Battery charging management optimization method based on neural network prediction

Abstract

The invention relates to a battery charging management optimization method based on neural network prediction, and the method comprises the steps: 1, obtaining the battery state data in the operation process of a battery, and enabling the battery state data to comprise the voltage, current, temperature and internal resistance of the battery; 2, constructing a time sequence according to the battery state data, and inputting the time sequence into a pre-trained fusion neural network model to obtain an optimal charging current prediction result; and 3, according to an optimal charging current prediction result, obtaining a dynamic safety protection threshold through a lightweight network model. The charging efficiency of the battery is improved by dynamically adapting the current intensity, so that the charging process always fits the current running state of the battery, and the over-charging and over-temperature risks caused by fixed parameters in a traditional constant-current and constant-voltage charging strategy are avoided; and meanwhile, the safety protection threshold is dynamically adapted according to the optimal charging current, so that the safety and high efficiency of the battery in the whole charging period are ensured.

Description

Technical Field

[0001] This invention relates to a battery charging management optimization method based on neural network prediction, which is applicable to battery charging control in scenarios such as energy storage systems, electric vehicle auxiliary power supplies, communication base station backup power supplies, and emergency microgrids, and belongs to the field of battery management technology. Background Technology

[0002] As a core component of various energy storage and power supply systems, the efficiency and safety of battery charging management directly affect the stable operation and service life of the entire system.

[0003] Batteries are widely used in critical scenarios such as energy storage systems, auxiliary power supplies for electric vehicles, backup power supplies for communication base stations, and emergency microgrids, becoming an important support for the stable operation of many fields. However, current charging management still generally adopts a fixed mode of constant current followed by constant voltage, and the safety protection threshold is always set to a constant value without being dynamically adjusted according to actual conditions. This traditional management mode has obvious limitations; it cannot fully release the charging potential of the battery, nor does it take into account the aging characteristics of the battery during charging, thus challenging the efficiency and safety of battery charging management.

[0004] Currently, the industry generally adopts a fixed charging mode of "constant current first, then constant voltage," and the safety protection thresholds (such as maximum charging current, maximum temperature, and maximum voltage) are all set to constant values. This does not take into account the dynamic state changes of the battery during the charging process, and has the following prominent technical problems:

[0005] Low charging efficiency: Fixed charging parameters cannot adapt to the dynamic state of the battery, such as battery aging, real-time temperature changes, and internal resistance fluctuations, resulting in significant energy loss and excessively long charging times. For example, a traditional 12V / 12AH battery pack takes 66 minutes to charge from 0% SOC to full capacity, and its charging efficiency urgently needs improvement.

[0006] Significant safety risks exist: Static safety thresholds ignore the dynamic changes in battery state during charging. When the battery ages, experiences localized overheating, or exhibits abnormal internal resistance, the fixed threshold cannot respond promptly, easily leading to overcharging, overheating, and other problems. In severe cases, this can cause battery bulging, fire, and even shorten battery life. Test data shows that under traditional mode, the maximum single-cell temperature can reach 51℃, and the maximum single-cell voltage can reach 13.9V, posing significant safety hazards.

[0007] Incomplete feature extraction: Existing charging control methods based on a single neural network can only extract local or global features from time-series data, and cannot simultaneously take into account the correlation between time-series dependencies and global features, resulting in insufficient accuracy in the prediction of the optimal charging current, which in turn affects the charging control effect.

[0008] Poor adaptability: Fixed charging strategies do not take into account the aging characteristics of the battery throughout its entire life cycle. As the number of battery cycles increases, charging performance gradually declines, and traditional methods cannot dynamically adjust parameters, further aggravating battery degradation. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention provides a battery charging management optimization method based on neural network prediction to improve the charging speed and charging safety of battery systems.

[0010] A battery charging management optimization method based on neural network prediction includes the following steps:

[0011] Step 1: Obtain battery status data during battery operation, including battery voltage, current, temperature, and internal resistance;

[0012] Step 2: Construct a time series based on the battery state data, and input the time series into a pre-trained fusion neural network model to obtain the optimal charging current prediction result. The fusion neural network model is a fusion structure of Transformer network and LSTM network.

