Energy storage battery module simulation method and system based on machine learning

Abstract

The invention relates to the technical field of energy storage battery simulation, and discloses an energy storage battery module simulation method and system based on machine learning. The method comprises the steps of collecting battery signal circuit data of a BMS system in charging and discharging circulation of an energy storage battery pack, extracting voltage fluctuation characteristics and current ripple characteristics of an energy storage high-voltage box, and constructing a multi-dimensional dynamic parameter matrix reflecting the internal state of a battery according to time sequence alignment; carrying out dimensionality reduction compression on the matrix through an encoder of the generative adversarial network to generate a battery state hidden variable; dynamic evolution modeling is carried out on hidden variables by utilizing a time sequence convolutional network, and hidden variable tracks of three charge and discharge cycles in the future are predicted; and performing reverse decoding to generate a simulation battery signal, and outputting a simulation data flow containing voltage, current and temperature parameters. According to the method, the multi-dimensional features and the time sequence modeling capability are fused, accurate simulation of the running state of the energy storage battery module is achieved, and the method is suitable for diversified application scenes.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery simulation technology, specifically to a machine learning-based method and system for simulating energy storage battery modules. Background Technology

[0002] With the rapid development of the new energy industry, energy storage batteries, as core components for energy storage and dispatch, are widely used in renewable energy grid connection, electric vehicles, microgrids, and other fields. The operating status of energy storage battery modules is directly related to the safety and reliability of the entire energy storage system. Therefore, accurate simulation of battery modules to model their dynamic response under different charging and discharging conditions has become an important aspect of energy storage technology research and development and application.

[0003] Traditional energy storage battery simulation methods are mostly based on physicochemical models, achieving simulation by establishing mechanistic equations such as electrochemical reactions and heat conduction within the battery. These methods require precise acquisition of a large amount of fundamental data, including the battery's material and structural parameters. However, in real-world applications, the battery's internal parameters dynamically change with increasing charge-discharge cycles and variations in ambient temperature, leading to decreased adaptability of the physical model and difficulty in guaranteeing simulation accuracy. Furthermore, the establishment of physical models is complex and computationally intensive, making it difficult to meet the demands of real-time simulation and rapid iteration.

[0004] In recent years, data-driven simulation methods have gradually become a research hotspot in the field of energy storage battery simulation due to their advantages of not relying on complex mechanisms and strong adaptability. However, existing data-driven methods still have many shortcomings: feature extraction is mostly focused on single electrical parameters, ignoring the synergistic effect of voltage and current features, making it difficult to comprehensively reflect the dynamic changes inside the battery; the temporal correlation of features is not fully considered, resulting in the constructed parameter matrix failing to accurately capture the temporal evolution of the battery state; when faced with high-dimensional battery signal data, traditional dimensionality reduction methods are prone to losing key dynamic information, affecting the accuracy of subsequent modeling; the time-series prediction capability is limited, mostly only able to achieve short-term single-cycle state prediction, failing to meet the actual needs of multi-cycle continuous simulation. In addition, the simulation data generated by existing methods often has single parameters, making it difficult to fully reproduce the comprehensive operating state of the battery under complex operating conditions, limiting the practical application value of the simulation results. These problems make it difficult for existing energy storage battery module simulation methods to balance accuracy, efficiency, and practicality, and cannot fully meet the high requirements of simulation technology in scenarios such as energy storage system design and operation and maintenance optimization. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based simulation method and system for energy storage battery modules to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a machine learning-based simulation method for energy storage battery modules, the method comprising: Collect battery signal circuit data of BMS system during charge and discharge cycles of energy storage battery pack, and extract voltage fluctuation characteristics and current ripple characteristics of energy storage high voltage box; By aligning voltage fluctuation characteristics and current ripple characteristics according to time series, a multi-dimensional dynamic parameter matrix reflecting the internal state of the battery is constructed. An encoder using an adversarial generative network is used to reduce the dimensionality of the multidimensional dynamic parameter matrix and generate latent battery state variables in the latent space. The battery state latent variables are dynamically evolved and modeled using a temporal convolutional network to predict the trajectory of latent variables for the next three charge-discharge cycles. The simulated battery signal is generated by reverse decoding based on the predicted latent variable trajectory, and the output is a simulated data stream containing voltage, current and temperature parameters.

[0007] Preferably, the battery signal circuit data of the BMS system includes the individual cell voltage range, inter-cluster temperature gradient, DC internal resistance change rate, polarization voltage decay, and capacity decay slope.

[0008] Preferably, the specific steps for constructing the multidimensional dynamic parameter matrix are as follows: The range of individual voltages is standardized using a sliding window, and the coefficient of variation of the range values ​​within the window is calculated. A temperature field distribution heatmap is generated based on the inter-cluster temperature difference gradient, and the area ratio of high-temperature regions in the heatmap is extracted. The DC internal resistance change rate and the polarization voltage decay are convolved in the time domain to generate a composite aging feature vector. By combining the capacity decay slope with the area ratio of the high-temperature region, a battery health status indicator is formed. The coefficient of variation, composite aging feature vector, and battery health status indicators are concatenated into a dynamic parameter matrix by timestamp.

[0009] Preferably, the latent variables of the battery state in the generated potential space include: Five residual convolutional modules are set in the encoder of the generative adversarial network, and a channel attention mechanism is connected after each convolutional layer. Multi-scale feature extraction of multi-dimensional dynamic parameter matrices is performed using spatial pyramid pooling layers; Bidirectional long short-term memory networks are used to capture the temporal dependencies of dynamic parameter matrices; Multi-scale features and temporal features are concatenated and then input into a fully connected layer to output 128-dimensional battery state latent variables.

[0010] Preferably, the step of predicting the latent variable trajectory for the next three charge-discharge cycles includes: Use dilated causal convolutional layers of temporal convolutional networks to capture long-term dependencies of latent variables; A gated recurrent unit network is superimposed at the output of the convolutional layer to establish a hidden variable state transition equation; The distribution of latent variable weights for future time steps is calculated using a multi-head self-attention mechanism; Monte Carlo sampling is used to generate multiple latent variable evolution paths, and the path with the highest probability density is selected as the predicted trajectory.

[0011] Preferably, the step of generating the simulated battery signal by reverse decoding based on the predicted latent variable trajectory includes: The predicted trajectory is input into the decoder of the adversarial generative network, and the signal dimension is gradually recovered through transposed convolutional layers; The rated parameter constraints of the energy storage high-voltage box are integrated during the decoding process; The gradient propagation of the decoder is stabilized using spectral normalization techniques; The output is connected to a sigmoid activation function to generate a standardized simulation signal.

[0012] Preferably, the method further includes: Real-time monitoring of the dynamic time-warped distance between simulated and measured signals; When the dynamic time warp distance exceeds the tolerance threshold, the retraining mechanism of the adversarial generative network is triggered. During retraining, the parameters of the temporal convolutional network are frozen, and only the weights of the adversarial generative network are updated.

