new energy vehicles (such as rapid acceleration, deceleration, and bumpy roads) and the dynamic changes caused by cell aging. This leads to low equalization accuracy and slow response, failing to fundamentally solve the problem of deteriorating cell consistency. Therefore, developing a dynamic equalization control method that adapts to complex operating conditions, is precise and efficient, and is self-adaptive throughout the entire lifecycle has become an urgent need for the development of new energy vehicle power battery technology.

[0003] In the prior art, Chinese patent CN110228397A discloses an electric vehicle power battery balancing system and its balancing method. Individual power batteries are connected in series. An embedded processor collects the voltage of the individual power batteries in the battery pack through a control voltage acquisition module. Based on the collected voltage of the individual power batteries, the embedded processor generates a balancing control signal and controls all battery balancing modules to work together in different balancing directions through the balancing control unit.

[0004] However, the existing technologies mentioned above only use the voltage of individual cells as the sole basis for acquisition and evaluation, lack multi-dimensional operational data support, have a single balancing judgment, lack dynamic adaptability of strategies, and do not adequately consider factors such as cell aging and temperature. The balancing accuracy and adaptability to operating conditions are limited, and the comprehensiveness and reliability of the balancing control of new energy vehicle battery packs need to be improved. Summary of the Invention

[0005] This application provides a dynamic equalization control method for new energy vehicle battery packs, which addresses the problem that the comprehensiveness and reliability of equalization control of new energy vehicle battery packs need to be improved in the prior art.

[0006] On the one hand, this application provides a dynamic equalization control method for a new energy vehicle battery pack, including the following steps: Step 1: Collect multi-dimensional data during the operation of the battery pack, including voltage data, current data, surface temperature data, vibration acceleration data, and aging status data.

[0007] Step 2: Perform noise suppression on the multi-dimensional data to obtain multi-dimensional denoised data, including denoised voltage data, denoised current data, denoised surface temperature data, denoised vibration acceleration data, and denoised aging state data.

[0008] Step 3: Based on the multi-dimensional denoised data, the improved Kalman filter algorithm based on the adaptive forgetting factor is used to estimate the state parameters of each cell in the battery pack, including state of charge, state of health, calibration voltage, and calibration temperature.

[0009] Step 4: Based on the state parameters of each cell, construct a multi-dimensional equalization priority evaluation model to determine whether equalization needs to be initiated and identify high-priority cells.

[0010] Step 5: Based on a hybrid algorithm of deep reinforcement learning and model predictive control, a dynamic balancing strategy for each high-priority cell is generated, including the optimal balancing power and the optimal balancing time.

[0011] Step 6: Based on the bidirectional buck-boost converter topology, an adaptive duty cycle adjustment algorithm is used to perform equalization control according to the dynamic equalization strategy.

[0012] In one possible implementation, in step two, a wavelet threshold denoising algorithm is used to suppress noise in the voltage data and the current data to obtain denoised voltage data and denoised current data.

[0013] The surface temperature data is noise-suppressed using a moving average filtering algorithm to obtain denoised surface temperature data.

[0014] The vibration acceleration data is noise-suppressed using the empirical mode decomposition algorithm to obtain denoised vibration acceleration data.

[0015] The Gaussian filtering algorithm is used to suppress noise in the aging state data to obtain denoised aging state data.

[0016] In one possible implementation, step three, the improved Kalman filter algorithm based on the adaptive forgetting factor, includes: Establish the state equation for the state vector and the observation equation for the observation vector.

[0017] Design an adaptive forgetting factor.

[0018] Based on the state equation, the observation equation, and the adaptive forgetting factor, Kalman filtering iteration is performed to output the state parameters of each cell in the battery pack.

[0019] In one possible implementation, step three, constructing the mapping relationship between the multi-dimensional denoised data and the state parameters, includes: The state of charge is corrected based on the denoised voltage data, which is also used to calculate the calibration voltage.

[0020] The fundamental change in the state of charge is calculated based on the denoised current data.

[0021] The calibration temperature is calculated based on the noise-reduced surface temperature data.

[0022] The health status is corrected based on the denoised vibration acceleration data and the denoised aging status data.

[0023] In one possible implementation, in step four, the corresponding state evaluation index is calculated based on the state parameters of each cell, including state of charge deviation, health state deviation, voltage deviation, and temperature deviation.

[0024] A multi-dimensional balanced priority evaluation model is constructed based on the state of charge deviation, the health state deviation, the voltage deviation, and the temperature deviation to calculate the comprehensive risk coefficient of each cell.

