Battery capacity prediction method based on multi-source asynchronous perception and time hybrid modeling

Abstract

The application discloses a battery capacity prediction method based on multi-source asynchronous sensing and time hybrid modeling, which comprises the following steps: collecting electrochemical parameters of a battery pack, and performing pretreatment to obtain time series samples and extract a capacity time series; constructing a multi-channel convolutional encoder family module, taking the time series samples as input, independently extracting features of the electrochemical parameters, and obtaining a spatial feature mapping; constructing a time hybrid modeling module, taking the capacity time series as input, and obtaining a capacity time sequence feature mapping; constructing a cross-parameter feature fusion layer, inputting the spatial feature mapping and the capacity time sequence feature mapping, and obtaining a health state feature; inputting the health state feature into a regression prediction head to output a predicted value of the battery pack; and constructing a main loss function to train the model. The application realizes an integrated prediction mechanism for the degradation process of the battery capacity of a substation, can effectively adapt to the complex operation environment of the substation, and has high engineering feasibility and application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power systems and condition assessment of energy storage devices, specifically to a method for predicting battery capacity based on multi-source asynchronous sensing and time-hybrid modeling. Background Technology

[0002] In recent years, with the rapid development of smart grids and energy storage systems, substation battery banks, as critical backup power equipment, directly affect the safe and reliable operation of relay protection and automation devices. The degradation process of battery capacity typically exhibits strong nonlinearity, time-varying characteristics, and multi-factor coupling, influenced by long-term factors such as ambient temperature, current surges, depth of charge / discharge, and internal resistance changes. Therefore, achieving high-precision modeling and prediction of battery capacity degradation trends has become a crucial technical issue for ensuring power system stability and operational economics.

[0003] Existing research largely relies on single time series modeling methods or shallow feature extraction networks, making it difficult to simultaneously address the asynchronicity and temporal dependence of multi-source signals. Typical methods include:

[0004] (1) Capacity decay fitting based on statistical models, such as exponential decay models, regression curves or equivalent circuit parameter estimation, are simple to calculate and highly interpretable, but have poor adaptability to complex operating conditions and dynamic disturbances.

[0005] (2) Prediction models based on recurrent neural networks (RNN) and long short-term memory networks (LSTM) can capture the long-term dependence of capacity changes over time through gating mechanisms, but they are easily affected by gradient accumulation error and information competition under multi-parameter asynchronous input conditions.

[0006] (3) The local time-series modeling method based on convolutional neural networks (CNN) can effectively extract short-term fluctuations and local change patterns of signals such as voltage, current, and temperature. However, its fixed receptive field structure is difficult to cover long-term degradation trends, and the information interaction capability between different channels is limited.

[0007] In recent years, some studies have attempted to introduce improved convolutional network structures (such as ConvNeXt and ResNet variants) and temporal hybrid modeling modules (such as TimeMixer) to balance the modeling capabilities of local and global features. However, conventional convolutional prediction models typically employ fixed convolutional kernels and static feature extraction methods, making it difficult to fully capture the nonlinear and dynamically changing structural features in multi-source monitoring signals, resulting in relatively limited deep representation capabilities and generalization performance. In contrast, ConvNeXt-Tiny introduces multi-scale deep convolutions, pointwise convolutions, layer normalization, and deeper residual channel designs at the structural level, giving it greater stability and adaptability when representing complex, noisy, and dynamically perturbed features. In particular, its modern convolutional structure, which is closer to the characteristics of the Transformer (including a larger receptive field and more flexible deep convolutional paths), demonstrates advantages that traditional convolutional models lack in capturing local pattern changes and long-term trend features in battery degradation.

[0008] On the other hand, traditional time series models such as LSTM mainly rely on gating structures to model time dependencies. Although they have a certain degree of sequence memory capability, their effective memory length is limited, they are prone to gradient decay, and they are difficult to simultaneously characterize long-term trend information and short-term local changes. The TimeMixer structure, through multi-scale time window mixing, cross-window dynamic residual paths, and implicit frequency domain modeling mechanisms, can more fully explore the correlation characteristics between long-range dependencies and complex perturbations. It is particularly suitable for processing degraded time series data with trend, non-stationarity, and multiple perturbation superposition characteristics. Summary of the Invention

[0009] The purpose of this invention is to provide a battery capacity prediction method based on multi-source asynchronous sensing and time-mixed modeling (Multi-Source Asynchronous Time-Mixing Network, or MSA-TMNet for short), in order to solve the following key problems existing in current models in power system scenarios:

[0010] The inconsistent sampling frequencies of multi-source asynchronous monitoring signals lead to time alignment errors and mismatches in degradation features; the differences in the roles of various electrochemical parameters (voltage, current, temperature, resistance, etc.) in capacity decay are not fully modeled; traditional recurrent networks suffer from memory decay and training instability in modeling long-term degradation trends; and the neglect of nonlinear coupling and dynamic importance between signals during the multi-parameter fusion stage results in decreased prediction accuracy.