[0013] Preferably, the specific steps of the retraining mechanism are as follows: Select charge / discharge cycle segments from historical data that match the current operating conditions; Construct an anomalous sample set containing voltage surge events and temperature surge events; The focus loss function is used to strengthen the training weights of outlier samples; Transfer learning allows the low-level feature extraction capabilities of a pre-trained model to be transferred to a new model.

[0014] Preferably, the method further includes: Establish a mapping library between simulation signals and the actual operating state of energy storage battery packs; When abnormal fluctuations in polarization voltage are detected in the simulation signal, the fault modes in the mapping relationship library are automatically associated. Generate a diagnostic report that includes fault location codes and severity levels.

[0015] Preferably, the present invention also includes a machine learning-based energy storage battery module simulation system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the machine learning-based energy storage battery module simulation method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Targeted acquisition of battery signal circuit data from the BMS system is employed, specifically extracting voltage fluctuation and current ripple features. These two types of features reflect the battery's energy transfer stability and circuit operating status, respectively. Their coordinated extraction comprehensively captures key dynamic information during battery operation, avoiding the limitation of single features failing to cover the complex internal states of the battery. After aligning these two types of features by time series, a multi-dimensional dynamic parameter matrix is ​​constructed, ensuring that each element in the matrix corresponds to the battery state at a specific time point. This accurately reconstructs the dynamic evolution trajectory of the battery state during charging and discharging, allowing subsequent modeling to be built upon data more closely aligned with actual operating conditions.

[0017] An encoder employing a generative adversarial network (GAN) is used for dimensionality reduction and compression. Compared to traditional dimensionality reduction methods, this encoder effectively removes redundant components from the data while preserving key dynamic information. The generated latent variables of the battery state reduce data dimensionality while fully retaining the core features of the battery's internal state. This reduces the computational burden for subsequent time-series modeling while ensuring modeling accuracy. Furthermore, the representation of latent variables in the latent space can overcome the dimensionality limitations of the original data, more intuitively reflecting the essential laws of battery state and providing a better input foundation for dynamic evolution modeling.

[0018] By using a temporal convolutional network to dynamically model the latent variables of battery state, this network possesses a strong ability to capture long-term dependencies, accurately identifying the changes in latent variables over charge-discharge cycles and achieving reliable predictions of the latent variable trajectories for the next three charge-discharge cycles. This multi-cycle forward-looking prediction can cover the complete dynamic process of battery short-term operation. Compared to methods that can only achieve single-cycle prediction, it better meets the actual needs of energy storage systems for long-term battery state assessment and can provide more valuable simulation data for the operation planning of battery modules.

[0019] The simulation data stream, which includes voltage, current, and temperature parameters, is generated by reverse decoding based on the predicted trajectory. These three types of parameters comprehensively cover the core indicators of battery operation, fully reproducing the overall operating state of the battery in future charge-discharge cycles. The simulation data stream can be directly applied to various scenarios such as performance testing, fault prediction, and control strategy optimization of battery modules. Whether in the laboratory R&D stage or in actual engineering applications, it can provide realistic simulation data support for related work. Furthermore, this method is entirely data-driven, without relying on complex physicochemical mechanism equations. It can adapt to different types and aging levels of energy storage battery modules, possessing broad applicability and effectively adapting to diverse charge-discharge conditions and energy storage application scenarios, thus improving the flexibility and practicality of energy storage battery module simulation. Attached Figure Description

[0020] Figure 1 This is a time-series characteristic diagram of battery temperature; Figure 2 A flowchart for constructing a multidimensional dynamic parameter matrix; Figure 3 This is a graph showing the trend of residual convolution module features over time. Figure 4 This is a flowchart for predicting the trajectory of latent variables. Detailed Implementation

[0021] 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.

[0022] Please see Figure 1 This invention provides a machine learning-based simulation method for energy storage battery modules, the method comprising: Efficient simulation is achieved by integrating battery management system signal processing and machine learning algorithms. Battery signal circuit data from the BMS system generated during charge-discharge cycles of the energy storage battery pack are acquired in real time, containing rich battery state information. Voltage fluctuation and current ripple features of the high-voltage storage box are extracted from the BMS system battery signal circuit data. Voltage fluctuation features reflect the dynamic changes in battery terminal voltage, while current ripple features capture high-frequency components in the charge-discharge current. The voltage fluctuation and current ripple features are aligned according to time series to form a data stream with consistent timestamps. A multi-dimensional dynamic parameter matrix reflecting the internal state of the battery is constructed by combining multiple feature dimensions. This multi-dimensional dynamic parameter matrix is ​​fed as input to the encoder part of the generative adversarial network (GAN). The encoder uses a deep neural network structure to reduce the dimensionality of the matrix, removing redundant information while retaining key state features, generating low-dimensional battery state latent variables in the latent space. These battery state latent variables capture the essential representation of the battery's internal state. A temporal convolutional network models the dynamic evolution of these latent variables, analyzes their change patterns over time, and predicts their trajectories for the next three charge-discharge cycles. The predicted latent variable trajectory is input into the decoder of the adversarial generative network. The decoder maps the latent variables back to the signal space through inverse transformation, generates a simulated battery signal, and outputs a simulated data stream containing voltage, current, and temperature parameters. The simulated data stream can be used for battery status monitoring and fault prediction.

[0023] Example 1: See Figure 2The battery signal circuit data of the BMS system includes individual cell voltage range, inter-cluster temperature gradient, DC internal resistance change rate, polarization voltage decay, and capacity decay slope. These parameters constitute the basic dataset for battery state monitoring. During operation, the energy storage battery pack collects the real-time voltage value of each individual cell through built-in voltage sensors. The battery management system periodically scans all individual cell voltages and records the maximum and minimum values. The individual cell voltage range is calculated from the difference between the maximum and minimum values, and the calculation frequency is consistent with the sampling period of the battery management system. Obtaining the inter-cluster temperature gradient requires deploying a temperature sensor array. The temperature sensors are evenly distributed on the surface and at key points inside the battery clusters. The battery management system reads the temperature sensor data and classifies it according to the physical grouping of the clusters. The inter-cluster temperature gradient is obtained by calculating the difference in average temperature between different clusters, and the time interval of the difference is set according to the thermal conductivity characteristics. The measurement of the DC internal resistance change rate depends on the voltage and current response during charging and discharging. The battery management system applies a short-time pulse current and records the voltage change waveform. The DC internal resistance value is calculated as the ratio of the voltage change to the current change. The DC internal resistance change rate is the linear regression slope of the DC internal resistance value over multiple consecutive charge-discharge cycles. Polarization voltage decay needs to be measured during the battery resting phase. The battery management system records the relaxation process of the open-circuit voltage during charge-discharge transitions. Polarization voltage decay is defined as the decay amplitude of the voltage relaxation curve at a specific time constant. The capacity decay slope is calculated by accumulating the capacity data for each charge-discharge cycle using the coulomb counting method. The capacity decay slope is the linear fit slope of the capacity sequence relative to the number of cycles.