[0013] Step 3: Based on the optimal charging current prediction results, obtain the dynamic safety protection threshold through a lightweight network model;

[0014] Step 4: Based on the optimal charging current prediction results and dynamic safety protection thresholds, change the battery charging topology control strategy to adjust the battery charging current and the safety thresholds during the charging process.

[0015] In step one, the acquisition of battery status data is accomplished by a high-precision operational amplifier sampling circuit. This circuit acquires the status data through sensors located at the cell interface and in the charging circuit. These sensors include those for acquiring temperature data T. c (t) is an NTC thermistor used to obtain the voltage data of a single battery cell, V. c (t) high-precision voltage divider resistors, used to acquire current data I c The current sensing resistor (t) acquires impedance state data Z in the range of 10Hz-100kHz via AC excitation. c (t) internal resistance detection module.

[0016] Step two specifically includes:

[0017] S2.1: Extract the global feature associations of the time series through the Transformer network with hidden state correction in the fusion neural network model;

[0018] S2.2: Extract the temporal dependency features in the temporal sequence through the LSTM network with attention feature correction in the fusion neural network model;

[0019] S2.3: Based on the global feature association and the time-dependent feature, generate the optimal charging current prediction result for the battery.

[0020] Step three specifically includes:

[0021] S3.1: Construct the optimal charging current and battery state data input vector;

[0022] S3.2: Based on the battery state data input vector, output a dynamic safety protection threshold using a lightweight MLP network.

[0023] Step four specifically includes:

[0024] S4.1: Based on the optimal charging current, obtain the battery equivalent load value;

[0025] S4.2: Determine the battery charging topology gain curve based on the battery equivalent load value and the battery dynamic safety protection threshold, adjust the PI controller parameters according to the gain curve, control the battery charging topology to output the optimal charging current, and set the dynamic safety protection threshold during the charging process.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] This invention acquires battery state data during battery operation, including voltage, current, temperature, and internal resistance. A time-series sequence is then constructed based on this data and input into a pre-trained fusion neural network model to obtain an optimal charging current prediction. Based on this prediction, a dynamic safety protection threshold is derived using a lightweight network model. The battery charging topology control strategy is then modified according to the optimal charging current prediction and the dynamic safety protection threshold, adjusting the charging current and safety thresholds during charging. This dynamically adapts the current intensity to improve charging efficiency, ensuring the charging process always matches the battery's current operating state and avoiding the overcharging and overheating risks associated with fixed parameters in traditional constant-current-then-constant-voltage charging strategies. Simultaneously, the dynamic adaptation of the safety protection threshold based on the optimal charging current ensures the battery's safety and efficiency throughout the entire charging cycle.

[0028] Significantly improved charging efficiency: By integrating neural networks to accurately predict the optimal charging current and dynamically adapt to the real-time state of the battery, the charging time is reduced by 44% compared to the traditional "constant current first, constant voltage second" mode (from 66 minutes to 37 minutes), 12% compared to a single Transformer network, and 31% compared to a single LSTM network, significantly improving charging efficiency.

[0029] Charging safety is significantly enhanced: the dynamic safety protection threshold can adapt to changes in battery status in real time, the maximum single cell temperature is controlled at 38℃ (25.5% lower than the traditional mode), the maximum single cell voltage is controlled at 13.6V (2.16% lower than the traditional mode), and the internal resistance at multiple frequency points is lower than that of the traditional mode, effectively avoiding the risks of overcharging and overheating, and extending battery life.

[0030] More comprehensive and accurate feature extraction: By combining the advantages of Transformer and LSTM networks, it captures both global feature associations and temporal dependent features, improving prediction accuracy compared to single network models and ensuring a high degree of matching between the optimal charging current and the battery state.

[0031] High adaptability and practicality: The dynamic adjustment mechanism can adapt to complex scenarios such as battery aging and changes in ambient temperature. The lightweight network design balances prediction accuracy and operating efficiency, making it easy to deploy in engineering. The distributed sensor layout and high-precision sampling circuit ensure the accuracy of data acquisition, providing reliable support for control strategies. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the battery charging management optimization method based on neural network prediction according to the present invention.

[0034] Figure 2 This is a schematic diagram illustrating the correction process of the fused neural network in the battery charging management optimization method based on neural network prediction of the present invention.