[0011] To achieve the above functions, this invention designs a battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling, executing the following steps S1-S6 to complete the battery pack capacity prediction of the target substation:

[0012] Step S1: Collect the electrochemical parameters of the battery pack during the operation of the target substation, including voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio and state entropy, and preprocess the electrochemical parameters to obtain a time series sample composed of the preprocessed electrochemical parameters, and extract the capacity time series.

[0013] Step S2: Based on the asynchronous sampling characteristics of each electrochemical parameter, a multi-channel convolutional encoder family module is constructed. Taking time series samples as input, each electrochemical parameter is extracted independently to obtain the spatial feature mapping of each electrochemical parameter.

[0014] Step S3: Based on the long-term degradation trend of battery pack capacity over time, a time-hybrid modeling module is constructed to obtain the capacity time-series feature mapping by taking the capacity time series as input.

[0015] Step S4: Construct a cross-parameter feature fusion layer, which concatenates the spatial feature map output by the multi-channel convolutional encoder family module with the capacity temporal feature map output by the temporal hybrid modeling module. Then, a feature channel attention mechanism is introduced to calculate the attention weights of different electrochemical parameter features. Finally, weighted fusion is performed to obtain the fused health status features.

[0016] Step S5: Input the fused health status features into the regression prediction head and output the predicted value of the battery pack, including the capacity at the next moment, the future multi-step capacity degradation sequence, and the remaining service life estimate.

[0017] Step S6: Based on the multi-channel convolutional encoder family module, temporal hybrid modeling module, cross-parameter feature fusion layer, and regression prediction head, construct a battery pack capacity prediction model, construct the main loss function, supervise the training of the battery pack capacity prediction model, obtain the trained battery pack capacity prediction model, and apply the trained battery pack capacity prediction model to complete the battery pack capacity prediction of the target substation.

[0018] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0019] First, unlike the single LSTM, CNN, or simple cascaded structures commonly used in existing research, traditional methods often directly stack multi-source sensor signals (such as voltage, current, and temperature) at the input for modeling. This ignores the differences in temporal resolution, dynamic changes, and degradation sensitivity of different physicochemical parameters, leading to feature aliasing, information interference, and insufficient model generalization. This invention designs a multi-channel convolutional encoder family of modules (MB-CNTE) to structurally achieve independent convolutional branch modeling of signals such as voltage, current, resistance, temperature, temperature-internal resistance coupling ratio (T / R), and state entropy, fully exploiting the unique local dynamics and non-stationary features of each electrochemical parameter. This design effectively avoids feature contamination caused by the fusion of heterogeneous signals at the input, significantly improving the discriminative power and stability of the features.

[0020] Secondly, this invention introduces a Temporal Hybrid Modeling (TTMB) module for capacity sequence modeling. By modeling with independent input capacity data, it preserves the physical continuity of capacity changes and degradation trend characteristics. The TTMB module proposes a hybrid convolution-attention joint modeling structure based on temporal pattern decoupling. This structure redesigns the decomposition and reconstruction of temporal features based on the TimeMixer framework: firstly, it uses multi-scale hybrid convolution to capture short-term mutations and local dynamics; then, it introduces a directional temporal attention operator to weight the contribution of degradation trends in different time periods, achieving explicit modeling of long-term dependent features of capacity degradation. This enables the collaborative capture of short-term fluctuations and long-term trends in capacity degradation, resulting in stronger stability and trend preservation capabilities in long-term prediction scenarios.

[0021] Furthermore, this invention proposes a cross-parameter feature fusion layer (CP-Fusion) in the multi-feature fusion stage. This layer concatenates and fuses multi-source physical features from MB-CNTE with the capacity temporal features of TTMB along the channel dimension, and adaptively allocates the weights of each parameter through a feature channel attention mechanism. This mechanism can dynamically highlight the contribution of key electrochemical parameters (such as temperature or internal resistance) in the degradation stage, suppress redundant parameter interference, and form a semantically consistent global health representation vector.

[0022] Furthermore, by introducing a trend-constrained loss function, this invention simultaneously constrains the changing trends of the predicted curve and the actual capacity degradation trajectory at the loss design level, so that the prediction results are consistent in both numerical error and degradation mode, thereby achieving a balance between the smoothness and physical rationality of capacity prediction.

[0023] In summary, the MB-CNTE–TTMB parallel structure and CP-Fusion fusion framework proposed in this invention achieve innovative breakthroughs in feature representation, temporal modeling, and semantic fusion. Compared with existing models, this method shows significant improvements in capacity prediction accuracy, trend preservation, early degradation identification capability, and model interpretability.