[0024] A sliding window standardization process is applied to the voltage range of individual cells. The time span of the sliding window is set to an integer multiple of a complete charge-discharge cycle. Standardization is achieved by first subtracting the mean from the range value within the window and then dividing by the standard deviation. The coefficient of variation (COP) is calculated by dividing the standard deviation of the range values ​​within the window by the mean. The COP reflects the intensity of voltage inconsistency fluctuations; a larger COP indicates increased dispersion within the battery module. Inter-cluster temperature gradient data is input into an image generation algorithm to generate a temperature field distribution heatmap. The heatmap uses color gradients to represent high and low temperature distributions. The proportion of high-temperature regions is automatically calculated based on a set temperature threshold, representing the percentage of pixels exceeding that threshold. The proportion of high-temperature regions is directly related to battery thermal management performance; an increased proportion indicates decreased heat dissipation efficiency.

[0025] The DC internal resistance change rate and polarization voltage decay are convolved in the time domain. The time-domain convolution uses a one-dimensional convolution kernel that slides along the time axis, and the kernel weights are initialized based on the physical meaning of the parameters. The dimension of the composite aging feature vector is the same as the number of convolution kernels, and each feature dimension captures the interaction pattern between the DC internal resistance change rate and polarization voltage decay. The capacity decay slope and the area ratio of the high-temperature region are fused using a weighted formula to form a battery health status index. The coefficients of the weighted formula are determined through regression analysis of historical data, and the battery health status index values ​​are normalized to between 0 and 1. The coefficient of variation, composite aging feature vector, and battery health status index are concatenated using millisecond-level timestamps to form a dynamic parameter matrix. The row index of the dynamic parameter matrix corresponds to the time series, and the column index contains the scalar of the coefficient of variation, the multi-dimensional data of the composite aging feature vector, and the scalar of the battery health status index. The dynamic parameter matrix is ​​stored using a circular buffer structure, with the latest data overwriting the oldest data to achieve real-time updates.

[0026] The synchronization of battery signal circuit data acquisition in the battery management system is ensured by a hardware clock signal, and all sensor data are stamped with a unified timestamp. The sliding window standardization of individual cell voltage ranges uses a recursive calculation method; the mean and standard deviation are updated when new data enters the window to avoid redundant calculations of the entire window. The generation of the temperature field distribution heatmap uses a bilinear interpolation algorithm to convert discrete temperature points into a continuous image, and the calculation of the area ratio of high-temperature regions uses image segmentation technology to identify continuous high-temperature regions. The time-domain convolution of the DC internal resistance change rate and polarization voltage decay is optimized in the frequency domain, and computational complexity is reduced through Fast Fourier Transform. The dimensionality reduction of the composite aging feature vector is achieved through principal component analysis, retaining the main variance components. The weighted fusion coefficients of the battery health status indicators are dynamically adjusted according to the battery type. The concatenation operation of the dynamic parameter matrix utilizes memory-mapped files for high-speed read and write operations, and timestamp alignment uses interpolation methods to compensate for the acquisition delays of different sensors.

[0027] The size of the sliding window is adaptively adjusted based on the battery's chemical characteristics. Shorter windows are used for power batteries to capture rapid changes, while longer windows are used for energy batteries to smooth fluctuations. The color depth of the temperature field distribution heatmap is positively correlated with the temperature resolution, and the calculation of the area ratio of high-temperature regions takes into account the weight of the battery's spatial layout. The convolution kernels for the DC internal resistance change rate and polarization voltage decay are designed as trainable parameters, and their shape is optimized during model training. The normalization of the composite aging feature vector uses a minimum-maximum scaling method, and the fusion formula for battery health indicators introduces a nonlinear transformation to enhance sensitivity. Missing values ​​in the dynamic parameter matrix are handled using forward and backward padding combined with complementary sensor redundant data. Matrix compression uses a lossy compression algorithm to balance accuracy and storage efficiency.

[0028] The battery management system's battery signal circuit data acquisition link is backed up by redundant circuits, and outlier filtering is added to the calculation of individual cell voltage range. An ambient temperature compensation coefficient is introduced to calculate the inter-cluster temperature gradient, eliminating interference from external heat sources. Multi-frequency point excitation is used to improve the accuracy of DC internal resistance change rate measurement, and a double-exponential model is used to approximate the actual relaxation curve when fitting polarization voltage decay. A temperature compensation factor is introduced to calculate the capacity decay slope, correcting for the impact of temperature on capacity. Smooth transitions are used at the window boundaries of the sliding window standardization process to avoid abrupt changes, and the calculation of the coefficient of variation adds confidence intervals to assess reliability. The resolution of the temperature field distribution heatmap is matched to the number of battery clusters, and the threshold for the area ratio of high-temperature regions is set based on the thermal stability of the battery materials. Symmetrical padding is used to maintain feature length in the boundary processing of time-domain convolution, and an exponentially weighted moving average is used to update the composite aging feature vector. The fusion weights of battery health status indicators are dynamically adjusted with the number of iterations, and a precision clock protocol is used to achieve microsecond-level alignment for the timestamp synchronization of the dynamic parameter matrix.

[0029] The acquisition of individual cell voltage range uses a differential amplifier circuit to improve the common-mode rejection ratio, and the measurement of inter-cluster temperature gradient uses a platinum resistance temperature sensor to ensure linearity. The calculation of DC internal resistance change rate incorporates a temperature compensation formula, and the measurement of polarization voltage decay controls the resting time to eliminate historical effects. The calculation of capacity decay slope eliminates the influence of shallow charge-discharge cycles, and the parameters for sliding window standardization are determined through offline calibration. The generation of the temperature field distribution heatmap optimizes the color mapping function to highlight temperature difference contrasts, and the calculation of the area ratio of high-temperature regions uses morphological operations to eliminate noise interference. The initialization of the convolution kernel in the time-domain convolution uses the Xavier method, and the dimensionality reduction of the composite aging feature vector retains 90% of the energy information. The output of battery health status indicators is smoothed by Kalman filtering, and the storage format of the dynamic parameter matrix adopts columnar storage to optimize query efficiency.

[0030] The battery management system (BMS) uses the CAN bus protocol to ensure real-time transmission of battery signal circuit data. The sampling rate for individual cell voltage range is adjusted according to the SOC range. The calculation of inter-cluster temperature gradient includes an axial gradient component, and the measurement of DC internal resistance change rate is standardized within the current pulse amplitude range. Polarization voltage decay analysis distinguishes between electrochemical and concentration polarization, and a piecewise linear model is used to fit the capacity decay slope. The sliding window normalization process uses gradual weighting for window switching, and robust statistics are added to the coefficient of variation calculation to reduce the impact of outliers. The rendering of the temperature field distribution heatmap is hardware-accelerated, and a shape factor is introduced to distinguish between concentrated and dispersed heat zones in the statistical analysis of the high-temperature area ratio. The Swish function is selected as the activation function for time-domain convolution to enhance nonlinearity, and cluster analysis of composite aging feature vectors identifies typical aging patterns. The calibration of battery health status indicators uses an online update mechanism, and the encrypted storage of the dynamic parameter matrix uses the AES-256 algorithm to protect data security.