[0025] When at least one cell in the battery pack has a comprehensive risk coefficient greater than a preset risk threshold, it is determined that equalization needs to be initiated. The cells are sorted from high to low according to their comprehensive risk coefficients, and the top few cells are selected as high-priority cells.

[0026] In one possible implementation, in step five, the hybrid algorithm based on deep reinforcement learning and model predictive control includes a deep reinforcement learning module and a model predictive control module.

[0027] The deep reinforcement learning module is used to construct an agent, with the state parameters of high-priority cells as the state space, the equalization power and equalization time as the action space, and the minimum deviation of the state of charge and the minimum energy loss after equalization as the reward function, and outputs the initial equalization strategy.

[0028] The model prediction and control module is used to establish a cell balancing response model. Taking the initial balancing strategy as input, it minimizes the objective function and performs rolling optimization to obtain the dynamic balancing strategy for each high-priority cell.

[0029] In one possible implementation, step six includes: The operating mode of the bidirectional buck-boost converter is determined based on the status parameters of the high-priority cells.

[0030] An adaptive duty cycle adjustment algorithm is used to calculate the initial duty cycle based on the operating mode and the optimal balanced power.

[0031] The duty cycle is dynamically adjusted using a PID control algorithm.

[0032] When the optimal balancing time is reached, the balancing control of high-priority cells is stopped.

[0033] In one possible implementation, step six is ​​followed by: Step 7: Evaluate the effectiveness of the balance control and optimize the parameters in steps 1, 3, and 5 based on the evaluation results.

[0034] In one possible implementation, in step seven, an evaluation index for the balancing effect is constructed based on the voltage consistency coefficient, the mean state-of-charge deviation rate, the balancing efficiency, and the temperature uniformity coefficient.

[0035] The particle swarm optimization algorithm is used to adjust the data collection frequency in step one, the initial forgetting factor in step three, and the weight coefficient of the reward function in step five, with the optimization objective of maximizing the equilibrium effect evaluation index.

[0036] The dynamic balance control method for a new energy vehicle battery pack disclosed in this application has the following advantages: By combining an improved Kalman filter algorithm, a multi-dimensional equalization priority evaluation model, a hybrid algorithm based on deep reinforcement learning and model predictive control, and an adaptive duty cycle adjustment algorithm, the comprehensiveness and reliability of equalization control of new energy vehicle battery packs are improved.

[0037] By establishing the state equation of the state vector and the observation equation of the observation vector, an adaptive forgetting factor is designed. Based on the state equation, the observation equation, and the adaptive forgetting factor, Kalman filtering iteration is performed to output the state parameters of each cell in the battery pack, thereby enhancing the adaptability of the filtering algorithm to the time-varying environment in the vehicle.

[0038] By constructing a mapping relationship between multi-dimensional denoised data and state parameters, including denoised vibration acceleration data and denoised aging state data, the estimation accuracy of the cell's state parameters is improved.

[0039] By constructing a multi-dimensional equilibrium priority assessment model based on state of charge deviation, health state deviation, voltage deviation, and temperature deviation, the comprehensive risk coefficient of each cell is calculated, avoiding the limitations of traditional single voltage index assessment and realizing the quantitative classification of equilibrium requirements.

[0040] The proposed hybrid algorithm based on deep reinforcement learning and model predictive control outputs an initial equilibrium policy through a deep reinforcement learning module and performs rolling optimization through a model predictive control module, thus resolving the contradiction between the poor real-time performance and insufficient optimization of traditional equilibrium policies.

[0041] By employing an adaptive duty cycle adjustment algorithm to calculate the initial duty cycle based on the operating mode and optimal balanced power, and then dynamically adjusting the duty cycle through a PID control algorithm, bidirectional energy flow is supported, thereby improving the balancing efficiency.

[0042] By constructing an evaluation index for the equilibrium effect, and using the particle swarm optimization algorithm to optimize the parameters in steps one, three, and five, this closed-loop optimization mechanism improves the long-term applicability of the method. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating a dynamic balancing control method for a new energy vehicle battery pack, provided in an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] like Figure 1 As shown in the figure, this application provides a dynamic equalization control method for a new energy vehicle battery pack, including the following steps: Step 1: Collect multi-dimensional data during the operation of the battery pack, including voltage data, current data, surface temperature data, vibration acceleration data, and aging status data.

[0047] Step 2: Perform noise suppression on the multi-dimensional data to obtain multi-dimensional denoised data, including denoised voltage data, denoised current data, denoised surface temperature data, denoised vibration acceleration data, and denoised aging state data.