[0024] Experimental results on publicly available battery datasets (such as NASA and CALCE) show that the present invention outperforms traditional CNN-LSTM, ConvMixer, and Transformer series models in terms of RMSE, MAE, and trend consistency metrics for long-term capacity prediction, verifying the superiority of the present invention in multi-source asynchronous signal modeling and capacity degradation prediction tasks.

[0025] This invention not only provides a high-precision prediction tool for the health status assessment of individual batteries, but can also be extended to the aggregation of degradation risks and life management of battery groups, providing reliable decision support for the intelligent operation and maintenance of substations and energy storage systems. Attached Figure Description

[0026] Figure 1 This is a flowchart of a battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to an embodiment of the present invention.

[0027] Figure 2 This is a network structure diagram of a battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling provided in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of the structure of a family of multi-channel convolutional encoder modules provided in an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of the structure of the time-mixing modeling module provided in an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of the cross-parameter feature fusion layer provided according to an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0032] The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling provided in this embodiment of the invention is illustrated in the flowchart below. Figure 1 Network structure diagram (refer to) Figure 2The model established by this method consists of three core components: a multi-channel convolutional encoder family module (MB-CNTE), a temporal hybrid modeling module (TTMB), and a cross-parameter feature fusion layer (CP-Fusion). Finally, a regression predictive decoder is used to jointly predict capacity and lifetime. Steps S1-S6 are executed to complete the battery pack capacity prediction for the target substation:

[0033] Step S1: Collect the electrochemical parameters of the battery pack during the operation of the target substation, including voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio and state entropy, and preprocess the electrochemical parameters to obtain a time series sample composed of the preprocessed electrochemical parameters, and extract the capacity time series.

[0034] The specific steps of step S1 are as follows:

[0035] Step S1.1: Collect electrochemical parameters of the battery bank during the operation of the target substation, including voltage. ), current (Current, ), internal resistance ), temperature Temperature-internal resistance coupling ratio T / R, and state entropy. Asynchronous time-series variables, such as...

[0036] Step S1.2: Preprocess each electrochemical parameter. Addressing issues such as inconsistent sampling frequencies, noise interference, and missing measurements in the original data, perform missing value interpolation, outlier removal, time alignment, and standardization to ensure alignment of electrochemical parameters on the time axis and consistency in the numerical domain, providing a reliable input foundation for subsequent deep modeling. This is done according to a fixed time window length. Construct time series samples as follows:

[0037] ;

[0038] in, Indicates electrochemical parameters Time series samples, electrochemical parameters This includes voltage V, current I, internal resistance R, temperature T, temperature-internal resistance coupling ratio T / R, and state entropy S; among which, , , , , , These are time series samples of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy, respectively. Indicates electrochemical parameters Feature dimensions at each time step;

[0039] Step S1.3: Extract the capacity time series as follows:

[0040] ;

[0041] In the formula, Represents a capacity time series, For a fixed time window length, Indicates at time step The observed battery pack capacity at time t, i.e., the value at time t in the current time window. Capacity sample at each time step; Used to represent the capacity decay trend of a battery over time, its data form is usually a one-dimensional real number sequence. Each element in the sequence corresponds to the remaining capacity value of the battery at a specific moment (such as year, month or test cycle), and is often expressed as a percentage (e.g. 100%, 95%, 80%), used to characterize the change in battery state of health (SOH).

[0042] Step S2: Based on the asynchronous sampling characteristics of each electrochemical parameter, a multi-channel convolutional encoder family module is constructed. Taking time series samples as input, each electrochemical parameter is extracted independently to obtain the spatial feature mapping of each electrochemical parameter.

[0043] The Multi-Branch ConvNeXt-Tiny Encoder (MB-CNTE) family of modules is detailed below:

[0044] The multi-channel convolutional encoder family module has six electrochemical parameter branches, each corresponding to a different electrochemical parameter. The multi-channel convolutional encoder family module takes time series samples as input and performs independent feature extraction on voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy.

[0045] Each electrochemical parameter branch includes multi-scale depthwise separable convolution (Depthwise-SeparableMixConv), layer normalization (LayerNorm), and GELU activation functions to capture local dynamic fluctuation patterns and non-stationary feature changes; the multi-source spatial feature mapping is obtained through a family of multi-channel convolutional encoder modules as follows:

[0046] ;

[0047] In the formula, , , , , , These are spatial characteristic mappings of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy, respectively. This indicates the processing of the multi-channel convolutional encoder family of modules; , , , , , These are time series samples of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy, respectively.

[0048] The above formula forms a multi-source asynchronous degradation feature space, providing an input basis for cross-parameter fusion.