[0035] Figure 3 This diagram illustrates the optimal charging current output of the battery charging management optimization method based on neural network prediction according to the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Reference Figure 1 A battery charging management optimization method based on neural network prediction includes the following steps:

[0038] Step 1: Obtain battery status data during battery operation, including battery voltage, current, temperature, and internal resistance;

[0039] Battery status data acquisition is accomplished by a high-precision operational amplifier sampling circuit. This circuit acquires status data through sensors located at the cell interface and in the charging circuit. These sensors include NTC thermistors, high-precision voltage divider resistors, current sensing resistors, and internal resistance detection modules. The NTC thermistor is used to acquire temperature data T. c (t), high-precision voltage divider resistors are used to obtain the voltage data V of a single battery cell. c (t), the current sensing resistor is used to acquire current data I. c (t)(including charging and discharging current), the internal resistance detection module acquires impedance state data Z in the range of 10Hz-100kHz through AC excitation. c (t)(impedance values ​​at multiple frequency points), and this data will be periodically transmitted to the computer.

[0040] The NTC thermistors are distributed. For a battery module consisting of 12 cells in series, 12 NTC thermistors are set up at the interfaces of different cells to acquire temperature data. A high-precision operational amplifier sampling circuit returns the temperature data to the computer for acquisition. The computer's sampling frequency is set to 10Hz, therefore the temperature data T c (t) is a 12-dimensional time series, for example, at time t, T c (t) can be represented as [T c1 (t), T c1 (t), T c3 (t), T c4 (t), T c5 (t), T c6 (t), T c7 (t), T c8 (t), T c9 (t), T c10 (t), T c11 (t), T c12[(t)], where each element corresponds to the temperature value of the battery cell at time t, in degrees Celsius.

[0041] High-precision voltage divider resistors are used to collect voltage data for each battery cell. A high-precision operational amplifier sampling circuit returns temperature data to the computer for acquisition. The computer's sampling frequency is 20Hz. For a battery module composed of 12 cells in series, the voltage data V... c (t) can be represented as [V c1 (t), V c2 (t), ..., V c12 (t), V total [(t)], the unit is volt, where V c1 (t) to V c12 (t) represents the voltage of each individual cell at time t, V. total (t) represents the total voltage of the module, therefore the voltage data V c (t) is a 13-dimensional time series.

[0042] The current sensing resistor is used to acquire current state data I. c (t), the high-precision operational amplifier sampling circuit returns the current data to the computer for acquisition. The computer's sampling frequency is 20Hz. The current data is the charging current. For a battery cluster composed of 12 cells, its data form is a one-dimensional time series. However, since the sampling time is continuous, it forms a set of sequences with multiple time points, such as [I(t-2Δt), I(t-Δt), I(t)], in amperes, where Δt is the sampling interval, and here Δt = 0.05s (corresponding to a sampling frequency of 20Hz).

[0043] The internal resistance detection module acquires impedance data Z via AC excitation. c (t), Z can be obtained in the range of 10Hz to 100kHz. c (f) Curve, with a sampling frequency of 1Hz, acquires an impedance sequence containing multiple frequency points every second, therefore Z c Z(t) is a multi-dimensional sequence, for example, at time t, Z c (t) can be represented as [Z c (10Hz, t), Z c (100Hz, t), Z c (1kHz, t), Z c (10kHz, t), Z c [(100kHz, t)], in ohms, includes impedance values ​​at different frequencies. The impedance sequence is ultimately obtained directly by the computer.

[0044] Step 2: Construct a time series based on the battery state data, and input the time series into a pre-trained fusion neural network model to obtain the optimal charging current prediction result. The fusion neural network model is a fusion structure of Transformer network and LSTM network.