[0048] Step 3: Based on the multi-dimensional denoised data, the improved Kalman filter algorithm based on the adaptive forgetting factor is used to estimate the state parameters of each cell in the battery pack, including state of charge, state of health, calibration voltage, and calibration temperature.

[0049] Step 4: Based on the state parameters of each cell, construct a multi-dimensional equalization priority evaluation model to determine whether equalization needs to be initiated and identify high-priority cells.

[0050] Step 5: Based on a hybrid algorithm of deep reinforcement learning and model predictive control, a dynamic balancing strategy for each high-priority cell is generated, including the optimal balancing power and the optimal balancing time.

[0051] Step 6: Based on the bidirectional buck-boost converter topology, an adaptive duty cycle adjustment algorithm is used to perform equalization control according to the dynamic equalization strategy.

[0052] Specifically, in this embodiment, in step one, for the N cells in the battery pack, a distributed acquisition system is used to complete the synchronous acquisition of multi-dimensional data for each cell, as follows: Voltage data is acquired through a voltage acquisition module and serves as a fundamental parameter for the real-time energy status of the battery cell.

[0053] Current data is acquired through a Hall current sensor, which characterizes the magnitude and direction of the cell's charging and discharging current.

[0054] Surface temperature data is acquired through a thermistor temperature sensor to characterize the thermal properties of the battery cell and parameters related to the risk of thermal runaway.

[0055] Vibration acceleration data is collected by a triaxial accelerometer to characterize the mechanical vibration and crush damage of the battery cells during vehicle operation.

[0056] Aging status data is obtained through an internal resistance tester and a cycle life database, serving as a priori evaluation parameter for the degree of cell degradation.

[0057] The multi-dimensional data in step one is denoted as .in, This represents the voltage data of the i-th cell at time t. This represents the current data of the i-th cell at time t. This represents the surface temperature data of the i-th cell at time t. This represents the vibration acceleration data of the i-th cell at time t. This represents the aging status data of the i-th cell at time t.

[0058] For example, in step two, a wavelet threshold denoising algorithm is used to suppress noise in the voltage data and the current data to obtain denoised voltage data and denoised current data.

[0059] The surface temperature data is noise-suppressed using a moving average filtering algorithm to obtain denoised surface temperature data.

[0060] The vibration acceleration data is noise-suppressed using the empirical mode decomposition algorithm to obtain denoised vibration acceleration data.

[0061] The Gaussian filtering algorithm is used to suppress noise in the aging state data to obtain denoised aging state data.

[0062] Specifically, in this embodiment, in step two, the wavelet threshold denoising algorithm uses the db4 wavelet as the wavelet basis function for voltage and current data; the number of decomposition layers is 3; the threshold function is a soft threshold function; and the threshold selection method is a general threshold.

[0063] The moving average filtering algorithm is applied to surface temperature data with a window length of 5. The filtering type is ordinary moving average filtering. The boundary treatment is as follows: the first and last two sampling points are filled with mirror images and then the moving average is calculated.

[0064] The Empirical Mode Decomposition (EMD) algorithm is used for vibration acceleration data. The intrinsic mode function is processed as follows: the first two and last two intrinsic mode functions are removed; the intrinsic mode functions of the intermediate order are retained; the reconstruction method is to add the retained intrinsic mode functions to obtain the denoised vibration acceleration data.

[0065] The Gaussian filtering algorithm is designed for aging data, with a filter kernel variance of 0.5. The filter kernel is a one-dimensional Gaussian kernel with an equivalent length of 5. The filtering method is to perform one-dimensional Gaussian convolution smoothing on the aging data sequence.

[0066] Multi-dimensional denoised data is denoted as .in, This represents the denoised voltage data of the i-th cell at time t. This represents the denoised current data of the i-th cell at time t. This represents the denoised surface temperature data of the i-th cell at time t. This represents the denoised vibration acceleration data of the i-th cell at time t. This represents the noise reduction and aging status data of the i-th cell at time t.

[0067] For example, in step three, the improved Kalman filter algorithm based on the adaptive forgetting factor includes: Establish the state equation for the state vector and the observation equation for the observation vector.

[0068] Design an adaptive forgetting factor.

[0069] Based on the state equation, the observation equation, and the adaptive forgetting factor, Kalman filtering iteration is performed to output the state parameters of each cell in the battery pack.

[0070] Specifically, in this embodiment, in step three, the state vector is defined as follows: .

[0071] in, This represents the state vector of the i-th cell at time t. This represents the state of charge of the i-th cell at time t. This represents the health status of the i-th cell at time t. This represents the calibration voltage of the i-th cell at time t. This represents the calibration temperature of the i-th cell at time t. That is, the state vector corresponds to the state parameters.