[0049] Reference Figure 3 The specific structure of the multi-channel convolutional encoder family of modules is as follows:

[0050] Input time series samples of various electrochemical parameters: Voltage series {V t}、 Current sequence {I t}, internal resistance sequence {R t}, temperature sequence {T t Temperature-internal resistance coupling ratio sequence {T / R} t} and the state entropy sequence {S t};

[0051] Each branch uses a family of one-dimensional ConvNeXt-Tiny structural units for deep feature encoding, with the specific hierarchical configuration as follows:

[0052] The time series samples of each electrochemical parameter are input with a dimension of L×1, where L is the fixed time window length;

[0053] First, for each time series sample of electrochemical parameter, a one-dimensional convolutional layer with 96 channels, a kernel size of 4, and a stride of 4 is used for processing. Then, layer normalization is performed to obtain an output with a dimension of L / 4 × 96.

[0054] Next, the data flows into three consecutive ConvNeXt modules for processing. Each ConvNeXt module has a dimension of 96, and the output dimension after processing remains L / 4×96.

[0055] Then, the data is downsampled through a one-dimensional convolutional layer with 192 channels, a kernel size of 2, and a stride of 2. The downsampled data is then fed into three ConvNeXt modules with a dimension of 192, and the final output dimension becomes L / 8×192.

[0056] Then, downsampling is performed again using a one-dimensional convolutional layer with 384 channels, a kernel size of 2, and a stride of 2. The data is then passed through nine ConvNeXt modules with a dimension of 384, and the output dimension is L / 16×384.

[0057] Next, the data is downsampled through a one-dimensional convolutional layer with 768 channels, a kernel size of 2, and a stride of 2. The downsampled data then passes through three ConvNeXt modules with a dimension of 768, and the output dimension becomes L / 32×768.

[0058] Finally, after passing through global average pooling, layer normalization, and a fully connected layer that maps the dimension from 768 to 128, a 128-dimensional spatial feature vector is obtained.

[0059] In summary, each branch outputs a 128-dimensional feature map:

[0060] ;

[0061] in, This represents the time series sample of the m-th electrochemical parameter. This represents the m-th electrochemical parameter branch processing of a multi-channel convolutional encoder family module. This represents the spatial characteristic mapping of the m-th electrochemical parameter;

[0062] The multi-channel convolutional encoder family of modules realizes deep convolutional modeling of multi-source asynchronous signals, effectively avoiding feature interference between various parameters and providing discriminative feature support for subsequent fusion.

[0063] Step S3: Based on the long-term degradation trend of battery pack capacity over time, a time-hybrid modeling module is constructed to obtain the capacity time-series feature mapping by taking the capacity time series as input.

[0064] The Temporal Hybrid Modeling Module (TTMB) is as follows:

[0065] The time-series hybrid modeling module takes capacity time series as input and proposes a "hierarchical temporal feature decoupling mechanism" for battery capacity degradation sequences. This mechanism first utilizes multi-scale hybrid convolution to extract rapid fluctuation patterns in capacity within a short time window. Then, it introduces a trend-aware temporal attention operator to weighted reconstruct the degradation contribution of different time periods, achieving explicit modeling of the long-term capacity decline process. Through a hierarchical structure, the model can simultaneously capture local non-stationary disturbances and long-term degradation trends, thereby constructing a time-dependent representation that better conforms to the physical laws of batteries. The extracted capacity time-series feature mapping is as follows:

[0066] ;

[0067] In the formula, Represents the temporal feature mapping of capacity. This indicates the fusion and hybrid convolution kernel processing. This indicates temporal attention operator processing; This represents a capacity time series.

[0068] The time-hybrid modeling module can effectively capture trend information and dynamic disturbance characteristics in the capacity degradation process while preserving the continuity and physical meaning of capacity changes.

[0069] The multi-channel convolutional encoder family of modules and the temporal hybrid modeling module operate in parallel in structure. The former focuses on the spatial representation of multi-source physical signals, while the latter focuses on the temporal variation of capacity. Together, they form a structured asynchronous-temporal joint representation system.

[0070] Reference Figure 4 The time-mixed modeling module is used for capacity time series {C} t The temporal evolution characteristics of the convolutional structure are modeled and memory is enhanced. This module contains a two-layer gated hybrid and multi-scale temporal convolutional structure, with the following specific configuration:

[0071] The input dimension is L×1, where L is the fixed time window length, representing the length of the input sequence;

[0072] First, the input data passes through a one-dimensional convolutional layer with 64 channels, a kernel size of 3, and a stride of 1, and then undergoes layer normalization to obtain an output with a dimension of L×64.

[0073] Subsequently, the data is passed through two gated hybrid blocks with a dimension of 64, while the output dimension remains unchanged;

[0074] Next, the length of the capacity time series is reduced to L / 2 by a downsampling one-dimensional convolutional layer with 128 channels, a kernel size of 2, and a stride of 2.

[0075] Then, the data is input into the temporal attention fusion module, which uses the attention mechanism to capture the features of key time steps and outputs a dimension of L / 2×128.

[0076] Then, three-scale convolution operations with kernel sizes of 3, 5 and 7 are performed in parallel, and the resulting features are concatenated to increase the output dimension to L / 2×384.