[0045] First, a time series is constructed based on the battery state data. The battery state data includes voltage data V. c (t), Current data I c (t), Temperature data T c (t), internal resistance data Z c (t). First, a unified time scale is needed, such as a sampling frequency of 10Hz. V c (t) is the voltage time series, I c (t) represents the current time series, with a sampling frequency of 20Hz, which is down-frequencyed to 10Hz by averaging the samples. c (t) represents the internal resistance sequence, sampled at a frequency of 1Hz, and upscaled to 10Hz via linear interpolation. Next, a synchronization timing window is constructed with a window length (seq_len) of 50 time steps (corresponding to 5 seconds of data) and a window sliding step size of 1 time step (corresponding to 100ms). The unified data is then concatenated with the window length, which can be represented as [seq_len, T...]. c1_u (t), ..., T c12_u (t), V c1_u (t), ..., V c12_u (t), V total_u (t), I c_u (t), Z c_u (10Hz, t), Z c_u (100Hz, t), Z c_u (1kHz, t), Z c_u (10kHz, t), Z c_u (100kHz, t)]. Then, the voltage, temperature, current, and internal resistance features are normalized by Z-score normalization to obtain the original target time series X, which includes the window length and 31 feature dimensions.

[0046]

[0047] Where X is the original target time series, and μ and σ are the mean and standard deviation of the feature.

[0048] Next, the time series is input into a pre-trained fusion neural network model. This fusion neural network model employs a fusion structure of Transformer and LSTM networks.

[0049] Step two specifically includes:

[0050] S2.1: Extract the global feature associations of the time series through the Transformer network with hidden state correction in the fusion neural network model;

[0051] The Transformer network includes input feature mapping, positional encoding, multi-head attention mechanism, hidden state correction, and encoder layer stacking. The Transformer network requires the input model dimension to be the same as the model's internal dimension; for example, if the model's internal dimension is 64, a linear layer maps the 31-dimensional original features to 64 dimensions. Positional encoding allows the Transformer network to perceive the temporal relationship of battery data. The multi-head attention mechanism calculates the attention weights of each time step with all other time steps, measuring the strength of dependencies between features. The original hidden state of the LSTM network assigns higher weights to key battery cells when calculating the original attention score to correct the original attention score. For example, choosing an 8-head attention mechanism maps the input features to Q (query), K (key), and V (value) through three linear layers. Each head independently calculates its attention weight. When calculating these attention weights, the original attention score is first corrected using the original hidden state, then the corrected attention weights for each head are calculated, and finally, the attention weights are concatenated. Stacking encoder layers can enhance the feature extraction capability of a single attention layer. For example, stacking two encoder layers can deepen the features. With the number of iteration samples set to 32, a global feature association summary that integrates all historical time step information of voltage, current, temperature, and internal resistance data within the window can be obtained.

[0052] For example, the input feature map contains a linear layer with an input dimension of 31 and an output dimension of 64.

[0053] Y X_N =Linear TS (Y X )

[0054] Among them, Linear TS () represents a linear layer, Y X These are the original input features after normalization.

[0055] The positional encoding adopts a sine-cosine positional encoding method, and the positional encoding is added to the mapped feature sequence, preserving the original features while injecting temporal positional information.

[0056] Y P =PE(Y X_N )

[0057] Y X_N_P =Y X_N +Y P

[0058] Where PE() is the sine / cosine position code, Y X_N Y is the input feature after mapping. P For position encoding.

[0059] The multi-head attention mechanism selects three linear layers to map the input to Q, K, and V. The 64-dimensional Q, K, and V are split into eight heads, each with eight dimensions.

[0060] Q Sp =Split(Linear) Q (Y X_N_P ))

[0061] K Sp =Split(Linear) K (Y X_N_P ))

[0062] V Sp =Split(Linear) V (Y X_N_P ))

[0063] Split() is the splitting function, Linear Q (), Linear K (), Linear V () represents a linear layer. Y X_N_P The input features are mapped with added location information.

[0064] The original attention score is calculated to measure the degree of matching between each Q and all K. The scores of the most critical battery cells in the original hidden features of the attention feature-corrected LSTM network are extracted as the original attention score correction parameters.

[0065]

[0066] Where T is the matrix transpose, Q Sp K Sp represents the Q and K values ​​of the 8 heads after splitting. s represents the most critical score among the original hidden features.

[0067] The corrected attention score is then normalized to obtain the attention weight. The attention weight is then used to perform a weighted summation on V to obtain the final single-head attention output.

[0068] Attention Sp (Q Sp ,K Sp V Sp = softmax(scores) n V Sp

[0069] scores n The adjusted attention score is represented by softmax(), which is a regression function. Sp The V value is the result of splitting the 8 heads.