[0072] The observation vector is defined as follows: .

[0073] in, This represents the observation vector of the i-th cell at time t, which corresponds to the multi-dimensional denoised data.

[0074] The state equations for establishing the state vector are as follows: .

[0075] in, Let represent the state transition matrix (4×4) from time t-1 to time t. The system noise driving matrix at time t-1 is represented by (4×4). Let t represent the state noise vector (4×1) of the i-th cell at time t.

[0076] The observation equations for the observation vectors are established as follows: .

[0077] in, The matrix (5×4) represents the mapping relationship between the multi-dimensional denoised data and the state parameters at time t. Let represent the observation noise vector (5×1) of the i-th cell at time t.

[0078] The adaptive forgetting factor is designed as follows: .

[0079] in, This represents the adaptive forgetting factor at time t (greater than 0 and less than or equal to 1). This represents the initial forgetting factor (0.95 in this example). This represents the forgetting rate balance coefficient (0.8 in this example). Let represent the observation prediction vector of the i-th cell at time t, denoted by . get( (This represents the state prediction vector of the i-th cell at time t). This represents the Euclidean norm.

[0080] Based on the state equation, the observation equation, and the adaptive forgetting factor, the Kalman filter iteration is performed as follows: State prediction: .

[0081] in, This represents the state estimate of the i-th cell at time t-1 after the state update.

[0082] Covariance prediction: .

[0083] in, Let represent the one-step prediction covariance matrix of the i-th cell at time t. This represents the optimal state estimation covariance matrix of the i-th cell at time t-1. The transpose of the state transition matrix. This represents the transpose of the system noise driving matrix. Let represent the system noise covariance matrix at time t-1.

[0084] Filter gain calculation: .

[0085] in, This represents the Kalman filter gain matrix of the i-th cell at time t. This represents the transpose of the mapping matrix. Let represent the observation noise covariance matrix at time t.

[0086] Status Update: .

[0087] in, This represents the optimal state estimate of the i-th cell at time t after the state update.

[0088] Covariance update: .

[0089] in, This represents the optimal state estimation covariance matrix after the state update of the i-th cell at time t. This represents the identity matrix (4×4).

[0090] For example, in step three, constructing the mapping relationship between the multi-dimensional denoised data and the state parameters includes: The state of charge is corrected based on the denoised voltage data, which is also used to calculate the calibration voltage.

[0091] The fundamental change in the state of charge is calculated based on the denoised current data.

[0092] The calibration temperature is calculated based on the noise-reduced surface temperature data.

[0093] The health status is corrected based on the denoised vibration acceleration data and the denoised aging status data.

[0094] Specifically, in this embodiment, the denoised voltage data The state of charge is corrected using an open-circuit voltage-state-of-charge calibration curve. Calibration voltage. The calculation formula is: .

[0095] in, This represents the voltage-temperature correction factor.

[0096] Denoising current data The fundamental change in state of charge is calculated using the ampere-hour integral method. : .

[0097] in, This represents the nominal capacity of the i-th cell.

[0098] Calibration temperature The calculation formula is: .

[0099] in, This indicates the correction value for temperature rise caused by vibration.

[0100] health status The calculation formula is: .

[0101] in, Represents the prior health status based on aging status data, by get, This represents the cycle life correction factor. This represents the vibration damage coefficient.

[0102] The state parameters output in step three are denoted as follows: ,in, This represents the final value of the state of charge of the i-th cell at time t. This represents the final health status value of the i-th cell at time t. This represents the calibration voltage of the i-th cell at time t. This represents the calibration temperature of the i-th cell at time t.

[0103] For example, in step four, the corresponding state evaluation index is calculated based on the state parameters of each cell, including state of charge deviation, health state deviation, voltage deviation, and temperature deviation.

[0104] A multi-dimensional balanced priority evaluation model is constructed based on the state of charge deviation, the health state deviation, the voltage deviation, and the temperature deviation to calculate the comprehensive risk coefficient of each cell.

[0105] When at least one cell in the battery pack has a comprehensive risk coefficient greater than a preset risk threshold, it is determined that equalization needs to be initiated. The cells are sorted from high to low according to their comprehensive risk coefficients, and the top few cells are selected as high-priority cells.

[0106] Specifically, in this embodiment, the state of charge deviation is as follows: .

[0107] in, This represents the state-of-charge deviation of the i-th cell at time t. This represents the average state of charge of all cells at time t.