[0031] Example 2: The latent variables of battery states in the latent space are generated using an encoder structure based on a generative adversarial network. The encoder receives a multidimensional dynamic parameter matrix from the preceding steps as input data. The dimension of the multidimensional dynamic parameter matrix is ​​the time step multiplied by the number of features. The first layer of the encoder, a residual convolutional module, performs a two-dimensional convolution operation on the input matrix. The kernel size is set to 3×3, the stride is set to 1, and zero padding is used at the edges to preserve the feature map size. A shortcut connection in the residual convolutional module adds the module input and convolution output element-wise. The sum is then passed through a batch normalization layer to adjust the data distribution. The momentum parameter of the batch normalization layer is set to 0.9. Activation function selection... The nonlinear transformation is used, with the negative slope coefficient set to 0.01 to prevent neuron death. The number of filters in the five residual convolutional modules increases sequentially: 64, 128, 256, 512, and 1024, respectively. The output feature map size of each residual convolutional module is halved by a max pooling layer.

[0032] The channel attention mechanism is embedded after the convolutional layer of each residual convolutional module. The channel attention mechanism first performs global average pooling on the feature map to generate channel statistics. Global average pooling compresses all feature values ​​of each channel into a single value. Two fully connected layers form a bottleneck structure to learn the correlation between channels. The first fully connected layer reduces the number of channels to 1 / 16 of the original number, and the second fully connected layer restores the original number of channels. The activation function converts the output of the fully connected layer into channel weights between 0 and 1. These weights are then multiplied and weighted by channel-level multiplication with the original feature map, highlighting the contribution of important channels. The spatial pyramid pooling layer receives the feature map output from the last residual convolutional module. This layer uses four parallel pooling branches to process the feature map, with kernel sizes set to 1×1, 2×2, 4×4, and global pooling, respectively. Each pooling branch produces a fixed-length feature vector. Feature vectors of different scales are concatenated along the channel dimension to form a multi-scale feature representation. This multi-scale feature representation is then dimensionality-reduced to 256 dimensions using a 1×1 convolution, preserving multi-scale information while reducing computational complexity. The bidirectional long short-term memory (LSTM) network processes the temporal dimension of the multi-dimensional dynamic parameter matrix. The bidirectional LSTM network consists of two LSTM layers: a forward LSTM layer and a backward LSTM layer. The forward LSTM layer processes sequential data in chronological order, while the backward LSTM layer processes the same data in reverse chronological order. Each LSTM layer has 128 hidden units. The hidden state at the final moment of the bidirectional long short-term memory network is concatenated with the forward and backward hidden states to form a 256-dimensional temporal feature vector, which captures the long-term temporal dependencies of the dynamic parameter matrix.

[0033] Multi-scale features and temporal features are concatenated along the feature dimension, and the concatenated feature vector is input into a fully connected layer for dimensionality reduction. The fully connected layer contains three linear transformation layers with 512, 256, and 128 neurons respectively. Each linear transformation layer is followed by a batch normalization layer and... Activation function. The final output layer uses a linear activation function to generate 128-dimensional battery state latent variables. The numerical range of the battery state latent variables is determined by... The function constraint is between -1 and 1.

[0034] The encoder of the generative adversarial network employs a gradient penalty strategy during training, with the gradient penalty term added to the loss function to constrain the gradient norm of the discriminator. The encoder weights are initialized using a He normal distribution, and the bias term is initialized to zero. The optimizer is the Adam algorithm, with a learning rate of 0.0001 and a beta parameter pair of (0.5, 0.999). The training batch size is fixed at 32 samples, and the encoder parameters are smoothed using an exponential moving average after each training epoch. The generation process of the battery state latent variables requires coordinated training with the decoder; the battery state latent variables output by the encoder are interpolated with the random noise vector in the latent space. The interpolated samples are used to calculate the gradient norm of the gradient penalty term, with a target value of 1.0. The convolutional layers of the encoder use spectral normalization to constrain the weight matrix. The constant, spectral normalization, rescales the weights at each forward propagation.

[0035] The short-circuit connections of the residual convolution module use 1×1 convolutions to match the dimension. When the number of input and output channels is inconsistent, the 1×1 convolution adjusts the number of channels while maintaining the spatial size. Group normalization in the residual convolution module replaces batch normalization for processing small batches of data; group normalization divides the channels into 32 groups for normalization calculation. The spatial weight branch of the channel attention mechanism adds a max-pooling path, and the features from the max-pooling path and the average pooling path are concatenated before being input into the fully connected layer. The output feature map of the spatial pyramid pooling layer undergoes L2 normalization to prevent excessively large feature norms from affecting training stability. Layer normalization of the bidirectional long short-term memory network is integrated into the recurrent unit, calculating the mean and variance at each time step. The dropout layer of the bidirectional long short-term memory network is placed between recurrent layers, with a dropout ratio set to 0.2 to prevent overfitting. The concatenation operation of multi-scale features and temporal features uses attention-weighted fusion; the attention weights learn the importance of each feature through a shallow neural network. The parameters of the fully connected layer are initialized using an orthogonal matrix method, which maintains the stability of the gradient norm. The distribution of the battery state latent variables is monitored using kernel density estimation, which ensures that the latent variables conform to a Gaussian prior distribution.

[0036] The encoder's training data augmentation employs time-series warping, which randomly scales the multidimensional dynamic parameter matrix. The gradient penalty term is calculated using random interpolation point sampling, with these points uniformly selected along the line connecting real and generated samples. The power iteration count for spectral normalization is set to one, and singular value estimation of the weight matrix is ​​performed during each forward propagation.

[0037] The activation function of the residual convolution module introduces learnable parameterization. The slope of the negative interval can be controlled by learnable parameters. The compression ratio of the channel attention mechanism is dynamically adjusted according to the layer depth, with shallower networks using a larger compression ratio to reduce the number of parameters. The pooling method of the spatial pyramid pooling layer is changed to adaptive pooling, which automatically adjusts the pooling kernel size to adapt to the input size. The gating mechanism of the bidirectional long short-term memory network adopts a coupled input forgetting gate structure, which reduces the number of parameters and improves training efficiency. The fusion of multi-scale features and temporal features adopts a gated attention mechanism, which learns the interaction weights between features. The sparsity constraint of the fully connected layer is achieved through L1 regularization, which promotes feature selection ability. The deentanglement representation of the battery state latent variables is achieved through... Objective function optimization The objective function balances reconstruction quality and representation independence.

[0038] Adversarial training of the encoder employs Distance metric Distance provides a more stable gradient signal. Depthwise separable convolution in the residual convolution module reduces computational complexity, decomposing standard convolution into depthwise convolution and pointwise convolution. The multi-head attention mechanism of the channel attention mechanism is expanded to eight heads, processing channel information in parallel. Dilated convolution in the spatial pyramid pooling layer introduces multi-scale receptive fields, expanding the receptive field without increasing parameters. Temporal convolution in the bidirectional long short-term memory network replaces the recurrent structure, processing the entire sequence in parallel to improve computational efficiency. Coordinate attention for multi-scale and temporal features introduces positional information, decomposing channel attention into horizontal and vertical directions. The switching activation function in the fully connected layer introduces sparse activation patterns, automatically learning the on / off states of neurons. Quantization encoding of battery state latent variables is achieved through a vector quantization layer, discretizing continuous latent variables into codebook indices.