[0070] Adjust the order of the head dimensions to concatenate single-head dimensions into a multi-head dimension output.

[0071] Attention(Q,K,V)=reshape(Attention Sp (Q Sp ,K Sp V Sp ))

[0072] Where reshape() is the concatenation function, Attention Sp (Q Sp ,K Sp V Sp ) represents the single-head attention output after splitting.

[0073] The multi-dimensional output becomes the output of an encoder layer after passing through a feedforward network (FFN).

[0074] EnCoder=FFN(Attention(Q,K,V))

[0075] FFN() is a feedforward network, and Attention(Q,K,V) is the multi-head attention output.

[0076] Two of the above encoder layers are stacked to form a feature enhancement pipeline, where the output of the previous layer is the input of the next layer. The output of the stacked encoder is extracted, normalized, and then the features from the last time step are taken as the final output of the Transformer branch.

[0077] TransFormer=EnCoder[EnCoder(Y i )]

[0078] EnCoder() is the encoder layer, which includes multi-head self-attention computation and a feedforward network. i Input any original feature.

[0079] The Transformer network output represents a summary of global correlations within the entire window, while the original hidden features of the LSTM network allow global correlations to be anchored to local time series. For example, the impact on the voltage at the current time is only considered when the temperature change at the previous time step exceeds a certain value.

[0080] S2.2: Extract the temporal dependency features in the temporal sequence through the LSTM network with attention feature correction in the fusion neural network model;

[0081] The core of LSTM is to memorize or forget temporal information through a gating mechanism. After adjusting the feature dimensions of the obtained temporal sequence, a two-layer unidirectional LSTM network is output. The two layers can capture more complex long-term trends, while the unidirectional approach avoids introducing future information leakage. The two-layer processing includes the original hidden state of the entire sequence. The attention weights are calculated from the original attention scores of the Transformer network and concatenated to the attention output, which corrects the original hidden state and weakens the local judgment of the original hidden features.

[0082] For example, the obtained time series sequence is passed through a linear layer to obtain a 64-dimensional feature sequence, which is then used as the input to an LSTM network, whose hidden dimension is 64.

[0083] Y X_T =Linear LT (Y X )

[0084] Among them, Linear LT () represents a linear layer, Y X These are the original input features.

[0085] Use a 2-layer unidirectional LSTM network.

[0086] Y X_T_LSTM =LSTM_2D(Y X_T )

[0087] Where LSTM_2D() is a 2-layer unidirectional LSTM network, Y X_T The input features are adjusted for dimensionality.

[0088] Finally, the original hidden state containing the entire sequence is obtained, and the attention output of the Transformer network is extracted and compressed into weights of the same dimension through convolution to correct the original hidden features.

[0089] w = Conv(Att)

[0090] Y X_T_LSTM_N =Y X_T_LSTM ×(1-w)

[0091] Where Y X_T_LSTM is the original hidden feature, w is the attention correction parameter, Conv() is the convolution function, and Att is the attention output.

[0092] Take the features from the last time step and perform Dropout regularization and layer normalization on them.

[0093] LSTM = LayerNorm(Dropout(Y X_T_LSTM_N ))

[0094] Dropout() is the regularization function, LayerNorm() is the layer normalization function, and Y... X_T_LSTM_N This is the corrected hidden state that includes the entire sequence.

[0095] The LSTM branch output integrates information from all historical time steps within the entire window, representing long-term temporal dependencies between data at different times. Meanwhile, the attention output of the Transformer network allows the temporal dependencies to avoid the influence of local noise, such as not considering the short-term effects of current fluctuations on temperature when the sensor malfunctions.

[0096] S2.3: Based on the global feature association and the time-dependent feature, generate the optimal charging current prediction result for the battery.

[0097] By simply splicing together the local temporal trend output by the LSTM network and the global feature output by the Transformer network, the fused features are spliced ​​together according to the feature dimension and then the fully connected layer learns autonomously to obtain the charging current.

[0098] For example, the output features of the LSTM network and the Transformer network are normalized to a distribution with a mean of 0 and a variance of 1 through layers.