[0108] The health status deviations are as follows: .

[0109] in, This represents the health status deviation of the i-th cell at time t. This represents the average health status of all battery cells at time t.

[0110] The voltage deviation is as follows: .

[0111] in, This represents the voltage deviation of the i-th cell at time t. This represents the average calibration voltage of all cells at time t. This indicates the maximum permissible voltage deviation.

[0112] The temperature deviation is as follows: .

[0113] in, This represents the temperature deviation of the i-th cell at time t. This represents the average calibration temperature of all cells at time t. This indicates the maximum permissible temperature deviation.

[0114] In this embodiment, the overall risk factor is as follows: .

[0115] in, This represents the overall risk coefficient of the i-th battery cell at time t. , , , These represent the weights of each state assessment indicator in the comprehensive risk coefficient, satisfying that the sum of the weights is 1, as determined by the existing entropy weight method.

[0116] When at least one battery cell has a comprehensive risk coefficient When the risk exceeds a preset risk threshold (0.15 in this embodiment), a balancing start command is triggered at time t, which will then balance all data that meet the comprehensive risk coefficient at time t. Cells with a risk coefficient greater than 0.15 are sorted from high to low according to their comprehensive risk coefficient, and the top M cells (M≤N / 5, where N represents the total number of cells) are selected as high-priority cells.

[0117] For example, in step five, the hybrid algorithm based on deep reinforcement learning and model prediction control is provided with a deep reinforcement learning module and a model prediction control module.

[0118] The deep reinforcement learning module is used to construct an agent, with the state parameters of high-priority cells as the state space, the equalization power and equalization time as the action space, and the minimum deviation of the state of charge and the minimum energy loss after equalization as the reward function, and outputs the initial equalization strategy.

[0119] The model prediction and control module is used to establish a cell balancing response model. Taking the initial balancing strategy as input, it minimizes the objective function and performs rolling optimization to obtain the dynamic balancing strategy for each high-priority cell.

[0120] Specifically, in this embodiment, the intelligent agent defines the state space S of the high-priority battery cell as follows: .

[0121] Define the action space A as: .

[0122] in, This represents the balanced power of the i-th cell. This represents the balancing time of the i-th cell.

[0123] Set the reward function R as follows: .

[0124] in, This represents the average state-of-charge deviation of high-priority cells after equalization. This represents the energy loss during the equilibrium process. This represents the total energy input in equilibrium. This indicates an over-equilibrium penalty term (triggered when the state of charge deviation is below 0.02, with a value of 1, and 0 when not triggered). , , The weights are 0.6, 0.3, and 0.1, respectively, satisfying the condition that the sum is 1. In this embodiment, the weights are 0.6, 0.3, and 0.1.

[0125] After the agent is trained, it takes the state space data of high-priority cells as input and outputs the initial equalization strategy. .in This represents the initial equalization power of the i-th cell. This represents the initial equalization time of the i-th cell.

[0126] In this embodiment, the cell equalization response model is defined as: .

[0127] in, Indicates the i-th cell. Predicted state of charge at time t, Indicates the prediction time step. This represents the balance coefficient (0.96 in this example).

[0128] The objective function is constructed to achieve optimal cell state consistency and minimum energy loss after equalization: .

[0129] in, Indicates the length of the prediction time domain. This represents the average state of charge of all high-priority cells in the prediction time domain. , As weights, in this embodiment, we take 0.8 and 0.2 respectively.

[0130] Using the initial equilibrium strategy as input, under the constraints... , Below (of which, Indicates the maximum equalization power. (Representing the maximum balancing time), the objective function is optimized using a quadratic programming algorithm to output the dynamic balancing strategy for each high-priority cell. ,in, This represents the optimal balanced power of the i-th cell. This represents the optimal balancing time for the i-th cell.

[0131] For example, step six includes: The operating mode of the bidirectional buck-boost converter is determined based on the status parameters of the high-priority cells.

[0132] An adaptive duty cycle adjustment algorithm is used to calculate the initial duty cycle based on the operating mode and the optimal balanced power.

[0133] The duty cycle is dynamically adjusted using a PID control algorithm.

[0134] When the optimal balancing time is reached, the balancing control of high-priority cells is stopped.

[0135] Specifically, in this embodiment, the state of charge of the high-priority cells is compared. and the average state of charge of the battery pack .

[0136] like In this case, the bidirectional buck-boost converter operates in buck mode, transferring excess energy from the battery cell to the battery pack's energy bus; if In this case, the bidirectional buck-boost converter operates in boost mode, drawing energy from the energy bus to replenish the battery cell. Otherwise, the bidirectional buck-boost converter does not operate.