[0039] See Figure 3This paper presents the dynamic variation trend of the feature values ​​of the five residual convolutional modules in the generative adversarial network encoder with time step. Specifically, the horizontal axis represents the time step (0 to 100), and the vertical axis represents the feature value. The different colored curves correspond to residual convolutional modules 1 to 5, respectively. The feature values ​​of each module exhibit complex fluctuation characteristics in the time dimension, reflecting the differences in the feature response of different modules in the temporal dimension during the feature extraction process of the multi-dimensional dynamic parameter matrix of the energy storage battery. This feature variation trend can be used to analyze the performance of each residual convolutional module in capturing the dynamic evolution information of the battery state, providing a feature-level temporal behavior basis for subsequent temporal convolutional networks to model the dynamic evolution of battery state latent variables. It also helps to understand the role mechanism of residual convolutional modules in multi-scale, temporally dependent feature extraction, thereby optimizing the encoder structure to improve the accuracy and robustness of battery state latent variable generation.

[0040] Example 3: See Figure 4 To predict the latent variable trajectory for the next three charge-discharge cycles, a temporal convolutional network architecture is used. The temporal convolutional network receives a 128-dimensional sequence of battery state latent variables generated by the encoder as input. The time length of the battery state latent variable sequence covers historical charge-discharge cycles, and the latent variables at each time step contain compressed battery state information. The dilated causal convolutional layers of the temporal convolutional network have dilation coefficients set to powers of 2, increasing progressively in the order of 1, 2, 4, and 8. The dilated causal convolutional layers maintain the causal relationship in time sequence, and the receptive field at the output time covers multiple historical time steps. The number of filters in the dilated causal convolutional layers is set to 256, and the kernel size is fixed at 3. After each dilated causal convolutional layer, layer normalization and summation are applied. Activation function. Layer normalization calculates the mean and variance of all features for a single sample. The activation function sets negative values ​​to zero to maintain feature sparsity. A gated recurrent unit (ROU) network is superimposed on the output of the dilated causal convolutional layer. The RNU network includes update and reset gate structures. The update gate controls the degree of retention of historical state information, while the reset gate determines the proportion of past states forgotten. The hidden state dimension of the RNU network is set to 512. The state transition equation of the RNU network describes the evolution of the hidden variables in the following form: in: This indicates the hidden state at the current moment. This indicates the hidden state at the previous moment. This indicates updating the output value of the gate. Indicates candidate activation values. Update gate. From the current input and the previous hidden state pass The function calculates the candidate activation values. use Function combined with reset gate Generation. This differential equation describes the continuity of the hidden state's evolution over time, with the left-hand side using state units and the right-hand side using consistent state units. A multi-head self-attention mechanism computes the latent variable weight distribution for future time steps, linearly mapping the input sequence to a query matrix, key matrix, and value matrix. The number of attention heads is set to 8, each with a dimension of 64. Scaled dot-product attention computes the similarity score between the query and the key. The similarity score is obtained through... The function is normalized to attention weights, and the attention weights are weighted and summed with the value matrix to obtain the output of each head. The multi-head outputs are concatenated along the feature dimension and then fused through a linear transformation. Position encoding is added to the input sequence to preserve temporal information.

[0041] Monte Carlo sampling generates multiple latent variable evolution paths from the predicted distribution, using a Gaussian distribution assumption to generate random perturbations. The sampling count is set to 100 times, with each sample generating a complete trajectory for the next three cycles. The trajectory probability density is calculated using kernel density estimation. The path with the highest probability density is selected as the final predicted trajectory, and the temporal resolution of the predicted trajectory is consistent with the input sequence. The temporal convolutional network is trained using historical charge / discharge data, with the training objective being to minimize the mean squared error between the predicted trajectory and the true latent variables. Optimizer selection... The algorithm employs a cosine annealing strategy, gradually decreasing the learning rate from 0.001. Gradient clipping limits the gradient norm to no more than 1.0 to prevent training instability. A missing layer is applied between convolutional layers, with a missing ratio set to 0.1 to reduce the risk of overfitting.

[0042] The dilated convolutional operation of the dilated causal convolutional layer skips intermediate time points, expanding the receptive field without increasing the number of parameters. The receptive field size is adjusted by the dilation coefficient and network depth, with the final layer's receptive field covering the entire historical sequence. Causal convolution ensures that the prediction at each time point relies only on historical information, without leaking future data. The parameters of the convolutional layer are initialized using a He normal distribution, with the bias term initialized to zero. The update and reset gates of the gated recurrent unit network are implemented using fully connected layers, with the weight matrix of the fully connected layer being the square of the input dimension plus the hidden dimension. The calculation of candidate activation values ​​introduces a reset gate to control the participation of historical states; when the reset gate is close to zero, past information is completely ignored. The hidden state of the gated recurrent unit network is initialized using a zero vector, and the recurrent connections update parameters through backpropagation over time. The positional encoding of the multi-head self-attention mechanism uses sine and cosine functions, generating a unique positional identifier for each time step. Relative positional encoding considers the distance relationship between time steps, enhancing the model's ability to capture temporal relationships. The visualization of attention weights shows the key time points that the model focuses on, corresponding to important moments in battery state transitions.

[0043] The random number generation for Monte Carlo sampling employs the Mason swirl algorithm, with a fixed random seed ensuring reproducible experimental results. Smoothing of the sampling path uses a moving average filter, with the filter window size matched to the charge / discharge cycle. Probability density calculation uses a Gaussian kernel function, and bandwidth selection is automatically determined using the Scott rule. Uncertainty quantification of the predicted trajectory is achieved through variance calculation of the sampling path, with a variance map displaying the change in prediction confidence over time. High-variance regions correspond to state transition points, while low-variance regions represent stable operating phases. Uncertainty information is used for risk assessment; a manual check mechanism is triggered when variance exceeds a threshold. The inference process of the temporal convolutional network is optimized for parallel computation, processing the entire sequence at once to improve efficiency. Convolution operations are converted to matrix multiplication using the im2col algorithm, and the BLAS library is used to accelerate computation. Memory layout optimization reduces cache misses, improving real-time inference performance. The loop computation of the gated recurrent unit network is unfolded with a fixed step size, and loop optimization techniques reduce function call overhead. Hidden state caching avoids redundant computation, and pre-allocated memory space reduces dynamic allocation costs. The key value cache of the multi-head self-attention mechanism stores intermediate results, accelerating the autoregressive generation process. Monte Carlo sampling uses vectorized operations for path generation, completing multiple samples in the same batch. Probability density calculations use logarithmic space to avoid numerical underflow, and addition of log probabilities replaces multiplication. Path selection is based on log probability comparisons, resulting in better numerical stability.