[0099] LSTM norm =LayerNorm LT (LSTM)

[0100] TransFormer norm =LayerNorm TS (TransFormer)

[0101] LSTM and Transformer are branch outputs, and LayerNorm is used for branch outputs. LT (),LayerNorm TS () is the layer normalization function.

[0102] The two branches of normalized features are concatenated directly with feature dimension 1. The outputs of both branches are 32 samples and 64 feature dimensions. After concatenation, the sample dimension remains unchanged, but the feature dimension becomes 128. The first 64 are LSTM features, and the last 64 are Transformer features.

[0103] Y F =concatenate(LSTM) normTransFormer norm )

[0104] LSTM norm TransFormer norm The `concatenate()` function is used to concatenate the outputs of different network branches after normalization.

[0105] After concatenation, the features are compressed and nonlinearly enhanced by a lightweight fully connected network (FC layer). The FC layer includes two linear layers, which compress the feature dimension to 32. The ReLU activation function is set between the linear layers to simulate the nonlinear mapping between battery state and charging current, and Dropout is set to 0.2.

[0106] Y FL =Linear FC (Y F )

[0107] Y FL_RL =ReLU(Y FL )

[0108] Y FL_RL_DP =Dropout FC (Y FL_RL )

[0109] Where Y F To output different network branches after splicing, Linear FC () linear layer, ReLU() is the activation function, Dropout FC () is the regularization function.

[0110] The compressed and nonlinearly enhanced features are mapped to 1D features through a linear layer to form the optimal charging current result.

[0111] Y O =Linear O (Y FL_RL_DP )

[0112] Where Y FL_RL_DP This represents the feature after compression and nonlinear enhancement. O Result of 1D optimal charging current

[0113] Step 3: Based on the optimal charging current prediction results, obtain the dynamic safety protection threshold through a lightweight network model;

[0114] Step three specifically includes:

[0115] S3.1: Construct the optimal charging current and battery state data input vector;

[0116] The optimal charging current is the optimal charging current Y obtained in S2.3. O The battery state data includes the voltage data V after unifying the time scale in step S20. c_u (t), Temperature data T c_u (t), internal resistance data Z c_u (t). The battery state data and the optimal charging current are concatenated into a 31-dimensional input vector, which can be represented as [Y]. O V c1_u (t), ..., V c12_u (t), V total_u (t), T c1_u (t), ..., T c12_u (t), Z c_u (10Hz, t), Z c_u (100Hz, t), Z c_u (1kHz, t), Z c_u (10kHz, t), Z c_u (100kHz, t)]. Then, the optimal charging current, voltage, temperature, and internal resistance features are normalized using Z-score normalization to obtain the original input vector X with 31 feature dimensions. g .

[0117]

[0118] Where X g Let μ be the original input vector, and σ be the mean and standard deviation of the input vector.

[0119] S3.2: Based on the battery state data input vector, output a dynamic safety protection threshold using a lightweight MLP network.

[0120] The lightweight MLP network comprises feature extraction, bottleneck compression, cross-cell correlation, and a physical constraint layer. To extract features from different physical quantities, the lightweight MLP network needs to split the input dimension while enhancing detailed correlations. For example, the input vector is split into three branches (voltage, temperature, and current + internal resistance), and these three branches are expanded to 32 dimensions. Bottleneck compression filters redundant features and reduces the number of parameters in subsequent layers. Cross-cell correlation captures the coupling relationship between cells, allowing the output cell thresholds to be adjusted accordingly. The physical constraint layer forces the output 31-dimensional thresholds to conform to the battery's physical characteristics and safety rules. Finally, a battery charging safety protection threshold is obtained that conforms to physical characteristics and safety rules and is correlated with optimal charging current, voltage, temperature, and internal resistance data.

[0121] For example, the input vector is split into 3 branches.

[0122]

[0123] The divide() function is used to split the data. Let be any original input vector after normalization.

[0124] Each branch is expanded to 32 dimensions after feature extraction.

[0125] [Y V_P ,Y T_P ,Y IZ_P ] = FeaturePick(Y V ,Y T ,Y IZ )

[0126] Where FeaturePick() is the feature extraction function, Y V ,Y T ,Y IZ These are the three branches after the split.

[0127] The results of the three branches are concatenated to obtain a 96-dimensional vector, which is then compressed to a 48-dimensional vector through bottleneck compression.