[0137] The initial duty cycle is calculated using an adaptive duty cycle adjustment algorithm: In buck mode, the initial duty cycle formula is: ,in, This indicates the energy bus voltage.

[0138] In boost mode, the initial duty cycle formula is: .

[0139] In this embodiment, the actual equalization power is used. With optimal balanced power The difference is the input deviation of the PID controller. Design an incremental PID algorithm to dynamically adjust the duty cycle: , , .

[0140] In the formula, This represents the duty cycle increment at the k-th sampling time. This represents the actual equalization power in the k-th sample. =0.4, =0.05, =0.02 is the PID control parameter. This represents the final duty cycle at the k-th sampling.

[0141] When the optimal equilibrium time is reached When this occurs, the equalization control of the corresponding high-priority battery cell is stopped.

[0142] For example, step six is ​​followed by: Step 7: Evaluate the effectiveness of the balance control and optimize the parameters in steps 1, 3, and 5 based on the evaluation results.

[0143] For example, in step seven, an evaluation index for the balancing effect is constructed based on the voltage consistency coefficient, the mean state of charge deviation rate, the balancing efficiency, and the temperature uniformity coefficient.

[0144] The particle swarm optimization algorithm is used to adjust the data collection frequency in step one, the initial forgetting factor in step three, and the weight coefficient of the reward function in step five, with the optimization objective of maximizing the equilibrium effect evaluation index.

[0145] Specifically, in this embodiment, the voltage consistency coefficient The calculation is as follows: .

[0146] in, This represents the average calibration voltage of all cells after equalization. This represents the standard deviation of voltage.

[0147] Mean of state of charge deviation rate The calculation is as follows: .

[0148] Balanced efficiency The calculation is as follows: .

[0149] in, This indicates an effective balance of energy.

[0150] Temperature uniformity coefficient The calculation is as follows: .

[0151] in, This represents the average calibration temperature of all cells after equalization. This represents the standard deviation of temperature.

[0152] Constructing a balanced performance evaluation index The calculation is as follows: .

[0153] in, , , , For the weights, in this embodiment, we take 0.3, 0.3, 0.2, and 0.2.

[0154] Evaluation indicators based on balanced effects With the goal of maximizing performance, the existing particle swarm optimization algorithm is used to iteratively optimize the data acquisition frequency in step one, the initial forgetting factor in step three, and the weight coefficients of the reward function in step five. In this embodiment, the number of particles is set to 20, the number of iterations is 50, and each particle corresponds to a combination of optimization variables. The position and velocity of the particles are randomly initialized to balance the performance evaluation indicators. The fitness value of the particles is used for iterative updates. When the number of iterations reaches 50 or the fitness value converges, the optimal parameters are output and synchronously updated to the parameter configurations in steps one, three, and five, thus achieving closed-loop optimization of the control method.

[0155] In one embodiment, to verify the effectiveness of the technical solution of this application, a battery pack simulation model was built in the MATLAB / Simulink environment. The simulation conditions are as follows: the battery pack consists of 20 ternary lithium-ion cells connected in series, with a nominal capacity of 50 Ah and a nominal voltage of 3.7 V; the initial SOC difference is ±5%, the initial SOH difference is ±3%, and the initial temperature difference is ±5°C; the operating condition is the CLTC-P cycle condition, the simulation duration is 3600 s, and the sampling period is 1 s.

[0156] Comparison Scheme 1: Single voltage data equalization scheme, that is, equalization is triggered only based on voltage deviation.

[0157] Comparison Scheme 2: Traditional Kalman Filter Algorithm.

[0158] Comparison Scheme 3: Pure model prediction control equilibrium algorithm and pure deep reinforcement learning equilibrium algorithm.

[0159] Table 1 shows the comparison results between the multi-dimensional data acquisition scheme and the single voltage data scheme in terms of equalization triggering and equalization effect: Table 1. Comparison of multi-dimensional data acquisition scheme and single voltage data scheme in terms of equalization triggering and equalization effect.

[0160] As shown in Table 1, the single voltage data scheme, which relies solely on voltage deviation for equalization judgment, is prone to false triggering or missed triggering when there are temperature inconsistencies, inconsistent aging, or changes in cell condition caused by vibration. This application introduces surface temperature data, vibration acceleration data, and aging state data to construct a multi-dimensional equalization priority evaluation model, which significantly improves the accuracy of equalization triggering, reduces the false triggering rate and missed triggering rate, and improves the consistency of SOC and temperature uniformity after equalization.