[0077] Next, through a composite module that includes multi-scale residual fusion, a linear layer that maps dimensions from 384 to 256, a GELU activation function, and layer normalization, the output dimension becomes L / 2×256.

[0078] Finally, after global context pooling and a linear layer that compresses the dimension from 256 to 128, a 128-dimensional temporal feature map is obtained:

[0079] ;

[0080] Represents the temporal feature mapping of capacity. This indicates that the time-mixed modeling module handles the processing. This represents a capacity time series.

[0081] The temporal hybrid modeling module can integrate local time awareness and global contextual memory to generate temporal embedded representations of capacity degradation trends.

[0082] Step S4: Construct a cross-parameter feature fusion layer, which concatenates the spatial feature map output by the multi-channel convolutional encoder family module with the capacity temporal feature map output by the temporal hybrid modeling module. Then, a feature channel attention mechanism is introduced to calculate the attention weights of different electrochemical parameter features. Finally, weighted fusion is performed to obtain the fused health status features.

[0083] The specific steps of step S4 are as follows:

[0084] Step S4.1: To achieve semantic complementarity and adaptive information weights between spatial and temporal features, a cross-parameter feature fusion layer CP-Fusion (Cross-Parameter Fusion) is constructed. This layer concatenates the spatial feature maps output by the multi-channel convolutional encoder family of modules with the capacity temporal feature maps output by the temporal fusion modeling module, as shown in the following equation:

[0085] ;

[0086] In the formula, Indicates splicing characteristics, , , , , , These are spatial characteristic mappings of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy, respectively. Represents capacity timing characteristics;

[0087] Step S4.2: Introduce a feature channel attention mechanism. The attention weights for different electrochemical parameters are as follows:

[0088] ;

[0089] in, Indicates electrochemical parameters attention weights of features Electrochemical parameters Feature mapping, To be related to electrochemical parameters The corresponding trainable weight matrix (or mapping matrix) is used to... This is mapped to the attention scoring space, thereby weighting the learning of the importance of different parameters.

[0090] Attention weight The specific calculation is as follows:

[0091] ;

[0092] in, This represents an index for all modes (all electrochemical parameter characteristic categories), used for Softmax normalization. It is related to electrochemical parameters The corresponding trainable linear mapping matrix is ​​used to calculate its attention score. Electrochemical parameters Feature mapping;

[0093] Step S4.3: Perform weighted fusion:

[0094] ;

[0095] in, Electrochemical parameters Attention weights This provides a globally unified health representation. By dynamically highlighting key parameters that significantly affect capacity degradation (such as temperature and internal resistance) and suppressing redundant features, a comprehensive representation vector with global health semantics is formed. .

[0096] The cross-parameter feature fusion layer adaptively adjusts the contribution ratio of each electrochemical parameter in the degradation characterization through attention weights, which can highlight key factors that are sensitive to capacity changes (such as temperature and internal resistance) and suppress the interference of noise parameters.

[0097] Reference Figure 5 The cross-parameter feature fusion layer integrates multi-source features and time-memory features to form a global health indicator. Its input consists of six electrochemical parameter feature vectors from MB-CNTE. With capacity characteristics from TTMB There are a total of 7 parameters to input. The specific structure is as follows:

[0098] The input dimension is 7×128;

[0099] Linear dimension upscaling is performed by linear projection layer and layer normalization, increasing the dimension from 128 to 256, with an output dimension of 7×256.

[0100] The interaction dependencies between electrochemical parameters are captured by a cross-parameter attention module; the cross-parameter attention module has 4 attention heads and a single head dimension of 64.

[0101] Feature reconstruction is performed using a residual structure and a feedforward network; the feature dimensions of the residual structure and the feedforward network are 256, 512, and 256, respectively, and GELU activation is used.

[0102] Attention weights are calculated using adaptive weighted fusion. The various electrochemical parameters are calculated using Softmax, and a corresponding attention weight is assigned to each electrochemical parameter:

[0103] ;

[0104] in, This represents an index for all modalities. It is related to electrochemical parameters The corresponding trainable linear mapping matrix, Electrochemical parameters Feature mapping;

[0105] The fusion feature is obtained by weighted summation:

[0106] ;

[0107] Subsequently, multi-scale one-dimensional convolution parallel convolution and feature concatenation are performed, with an output dimension of 1×384; in the multi-scale one-dimensional convolution, the kernel sizes are 1, 3, and 5 respectively; the stride is 1; and the number of channels is 128;

[0108] After multi-scale residual fusion and linear mapping layer, the dimension is reduced from 384 to 256. After GELU activation and layer normalization, the output dimension is 1×256.

[0109] Cross-timescale feature reconstruction is achieved through a Transformer encoder. The Transformer encoder has a feature dimension of 256, 4 attention heads, and 2 layers. Finally, the dimension is reduced from 256 to 128 through global average pooling and linear transformation. After layer normalization, a health status feature with a dimension of 128 is obtained.