[0128] Y V_T_IZ =concatenate(Y V_P ,Y T_P ,Y IZ_P )

[0129] Where concatenate() is the concatenation function, Y V_P ,Y T_P ,Y IZ_P These are the three branches after feature extraction.

[0130] Cross-monoassociation reduces the compressed vector to 32 dimensions.

[0131] Y V_T_IZ_Cr =Cross(Y V_T_IZ )

[0132] Cross() is the correlation function, Y V_T_IZ This is the concatenated vector.

[0133] The physical constraint layer includes a linear layer that maps a 32-dimensional vector to a 31-dimensional original dynamic protection threshold.

[0134] Y g_Ori =Linear PhyCons (Y V_T_IZ_Cr )

[0135] The constraint rules are dynamically adjusted using the original input parameters, and the constraint process is implemented using tensor truncation (clamp).

[0136] [Y C_V ,Y C_Vtot ,Y C_T ,Y C_I ,Y C_Z ] = Pick(X g )

[0137] [Y V_Ori ,Y T_Ori ,Y I_Ori ,Y R_Ori ] = Pick(Y g_Ori )

[0138] Y V_thre =Clamp(Y C_V ,Y V_Ori )

[0139] Y Vtot_thre =Clamp(Y C_Vtot Sum(Y) V_Ori ))

[0140] Y T_thre =Clamp(Y C_T ,Y T_Ori )

[0141] Y I_thre =Clamp(Y C_I ,Y I_Ori )

[0142] Y Z_thre =Clamp(Y C_Z ,Y Z_Ori )

[0143] Pick() is the extraction function, Clamp() is the tensor truncation function, and Y... g_Ori Y is the original dynamic protection threshold. C_V ,Y C_Vtot ,Y C_T ,Y C_I ,Y C_Z These represent the voltage, temperature, current, and internal resistance values ​​extracted from the original vector, respectively, and Y. V_Ori ,Y T_Ori ,Y I_Ori ,Y Z_Ori These are the voltage, temperature, current, and internal resistance thresholds extracted from the original dynamic protection thresholds.

[0144] The final output is a 31-dimensional battery charging safety protection threshold, including the maximum charging current Y. I_thre Maximum single-cell voltage Y V_thre Maximum cluster voltage Y Vtot_thre Maximum single-cell temperature Y T_thre Maximum internal resistance Y at different frequenciesZ_thre .

[0145] Step 4: Based on the optimal charging current prediction results and dynamic safety protection thresholds, change the battery charging topology control strategy to adjust the battery charging current and the safety thresholds during the charging process.

[0146] Step four specifically includes:

[0147] S4.1: Based on the optimal charging current, obtain the battery equivalent load value;

[0148] The equivalent load value of the battery is calculated based on the predicted optimal charging current and the current voltage of the battery cluster consisting of 12 cells. The calculation formula is as follows:

[0149] R = V total (t) / Y O

[0150] Where V total (t) represents the initial battery cluster voltage, Y O The optimal charging current is given by R, which represents the current equivalent load on the battery.

[0151] S4.2: Determine the battery charging topology gain curve based on the battery equivalent load value and the battery dynamic safety protection threshold, adjust the PI controller parameters according to the gain curve, control the battery charging topology to output the optimal charging current, and set the dynamic safety protection threshold during the charging process.

[0152] The battery charging topology gain curve is related to the battery's equivalent load value.

[0153] CurseM = f(R)

[0154] Where f() is the functional relationship between the battery charging topology gain curve and the battery equivalent load value, and R is the current battery equivalent load.

[0155] Based on the battery charging topology gain curve, the PI controller parameters are adjusted according to the gain curve to control the battery charging topology to output the optimal charging current.

[0156] [Kp,Ki]=CurseM(Y O )

[0157] e(t) = Y O -I(t)

[0158] I(t+1)=I(t)+Kp(e(t)-e(t-1))+Kie(t)

[0159] Where t represents the current time, I(t) represents the charging current at the current time, I(t+1) represents the charging current at the next time, e(t) represents the current error at the current time, e(t-1) represents the current error at the previous time, Kp and Ki are the proportional and integral coefficients of the corresponding charging gain curves, and Y... O This is the optimal charging current.