[0044] Post-processing of the predicted trajectory includes outlier filtering and smoothing. Outlier detection uses the Isolation Forest algorithm. Smoothing employs the Savitzky-Gore filter to remove noise while preserving the trajectory shape characteristics. The rate of change of the trajectory derivative is calculated, with peak values ​​corresponding to abrupt state changes. Monitoring metrics for the temporal convolutional network include receptive field size and gradient norm. The receptive field size ensures coverage of the complete cyclic pattern. The gradient norm monitors training stability; a sudden increase indicates a risk of gradient explosion. The activation value distribution histogram displays changes in feature distribution; normalization parameters are adjusted when the distribution shifts.

[0045] Saturation state monitoring of gates in the gated circulation unit is achieved through Output value statistics show that values ​​close to 0 or 1 indicate saturation. Saturation leads to gradient vanishing, requiring adjustments to initialization or activation functions. Hidden state norm growth monitors gradient explosion, while norm pruning controls the numerical range. The attention graph visualization of the multi-head self-attention mechanism displays time-dependent patterns: diagonal patterns represent local dependencies, and dispersed patterns represent global dependencies. Attention head specialization analysis shows that each head focuses on different time scales, some on short-term fluctuations, others on long-term trends. Monte Carlo sampling convergence is monitored through path variance; stable variance indicates sufficient sampling. Path diversity metrics ensure sufficient exploration and avoid getting trapped in local optima. Sampling efficiency monitors the acceptance rate; if the acceptance rate is too low, the proposal distribution is adjusted. Predicted trajectories are evaluated using multiple metrics, including mean squared error, mean absolute error, and dynamic time warping distance. Dynamic time warping distance handles time scaling variations and is more suitable for evaluating temporal shape similarity. Predicted interval coverage measures the degree of uncertainty calibration; ideally, the value should be close to the confidence level.

[0046] The multi-step prediction of the temporal convolutional network employs a recursive strategy, with each prediction serving as input for the next. Accumulated error from recursive prediction is mitigated through planned sampling, which combines teacher-mandated and free-running modes. The planned sampling ratio gradually decreases during training, eventually relying entirely on the model's own predictions. Gradient flow in the gated recurrent unit network is prevented from exponential gradient growth by temporally pruning the gradient norm. Gradient pruning preserves direction while adjusting size to maintain training stability. Hidden state normalization controls the numerical range, preventing activation function saturation. The multi-head self-attention mechanism uses linear attention to approximate long sequences, reducing complexity from quadratic to linear. A soft maximum kernel is chosen to maintain attention properties, and random feature mapping accelerates computation. Long sequence processing capabilities are extended to the entire battery lifespan. Monte Carlo sampling reduces variance through importance sampling, and a proposal distribution focuses on high-probability regions. Importance weights correct bias, and unbiased estimation improves sampling efficiency. The adaptive proposal distribution adjusts based on historical samples, gradually approximating the target distribution. Applications of the predicted trajectory include anomaly detection and lifetime prediction; anomaly detection is achieved by comparing predicted and actual values. Residual sequence analysis identifies anomalous patterns, and mutation point detection pinpoints anomalies. Lifespan prediction is based on trajectory trend extrapolation, and remaining service life estimation supports maintenance decisions.

[0047] Example 4: The process of generating simulated battery signals by reverse decoding based on predicted latent variable trajectories is completed by the decoder component of the generative adversarial network. The decoder receives the predicted latent variable trajectories for the next three charge-discharge cycles from the temporal convolutional network. The latent variable trajectories contain a 128-dimensional vector sequence, with the sequence length corresponding to the target simulation time period. The first transposed convolutional layer of the decoder upsamples the latent variable sequence to a 256-dimensional feature map. The transposed convolutional kernel size is set to 4, the stride is set to 2, and the same padding is used to maintain size alignment. The output of the transposed convolutional layer adjusts the data distribution through a batch normalization layer. The momentum parameter of the batch normalization layer is set to 0.9, and the epsilon parameter is set to 1e-5 to prevent division by zero errors. The activation function is applied to the normalized features, with the negative slope coefficient set to 0.2 to maintain gradient flow.

[0048] The decoding process incorporates the rated parameter constraints of the energy storage high-voltage box, including the maximum allowable voltage, continuous operating current range, and temperature limit. These rated parameter constraints are implemented using a differentiable function, mapping the decoder output to the physically feasible region. Voltage constraints use a hyperbolic tangent function to limit the output to between 0 and the maximum voltage, while current constraints employ a scaled... The function ensures correct sign. Temperature constraints map the network output to the actual temperature range through a linear transformation, and the gradient of the constraint function is backpropagated using the chain rule. Spectral normalization is applied to all weight matrices of the decoder, calculating the maximum singular value of the weight matrix during each forward propagation. A power-law iteration algorithm is used for singular value estimation, with a fixed number of iterations (1) to balance computational accuracy and efficiency. Spectral normalization divides the weight matrix by the estimated singular value, constraining... The training process is stable under constant conditions. The stability of gradient propagation is further enhanced by a gradient penalty term, which calculates the gradient norm of the discriminator with respect to randomly interpolated samples.

[0049] Output connection The activation function generates a standardized simulation signal. The function compresses the output to the range of 0 to 1. The standardized simulation signal is converted into actual physical quantities through an inverse normalization transformation; the inverse normalization parameters are derived from statistics in the training dataset. The simulation data stream contains time-series data of three parameters: voltage, current, and temperature, with the data sampling rate consistent with the original acquisition system. The dynamic time warping distance between the simulated signal and the measured signal is monitored in real time, and this distance is used to calculate the minimum alignment cost between the two time series. The tolerance threshold for the dynamic time warping distance is set based on the historical error distribution, and the threshold is updated using a sliding window statistical method. When the dynamic time warping distance exceeds the tolerance threshold, the retraining mechanism of the Generative Adversarial Network (GAN) is triggered, using recently acquired real data. During retraining, the parameters of the temporal convolutional network are frozen, and only the weights of the GAN are updated. Parameter freezing is achieved by setting the `requires_grad` flag; the optimizer only calculates gradients for parameters where `requires_grad` is True. The retraining of the GAN uses a small learning rate, set to one-tenth of the initial training rate, to avoid destroying the learned feature representations (see Table 1).

[0050] Table 1 shows the detailed parameter configurations for each layer of the decoder.

[0051] The decoder's transposed convolutional layers progressively increase the feature map size while decreasing the number of channels, ultimately producing a 3-channel simulated signal at the output layer. After each transposed convolutional layer, spectral normalization is applied to constrain the weight norm. The power iteration of spectral normalization uses randomly initialized left and right vectors. The power iteration process includes matrix multiplication and vector normalization steps, converging to obtain an estimate of the maximum singular value. The nominal parameter constraints are implemented using a differentiable projection layer, which maps the network output to the feasible region while maintaining gradient computability. Voltage constraints use a scaled hyperbolic tangent function, with a scaling factor equal to half the maximum voltage value. Current constraints handle the sign issue using a piecewise approach. Function combination, positive current use Functions, negative current uses negative The function uses a linear mapping to map the [0,1] interval to the actual temperature range, with mapping parameters determined based on the sensor's range. Dynamic time-warped distance calculation employs a dynamic programming algorithm, creating a cost matrix to store the cumulative distance. The search for curved paths uses monotonicity and continuity constraints; after finding the minimum cost path, the average distance is calculated. Adaptive adjustment of the tolerance threshold is based on the mean and standard deviation of recent errors, with the threshold set to the mean plus three times the standard deviation.