[0110] Step S5: Input the fused health status features into the regression prediction head and output the predicted value of the battery pack, including the capacity at the next moment, the future multi-step capacity degradation sequence, and the remaining service life estimate.

[0111] The specific method for step S5 is as follows:

[0112] The fused health status features are input into the regression head, which outputs predicted values, including the capacity at the next time step, the future multi-step capacity degradation sequence, and the remaining useful life (RUL) estimate.

[0113] ;

[0114] In the formula, This indicates the regression prediction head processing. This represents the model's final predicted output;

[0115] Final prediction output Corresponding to three output formats:

[0116] The capacity will be a scalar output at the next moment. ;

[0117] Multi-step capacity degenerates into sequential output. ;

[0118] The remaining service life is a scalar output. .

[0119] in The number of time steps for future prediction determines the output length of the multi-step capacity degradation sequence; The real number field is used to represent prediction results where the output is a continuous numerical value.

[0120] In the formula, This indicates that the regression prediction head is processed, and the final capacity prediction value is generated through [Linear(128→1)], as shown in the following formula:

[0121] ;

[0122] In the formula, This represents the predicted value for the next time step, i.e., the predicted capacity for the next time step. This is a linear mapping matrix for the regression prediction head, used to integrate the fused health status features. Mapped to output space; These are the corresponding bias terms used to adjust the offset of the linear mapping result. Both are trainable parameters and are optimized together with the rest of the network during model training.

[0123] Step S6: Based on the multi-channel convolutional encoder family module, the temporal hybrid modeling module, the cross-parameter feature fusion layer, and the regression prediction head, a battery pack capacity prediction model is constructed. This model realizes asynchronous modeling of multi-source local features and capacity temporal features through parallel MB-CNTE TTMB branches, and then completes cross-parameter attention weighting and temporal hybrid fusion in the CP-Fusion module, finally outputting the capacity prediction value.

[0124] Construct a master loss function to supervise the training of the battery pack capacity prediction model, obtain a trained battery pack capacity prediction model, and apply the trained battery pack capacity prediction model to complete the battery pack capacity prediction of the target substation.

[0125] In step S6, mean squared error (MSE) is used as the main loss function, and a trend constraint term is introduced to maintain the similarity between the predicted sequence and the actual degradation trajectory at the trend level: The constructed main loss function is as follows:

[0126] ;

[0127] in, Represents the main loss function. This represents the mean squared error loss function. This represents the trend-constrained loss function. The battery pack capacity prediction model is trained using end-to-end backpropagation, with the weighting factor being the battery pack capacity prediction model.

[0128] The mean squared error (MSE) loss function is as follows:

[0129] ;

[0130] Where N represents the length of the time series used to calculate the trend constraint. and These represent the predicted capacity and the actual capacity, respectively.

[0131] To enhance the consistency of the predicted curve with the actual degradation trajectory in terms of trend, a trend constraint loss term is introduced, where the trend constraint loss function... as follows:

[0132] ;

[0133] Where N represents the length of the time series used to calculate the trend constraint, that is, the total number of sampling points in the predicted or actual series that participate in the trend calculation; and These represent the predicted capacity value and the actual capacity value at time step [1]. The value of ; and These represent the predicted capacity value and the actual capacity value at time step [1]. The value of .

[0134] By minimizing this loss function, end-to-end optimization of the model is achieved at both the trend and numerical levels, ultimately resulting in a battery pack capacity prediction model that combines multi-parameter sensing and time-series trend capture capabilities. This battery pack capacity prediction model can fully utilize multi-source monitoring signals and capacity evolution memory to achieve high-precision and robust battery capacity prediction.

[0135] In summary, this invention proposes a structured optimization scheme for asynchronous multi-parameter monitoring data of batteries, fully leveraging the complementary advantages of time-series models and the TimeMixer structure in feature representation, time-dependent modeling, and stability, thus forming a prediction system more adapted to real-world scenarios. Specifically, it includes:

[0136] (1) A guided multi-parameter fusion capacity degradation prediction model is proposed:

[0137] To address the problem that traditional battery life prediction models rely on single parameters for capacity features and struggle to generalize across different battery types or datasets, this invention proposes a multi-parameter electrochemical feature-guided capacity representation method. The method uses capacity as the core degradation characterization variable and constructs a feature representation mechanism driven by key operating parameters such as voltage, current, temperature, and internal resistance. Multiple input temporal channels and cross-parameter coupling units are introduced into the network structure, enabling a unified mapping of the dynamic changes of different physical quantities in the feature space. This approach achieves adaptive modeling and transfer representation of capacity features under multi-parameter conditions, allowing the model to capture consistent characteristics of the degradation mechanism. In the feature extraction stage, each parameter undergoes local dynamic pattern capture via independent convolutional temporal pathways, maintaining the feature independence of each electrochemical process. In the fusion stage, a cross-parameter attention mechanism guided by capacity features is designed, dynamically allocating weights among channels such as voltage, current, temperature, and internal resistance to enhance the synergistic effect of different electrochemical parameters on degradation characterization. While maintaining the dominance of the capacity center feature, adaptive coupling between parameters is achieved, avoiding information loss and underfitting problems under single-parameter input. Through this guided multi-parameter fusion strategy, the model can maintain predictive stability across battery data of different types and sampling conditions, and has good generalization ability and physical consistency, providing a unified capacity degradation characterization framework for cross-dataset migration prediction.