[0160] Set the dynamic safety protection threshold during the charging process.

[0161] V(t)≤Y V_thre (t)

[0162] V tot (t)≤Y Vtot_thre (t)

[0163] T(t)≤Y T_thre (t)

[0164] I(t)≤Y I_thre (t)

[0165] Z(t)≤Y Z_thre (t)

[0166] Where V(t), V tot Y(t), T(t), I(t), and Z(t) represent the individual cell voltage, cell cluster voltage, individual cell temperature, battery charging current, and individual cell internal resistance during the charging process, respectively. V_thre (t), Y Vtot_thre (t), Y T_thre (t), Y I_thre (t), Y Z_thre (t) represents the current battery cell voltage, battery cluster, individual cell temperature, battery charging current, and individual cell internal resistance protection threshold.

[0167] In this embodiment, based on the proposed charging management optimization method, the charging time and battery state data of three other different methods were tested. The charging battery cluster parameters were 12 batteries connected in series, with a single battery rated voltage of 12V, a rated capacity of 12AH, an ambient temperature of 25 degrees Celsius, and an initial SOC value of 0%. The charging time and battery state data results are shown in Table 1 below:

[0168] Table 1

[0169]

[0170] As shown in Table 1, the charging management optimization method proposed in this study can effectively accelerate the charging time of battery clusters. While maintaining a shorter charging time, compared to single Transformer networks, single LSTM networks, and traditional constant-current followed by constant-voltage charging methods, the battery state during charging is safer. For example, the temperature is lower, the individual cell voltage is lower, and the maximum internal resistance of the individual cell at different frequencies is also lower. Simultaneously, a dynamic safety threshold is provided, fully considering the aging characteristics of the battery and charging efficiency, ensuring charging safety while improving charging efficiency.

[0171] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A battery charging management optimization method based on neural network prediction, characterized in that: Includes the following steps: Step 1: Obtain battery status data during battery operation, including battery voltage, current, temperature, and internal resistance; Step 2: Construct a time series based on the battery state data, and input the time series into a pre-trained fusion neural network model to obtain the optimal charging current prediction result. The fusion neural network model is a fusion structure of Transformer network and LSTM network. Step 3: Based on the optimal charging current prediction results, obtain the dynamic safety protection threshold through a lightweight network model; Step 4: Based on the optimal charging current prediction results and dynamic safety protection thresholds, change the battery charging topology control strategy to adjust the battery charging current and the safety thresholds during the charging process.

2. The battery charging management optimization method based on neural network prediction according to claim 1, characterized in that: In step one, the acquisition of battery status data is accomplished by a high-precision operational amplifier sampling circuit. This circuit acquires the status data through sensors located at the cell interface and in the charging circuit. These sensors include those for acquiring temperature data T. c (t) is an NTC thermistor used to obtain the voltage data V of a single battery cell. c (t) high-precision voltage divider resistors, used to acquire current data I c The current sensing resistor (t) is used to acquire impedance state data Z in the range of 10Hz-100kHz via AC excitation. c (t) internal resistance detection module.

3. The battery charging management optimization method based on neural network prediction according to claim 1, characterized in that: Step two specifically includes: S2.1: Extract the global feature associations of the time series through the Transformer network with hidden state correction in the fusion neural network model; S2.2: Extract the temporal dependency features in the temporal sequence through the LSTM network with attention feature correction in the fusion neural network model; S2.3: Based on the global feature association and the time-dependent feature, generate the optimal charging current prediction result for the battery.

4. The battery charging management optimization method based on neural network prediction according to claim 1, characterized in that: Step three specifically includes: S3.1: Construct the optimal charging current and battery state data input vector; S3.2: Based on the battery state data input vector, output a dynamic safety protection threshold using a lightweight MLP network.

5. The battery charging management optimization method based on neural network prediction according to claim 1, characterized in that: Step four specifically includes: S4.1: Based on the optimal charging current, obtain the battery equivalent load value; S4.2: Determine the battery charging topology gain curve based on the battery equivalent load value and the battery dynamic safety protection threshold, adjust the PI controller parameters according to the gain curve, control the battery charging topology to output the optimal charging current, and set the dynamic safety protection threshold during the charging process.

Images