[0052] The retraining mechanism uses a first-in, first-out (FIFO) queue for data management, with a fixed queue size of the most recent 1000 samples. Retraining batches are randomly sampled from the queue, with sampling weights positively correlated with the error magnitude. Retraining of the adversarial generative network uses a gradient-penalized approach. The distance and gradient penalty coefficient were set to 10. The discriminator's training iterations were set to five times that of the generator to balance their training progress. Gradient flow analysis of the decoder showed that spectral normalization effectively controlled the gradient norm, which remained stable between 0.8 and 1.2. Activation value distribution was monitored using an exponential moving average, with normalization parameters dynamically adjusted as the distribution shifted. Output signal quality was assessed using a multi-scale structural similarity metric, considering the signal's structural information rather than simple pixel differences.

[0053] The simulation data stream time alignment uses interpolation to compensate for processing delays, and the interpolation algorithm preserves the signal shape characteristics. The computational complexity of the real-time monitoring system is optimized using a sliding window approximation, with the window size synchronized to the charge / discharge cycle. The anomaly detection mechanism calculates signal anomaly scores based on Mahalanobis distance; detailed analysis is triggered when the score exceeds a threshold. Convergence judgment during retraining is based on changes in the validation set loss; premature termination occurs if there is no improvement for several consecutive cycles. Model version management records performance metrics for each retraining iteration; performance degradation leads to a rollback to a previous version. Quality filtering of the retraining data uses the Isolation Forest algorithm to remove outlier samples and ensure training data quality.

[0054] Post-processing of the simulated signal includes smoothing filtering and outlier correction. The smoothing filter employs a moving average algorithm with a window width matched to the signal characteristics. Outlier detection is based on statistical outlier detection, using absolute median difference to estimate data distribution. The correction algorithm uses nearest-neighbor interpolation to replace outliers, maintaining signal continuity. Decoder computation graph optimization uses operation fusion technology to merge continuous linear operations into a single operation. Memory allocation optimization reuses intermediate result buffers, reducing dynamic memory allocation overhead. Parallel computing utilizes a multi-core CPU architecture, distributing batch samples across different cores for processing. The real-time monitoring interface visually compares the simulated and measured signals, highlighting differences in prominent colors. Historical error trend graphs show the evolution of system performance, aiding in the assessment of model degradation. The parameter adjustment interface allows engineers to fine-tune constraints to adapt to different operating conditions. Alarm logs record all triggered events, supporting post-event analysis and fault diagnosis.

[0055] Example 5: The specific implementation process of the retraining mechanism is based on a sudden voltage fluctuation event in the battery modules of an energy storage power station. During the cold wave in the winter of 2023, the power station monitoring system detected an abnormal deviation in the simulation signal of battery container No. 3. The dynamic time warping distance monitoring module recorded a deviation value exceeding the tolerance threshold within three consecutive charge-discharge cycles. The deviation value increased from an initial 0.12 to 0.45, triggering the retraining mechanism of the generative adversarial network. The system automatically selected charge-discharge cycle segments that matched the current operating conditions from the historical database. The selection criteria included an ambient temperature range of -10℃ to 0℃, a SOC fluctuation amplitude greater than 30%, and a charge-discharge rate between 0.5C and 1C. The data selection algorithm was based on distance calculation in the multi-dimensional feature space, selecting the 500 historical cycle segments with the smallest Euclidean distance to the current operating state as the retraining dataset.

[0056] When constructing the anomalous sample set containing voltage surge events and temperature surge events, the system scanned all recorded anomalous events from the past two years. Voltage surge events were defined as operating conditions with a voltage change rate exceeding 5V / s, and temperature surge events were defined as operating records with a temperature change rate greater than 3℃ / min. The anomalous sample set contained 127 special event segments, each containing a complete data sequence of 10 minutes before and after the event. These anomalous samples were mixed with normal samples at a 1:3 ratio to form a balanced training dataset, preventing the model from overemphasizing rare events. The focus loss function adjusted the sample weights during retraining, with the adjustment factor γ set to 2.0, automatically increasing the loss weight of difficult samples. The mathematical expression of the focus loss function reduced the loss contribution of correctly classified samples while maintaining a larger weight for misclassified samples. During loss calculation, the model's attention to voltage surge events increased by 3.2 times, and its sensitivity to temperature surge events increased by 2.7 times, effectively strengthening the training influence of anomalous samples.

[0057] Transfer learning techniques transfer the low-level feature extraction capabilities of the pre-trained model to the new model, employing a full weight copying strategy for the feature extraction layer during parameter transfer. The first four convolutional weights of the encoder in the Generative Adversarial Network (GAN) are directly loaded from the pre-trained model, while the final fully connected layer undergoes random initialization. The parameters of the temporal convolutional network remain frozen throughout the retraining process; gradient calculation skips these layers, updating only the trainable parameters of the GAN. Retraining is performed on edge computing nodes, with a training epoch of 100 and a batch size of 16. A step-down learning rate strategy is used, initially set to 0.0001, decreasing to 0.5 times the original value every 20 epochs. A gradient clipping threshold of 1.0 is set to prevent gradient explosion. Monitoring during training shows that the dynamic temporal warping distance of the retrained model on the validation set decreased from 0.45 to 0.08, falling below the tolerance threshold of 0.15.

[0058] When establishing the mapping relationship library between simulated signals and the actual operating status of the energy storage battery pack, the system integrated three years of maintenance records and fault reports. The mapping relationship library adopts a graph database storage structure, where nodes represent fault modes and edges represent the correlation strength between signal features and faults. The fault mode coding refers to the definition of the international standard IEC62619 and includes 12 typical faults such as internal short circuit, lithium plating, and diaphragm aging. Each fault mode is associated with 10 to 20 feature signal modes, forming a complete knowledge graph. When abnormal fluctuations in polarization voltage are detected in the simulated signal, the pattern matching algorithm calculates the similarity between the current signal and the fault mode in the mapping relationship library. The similarity calculation adopts a hybrid algorithm combining dynamic time warping distance and cosine similarity, with weights of 0.7 and 0.3. The detection threshold for abnormal fluctuations in polarization voltage is set to ±15% of the baseline value, and the diagnostic process is triggered if the duration exceeds 3 sampling periods.

[0059] The automatic association mapping of fault modes in the database employs a rule-based inference engine. This engine first eliminates the influence of ambient temperature, then checks voltage-current coordination, and finally matches historical fault modes. During the association process, the system prioritizes matching fault types that occurred within the last three months, while also considering the battery's health status index. The system determines successful fault mode identification when the fault matching confidence level reaches 0.75 or higher. When generating a diagnostic report containing fault location codes and severity levels, the system uses a standardized template. The fault location code includes three levels of information: battery container number, module location, and individual cell serial number. Severity levels are categorized into observation, warning, and alarm. The diagnostic report automatically includes a timestamp, environmental parameters, and detailed data for the preceding and following hours for maintenance personnel to analyze and make decisions. After generation, the report is sent to the monitoring center via an encrypted channel, simultaneously triggering a local audible and visual alarm. Retraining performance evaluation shows that the model accurately predicted two voltage fluctuation events during the subsequent 30 days of operation, with a warning time lead of 15 minutes. The fault diagnosis system successfully identified an early lithium plating fault, classifying it as a warning level, allowing maintenance personnel to adjust charging strategies in a timely manner and prevent the incident from escalating. The mapping relationship database is continuously updated during operation. Newly identified fault modes and their characteristic signals are added to the database, enabling it to achieve self-improvement.