[0138] (2) Introduction of trend-constrained loss function:

[0139] To enhance the robustness of the model under complex operating conditions, this invention introduces a trend consistency constraint term in the loss function design. This term constrains the direction of change in the predicted sequence to align with the actual degradation trend. During optimization, this constraint term makes directional judgments on the predicted increments at adjacent time steps. When the predicted trend deviates from the historical degradation direction, a gradient penalty mechanism suppresses reverse fluctuations, thereby limiting the model's over-response to instantaneous anomalies or measurement noise. In battery life prediction scenarios, capacity degradation exhibits a long-term monotonically decreasing characteristic. However, field monitoring data is often affected by factors such as temperature disturbances, electrochemical measurement errors, and uneven sampling, making it prone to local non-stationary fluctuations. Traditional supervision objectives based primarily on mean square error cannot distinguish between "physical degradation signals" and "random noise," leading to local oscillations or degradation direction reversals during the model fitting phase. This invention, through a trend constraint mechanism, enables the model to possess directional constraint capabilities while learning the overall degradation trajectory. This means that while maintaining accuracy, it automatically smooths short-term anomalies, forming a capacity prediction curve that better conforms to physical laws. Essentially, it embeds a "temporal direction prior" into the optimization objective, achieving physical consistency and noise resistance in the predicted output. After constraints, the model can still maintain the continuity and stability of the degradation trend even in the presence of abnormal measurement points, uneven sampling intervals, or cross-device data offsets, thereby significantly improving the robustness and usability of the prediction system in complex engineering environments.

[0140] Through the above structural and optimized design, the battery capacity degradation prediction model proposed in this invention achieves synergistic enhancement in the feature modeling layer and the target constraint layer:

[0141] At the feature modeling level, a guided multi-parameter fusion mechanism is employed to fully explore the coupling relationships among multiple electrochemical features such as voltage, current, temperature, and internal resistance. This achieves a unified characterization of the capacity degradation process and cross-parameter adaptive modeling, significantly improving the model's generalization ability under different battery types and sampling conditions. At the objective constraint level, the introduction of a trend consistency loss function provides physically guided degradation trend constraints for the model learning process, effectively suppressing prediction oscillations caused by environmental temperature fluctuations, sampling errors, and random load disturbances. This ensures that the output results maintain accuracy while possessing stronger noise resistance and directional stability. In summary, the prediction system constructed in this invention combines multi-parameter feature expression capabilities, time series robustness, and physical consistency. It can continuously output smooth and reliable capacity degradation prediction results in complex substation environments, providing a highly stable and transferable technical solution for battery health status assessment and life prediction.

[0142] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A battery capacity prediction method based on multi-source asynchronous perception and time hybrid modeling, characterized in that, Perform the following steps S1-S6 to complete the battery pack capacity prediction for the target substation: Step S1: Collect the electrochemical parameters of the battery pack during the operation of the target substation, including voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio and state entropy, and preprocess the electrochemical parameters to obtain a time series sample composed of the preprocessed electrochemical parameters, and extract the capacity time series. Step S2: Based on the asynchronous sampling characteristics of each electrochemical parameter, a multi-channel convolutional encoder family module is constructed. Taking time series samples as input, each electrochemical parameter is extracted independently to obtain the spatial feature mapping of each electrochemical parameter. The multi-channel convolutional encoder family of modules in step S2 is as follows: The multi-channel convolutional encoder family module has six electrochemical parameter branches, each corresponding to a different electrochemical parameter. The multi-channel convolutional encoder family module takes time series samples as input and performs independent feature extraction on voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy. Each branch of electrochemical parameters includes multi-scale deep separable convolution, layer normalization, and GELU activation functions. The multi-source spatial feature mapping is obtained through a family of multi-channel convolutional encoder modules as follows: ； wherein, , , , , , are spatial feature maps of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio and state entropy, respectively; denotes a multi-channel convolutional encoder family module processing; , , , , , are time series samples of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio and state entropy, respectively; Step S3: Based on the long-term degradation trend of battery pack capacity over time, a time-hybrid modeling module is constructed to obtain the capacity time-series feature mapping by taking the capacity time series as input. The time-mixing modeling module in step S3 is as follows: The temporal hybrid modeling module takes the capacity time series as input, fuses hybrid convolutional kernels and temporal attention operators, and extracts the capacity temporal feature map as follows: ； In the formula, represents a capacity timing characteristic mapping, represents a fusion mixed convolution kernel processing, represents a time domain attention operator processing; represents a capacity time sequence; Step S4: Construct a cross-parameter feature fusion layer, which concatenates the spatial feature map output by the multi-channel convolutional encoder family module with the capacity temporal feature map output by the temporal hybrid modeling module. Then, a feature channel attention mechanism is introduced to calculate the attention weights of different electrochemical parameter features. Finally, weighted fusion is performed to obtain the fused health status features. Step S5: Input the fused health status features into the regression prediction head and output the predicted value of the battery pack, including the capacity at the next moment, the future multi-step capacity degradation sequence, and the remaining service life estimate. Step S6: Based on the multi-channel convolutional encoder family module, temporal hybrid modeling module, cross-parameter feature fusion layer, and regression prediction head, construct a battery pack capacity prediction model, construct the main loss function, supervise the training of the battery pack capacity prediction model, obtain the trained battery pack capacity prediction model, and apply the trained battery pack capacity prediction model to complete the battery pack capacity prediction of the target substation.