[0161] Table 2 shows the comparison results of the state parameter estimation error between the adaptive forgetting factor improved Kalman filter algorithm of this application and the traditional Kalman filter algorithm: Table 2. Comparison of state parameter estimation errors between the adaptive forgetting factor improved Kalman filter algorithm and the traditional Kalman filter algorithm.

[0162] As shown in Table 2, this application significantly reduces the estimation errors of SOC and SOH, improves the accuracy of calibration voltage and calibration temperature, and accelerates the filter convergence speed by designing an adaptive forgetting factor and incorporating vibration acceleration and aging state data into the state equation to correct the health state. In vehicle-mounted time-varying environments, the improved Kalman filter algorithm can better track changes in cell state.

[0163] Table 3 shows the comparison results of the hybrid algorithm based on deep reinforcement learning and model predictive control in this application with the pure model predictive control algorithm, the pure deep reinforcement learning algorithm, and the traditional rule-based equilibrium algorithm in terms of real-time performance and optimization effect: Table 3. Comparison of real-time performance and optimization results between the hybrid algorithm based on deep reinforcement learning and model predictive control, the pure model predictive control algorithm, the pure deep reinforcement learning algorithm, and the traditional rule-based equilibrium algorithm.

[0164] As shown in Table 3, the pure deep reinforcement learning algorithm has a fast decision-making speed, but its optimization effect is generally poor; the pure model predictive control algorithm has a better optimization effect, but its computation time is long and its real-time performance is insufficient. The hybrid algorithm in this application first outputs the initial equilibrium strategy from the deep reinforcement learning module, and then performs rolling optimization by the model predictive control module. The average single-step decision time is 8 ms, which meets the real-time requirements of vehicle in-vehicle systems. At the same time, it has the shortest equilibrium completion time, the lowest average SOC deviation rate after equilibrium, the least energy loss, and no over-equilibrium occurs, thus achieving a balance between real-time performance and optimization.

[0165] This application's embodiments improve the comprehensiveness and reliability of the balanced control of new energy vehicle battery packs by combining an improved Kalman filter algorithm, a multi-dimensional equalization priority evaluation model, a hybrid algorithm based on deep reinforcement learning and model predictive control, and an adaptive duty cycle adjustment algorithm.

[0166] By establishing the state equation of the state vector and the observation equation of the observation vector, an adaptive forgetting factor is designed. Based on the state equation, the observation equation, and the adaptive forgetting factor, Kalman filtering iteration is performed to output the state parameters of each cell in the battery pack, thereby enhancing the adaptability of the filtering algorithm to the time-varying environment in the vehicle.

[0167] By constructing a mapping relationship between multi-dimensional denoised data and state parameters, including denoised vibration acceleration data and denoised aging state data, the estimation accuracy of the cell's state parameters is improved.

[0168] By constructing a multi-dimensional equilibrium priority assessment model based on state of charge deviation, health state deviation, voltage deviation, and temperature deviation, the comprehensive risk coefficient of each cell is calculated, avoiding the limitations of traditional single voltage index assessment and realizing the quantitative classification of equilibrium requirements.

[0169] The proposed hybrid algorithm based on deep reinforcement learning and model predictive control outputs an initial equilibrium policy through a deep reinforcement learning module and performs rolling optimization through a model predictive control module, thus resolving the contradiction between the poor real-time performance and insufficient optimization of traditional equilibrium policies.

[0170] By employing an adaptive duty cycle adjustment algorithm to calculate the initial duty cycle based on the operating mode and optimal balanced power, and then dynamically adjusting the duty cycle through a PID control algorithm, bidirectional energy flow is supported, thereby improving the balancing efficiency.

[0171] By constructing an evaluation index for the equilibrium effect, and using the particle swarm optimization algorithm to optimize the parameters in steps one, three, and five, this closed-loop optimization mechanism improves the long-term applicability of the method.

[0172] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0173] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A dynamic equalization control method for a new energy vehicle battery pack, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional data during the operation of the battery pack, including voltage data, current data, surface temperature data, vibration acceleration data, and aging status data; Step 2: Perform noise suppression on the multi-dimensional data to obtain multi-dimensional denoised data, including denoised voltage data, denoised current data, denoised surface temperature data, denoised vibration acceleration data, and denoised aging state data. Step 3: Based on the multi-dimensional denoised data, the improved Kalman filter algorithm based on the adaptive forgetting factor is used to estimate the state parameters of each cell in the battery pack, including state of charge, state of health, calibration voltage, and calibration temperature. Step 4: Based on the state parameters of each cell, construct a multi-dimensional equalization priority evaluation model to determine whether equalization needs to be initiated and identify high-priority cells. Step 5: Based on a hybrid algorithm of deep reinforcement learning and model predictive control, generate a dynamic balancing strategy for each high-priority cell, including optimal balancing power and optimal balancing time. Step 6: Based on the bidirectional buck-boost converter topology, an adaptive duty cycle adjustment algorithm is used to perform equalization control according to the dynamic equalization strategy.