[0060] The entire retraining mechanism employs an automated pipeline design, requiring no manual intervention from bias detection to model updates. The monitoring interface displays retraining progress and performance metrics in real time, and operations personnel can view detailed logs at any time. The system retains a copy of the model before each retraining session, allowing for rapid rollback to a stable version in case of performance degradation. Computational resource allocation during retraining utilizes a dynamic adjustment strategy, with training tasks automatically scheduled during system idle times, without affecting real-time monitoring. The abnormal sample set is updated monthly, with the system automatically collecting new abnormal events and retraining. The parameters of the focus loss function are dynamically adjusted based on the sample distribution to maintain focus on difficult samples. The transfer learning strategy periodically evaluates the applicability of the base model, initiating a global retraining when the base model becomes outdated. Cross-validation is used to validate the mapping relation library, ensuring the reliability of newly added failure modes.

[0061] The diagnostic reports conform to industry standards, including necessary technical parameters and maintenance recommendations. Upon report generation, they are automatically linked to the work order system to create corresponding maintenance work orders. Faults with an alarm severity level directly notify the on-duty engineer, initiating emergency response procedures. During the entire system operation, a total of 47 diagnostic reports were generated, accurately identifying 39 potential faults, with a false alarm rate controlled below 5%. The reliability of the retraining mechanism was verified through continuous operation testing; the system triggered retraining 12 times within six months. The average training time was 45 minutes, with a minimum of 20 minutes and a maximum of 2 hours. The retrained models consistently achieved the expected performance on the validation set, with no instances of performance degradation. Resource monitoring showed that the computational resources consumed during the retraining process were within the system design limits and did not affect the normal operation of other functions.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based simulation method for energy storage battery modules, characterized in that, include: Collect battery signal circuit data of BMS system during charge and discharge cycles of energy storage battery pack, and extract voltage fluctuation characteristics and current ripple characteristics of energy storage high voltage box; By aligning voltage fluctuation characteristics and current ripple characteristics according to time series, a multi-dimensional dynamic parameter matrix reflecting the internal state of the battery is constructed. An encoder using an adversarial generative network is used to reduce the dimensionality of the multidimensional dynamic parameter matrix and generate latent battery state variables in the latent space. The battery state latent variables are dynamically evolved and modeled using a temporal convolutional network to predict the trajectory of latent variables for the next three charge-discharge cycles. The simulated battery signal is generated by reverse decoding based on the predicted latent variable trajectory, and the output is a simulated data stream containing voltage, current and temperature parameters.

2. The machine learning-based simulation method for energy storage battery modules according to claim 1, characterized in that, The battery signal circuit data of the BMS system includes the individual cell voltage range, inter-cluster temperature gradient, DC internal resistance change rate, polarization voltage decay, and capacity decay slope.

3. The machine learning-based simulation method for energy storage battery modules according to claim 2, characterized in that, The specific steps for constructing the multidimensional dynamic parameter matrix are as follows: The range of individual voltages is standardized using a sliding window, and the coefficient of variation of the range values ​​within the window is calculated. A temperature field distribution heatmap is generated based on the inter-cluster temperature difference gradient, and the area ratio of high-temperature regions in the heatmap is extracted. The DC internal resistance change rate and the polarization voltage decay are convolved in the time domain to generate a composite aging feature vector. By combining the capacity decay slope with the area ratio of the high-temperature region, a battery health status indicator is formed. The coefficient of variation, composite aging feature vector, and battery health status indicators are concatenated into a dynamic parameter matrix by timestamp.

4. The machine learning-based simulation method for energy storage battery modules according to claim 3, characterized in that, The latent variables of the battery state in the generated potential space include: Five residual convolutional modules are set in the encoder of the generative adversarial network, and a channel attention mechanism is connected after each convolutional layer. Multi-scale feature extraction of multi-dimensional dynamic parameter matrices is performed using spatial pyramid pooling layers; Bidirectional long short-term memory networks are used to capture the temporal dependencies of dynamic parameter matrices; Multi-scale features and temporal features are concatenated and then input into a fully connected layer to output 128-dimensional battery state latent variables.

5. The machine learning-based simulation method for energy storage battery modules according to claim 4, characterized in that, The steps for predicting the latent variable trajectory for the next three charge-discharge cycles include: Use dilated causal convolutional layers of temporal convolutional networks to capture long-term dependencies of latent variables; A gated recurrent unit network is superimposed at the output of the convolutional layer to establish a hidden variable state transition equation; The distribution of latent variable weights for future time steps is calculated using a multi-head self-attention mechanism; Monte Carlo sampling is used to generate multiple latent variable evolution paths, and the path with the highest probability density is selected as the predicted trajectory.

6. The machine learning-based simulation method for energy storage battery modules according to claim 5, characterized in that, The process of generating simulated battery signals by reverse decoding based on predicted latent variable trajectories includes: The predicted trajectory is input into the decoder of the adversarial generative network, and the signal dimension is gradually recovered through transposed convolutional layers; The rated parameter constraints of the energy storage high-voltage box are integrated during the decoding process; The gradient propagation of the decoder is stabilized using spectral normalization techniques; The output is connected to a sigmoid activation function to generate a standardized simulation signal.

7. The machine learning-based simulation method for energy storage battery modules according to claim 6, characterized in that, Also includes: Real-time monitoring of the dynamic time-warped distance between simulated and measured signals; When the dynamic time warp distance exceeds the tolerance threshold, the retraining mechanism of the adversarial generative network is triggered. During retraining, the parameters of the temporal convolutional network are frozen, and only the weights of the adversarial generative network are updated.

8. The machine learning-based simulation method for energy storage battery modules according to claim 7, characterized in that, The specific steps of the retraining mechanism are as follows: Select charge / discharge cycle segments from historical data that match the current operating conditions; Construct an anomalous sample set containing voltage surge events and temperature surge events; The focus loss function is used to strengthen the training weights of outlier samples; Transfer learning transfers the low-level feature extraction capabilities of a pre-trained model to a new model.

9. The machine learning-based simulation method for energy storage battery modules according to claim 8, characterized in that, Also includes: Establish a mapping library between simulation signals and the actual operating state of energy storage battery packs; When abnormal fluctuations in polarization voltage are detected in the simulation signal, the fault modes in the mapping relationship library are automatically associated. Generate a diagnostic report that includes fault location codes and severity levels.

10. A machine learning-based energy storage battery module simulation system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the machine learning-based energy storage battery module simulation method according to any one of claims 1 to 9.

Images