2. The dynamic equalization control method for a new energy vehicle battery pack according to claim 1, characterized in that, In step two, a wavelet threshold denoising algorithm is used to suppress noise in the voltage data and the current data to obtain denoised voltage data and denoised current data. The surface temperature data is noise-suppressed by a moving average filtering algorithm to obtain denoised surface temperature data. The vibration acceleration data were noise-suppressed using the empirical mode decomposition algorithm to obtain denoised vibration acceleration data. The Gaussian filtering algorithm is used to suppress noise in the aging state data to obtain denoised aging state data.

3. The dynamic equalization control method for a new energy vehicle battery pack according to claim 1, characterized in that, In step three, the improved Kalman filter algorithm based on the adaptive forgetting factor includes: Establish the state equation for the state vector and the observation equation for the observation vector; Design an adaptive forgetting factor; Based on the state equation, the observation equation, and the adaptive forgetting factor, Kalman filtering iteration is performed to output the state parameters of each cell in the battery pack.

4. The dynamic balance control method for a new energy vehicle battery pack according to claim 1, characterized in that, Step three involves constructing the mapping relationship between the multi-dimensional denoised data and the state parameters, including: The state of charge is corrected based on the denoised voltage data, which is also used to calculate the calibration voltage. The fundamental change in the state of charge is calculated based on the denoised current data; The calibration temperature is calculated based on the noise-reduced surface temperature data; The health status is corrected based on the denoised vibration acceleration data and the denoised aging status data.

5. The dynamic equalization control method for a new energy vehicle battery pack according to claim 1, characterized in that, In step four, the corresponding state evaluation index is calculated based on the state parameters of each cell, including state of charge deviation, state of health deviation, voltage deviation, and temperature deviation. A multi-dimensional balanced priority evaluation model is constructed based on the state of charge deviation, the health state deviation, the voltage deviation, and the temperature deviation to calculate the comprehensive risk coefficient of each cell. When at least one cell in the battery pack has a comprehensive risk coefficient greater than a preset risk threshold, it is determined that equalization needs to be initiated. The cells are sorted from high to low according to their comprehensive risk coefficients, and the top few cells are selected as high-priority cells.

6. The dynamic equalization control method for a new energy vehicle battery pack according to claim 1, characterized in that, In step five, the hybrid algorithm based on deep reinforcement learning and model prediction control includes a deep reinforcement learning module and a model prediction control module. The deep reinforcement learning module is used to construct an intelligent agent, with the state parameters of high-priority cells as the state space, the equalization power and equalization time as the action space, and the minimum deviation of the state of charge after equalization and the minimum energy loss as the reward function, and outputs the initial equalization strategy. The model prediction and control module is used to establish a cell balancing response model. Taking the initial balancing strategy as input, it minimizes the objective function and performs rolling optimization to obtain the dynamic balancing strategy for each high-priority cell.

7. The dynamic equalization control method for a new energy vehicle battery pack according to claim 1, characterized in that, Step six includes: The operating mode of the bidirectional buck-boost converter is determined based on the status parameters of high-priority cells. An adaptive duty cycle adjustment algorithm is used to calculate the initial duty cycle based on the operating mode and the optimal balanced power. The duty cycle is dynamically adjusted using a PID control algorithm. When the optimal balancing time is reached, the balancing control of high-priority cells is stopped.

8. The dynamic equalization control method for a new energy vehicle battery pack according to claim 1, characterized in that, Step six also includes: Step 7: Evaluate the effectiveness of the balance control and optimize the parameters in steps 1, 3, and 5 based on the evaluation results.

9. The dynamic equalization control method for a new energy vehicle battery pack according to claim 8, characterized in that, In step seven, an evaluation index for the balancing effect is constructed based on the voltage consistency coefficient, the mean state-of-charge deviation rate, the balancing efficiency, and the temperature uniformity coefficient. The particle swarm optimization algorithm is used to adjust the data collection frequency in step one, the initial forgetting factor in step three, and the weight coefficient of the reward function in step five, with the optimization objective of maximizing the equilibrium effect evaluation index.

Images