2. The method for predicting the capacity of the battery based on multi-source asynchronous perception and time hybrid modeling according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Collect the electrochemical parameters of the battery pack during the operation of the target substation, including voltage V, current I, internal resistance R, temperature T, temperature-internal resistance coupling ratio T / R, and state entropy S; Step S1.2: Preprocess each electrochemical parameter, including missing value imputation, outlier removal, time alignment and standardization, and according to a fixed time window length. Construct time series samples as follows: ； in, Indicates electrochemical parameters Time series samples, electrochemical parameters This includes voltage V, current I, internal resistance R, temperature T, temperature-internal resistance coupling ratio T / R, and state entropy S; among which, , , , , , These are time series samples of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy, respectively. Indicates electrochemical parameters Feature dimensions at each time step; Step S1.3: Extract the capacity time series as follows: ； In the formula, Represents a capacity time series, For a fixed time window length, Indicates at time step The observed capacity of the battery pack at a given time.

3. The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to claim 2, characterized in that, The specific structure of the multi-channel convolutional encoder family of modules is as follows: The time series samples of each electrochemical parameter are input with a dimension of L×1, where L is the fixed time window length; First, for each time series sample of electrochemical parameter, a one-dimensional convolutional layer with 96 channels, a kernel size of 4, and a stride of 4 is used for processing. Then, layer normalization is performed to obtain an output with a dimension of L / 4×96. Next, the data flows into three consecutive ConvNeXt modules for processing. Each ConvNeXt module has a dimension of 96, and the output dimension after processing remains L / 4×96. Then, the data is downsampled through a one-dimensional convolutional layer with 192 channels, a kernel size of 2, and a stride of 2. The downsampled data is then fed into three ConvNeXt modules with a dimension of 192, and the final output dimension becomes L / 8×192. Then, downsampling is performed again using a one-dimensional convolutional layer with 384 channels, a kernel size of 2, and a stride of 2. The data is then passed through nine ConvNeXt modules with a dimension of 384, and the output dimension is L / 16×384. Next, the data is downsampled through a one-dimensional convolutional layer with 768 channels, a kernel size of 2, and a stride of 2. The downsampled data then passes through three ConvNeXt modules with a dimension of 768, and the output dimension becomes L / 32×768. Finally, after passing through global average pooling, layer normalization, and a fully connected layer that maps the dimension from 768 to 128, a 128-dimensional spatial feature vector is obtained.

4. The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to claim 3, characterized in that, The time-mixing modeling module in step S3 is as follows: The temporal hybrid modeling module takes the capacity time series as input, fuses hybrid convolutional kernels and temporal attention operators, and extracts the capacity temporal feature map as follows: ； In the formula, Represents the temporal feature mapping of capacity. This indicates the fusion and hybrid convolution kernel processing. This indicates temporal attention operator processing; This represents a capacity time series.

5. The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to claim 4, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Construct a cross-parameter feature fusion layer, which concatenates the spatial feature map output by the multi-channel convolutional encoder family module with the capacity temporal feature map output by the temporal hybrid modeling module, as shown in the following formula: ； In the formula, Indicates splicing characteristics, , , , , , These are spatial characteristic mappings of voltage, current, internal resistance, temperature, temperature-internal resistance coupling ratio, and state entropy, respectively. Represents the temporal feature mapping of capacity; Step S4.2: Introduce a feature channel attention mechanism. The attention weights for different electrochemical parameters are as follows: ； in, Indicates electrochemical parameters attention weights of features To be related to electrochemical parameters The corresponding trainable weight matrix; Electrochemical parameters Feature mapping; Attention weight The specific calculation is as follows: ； in, This represents an index for all modalities. It is related to electrochemical parameters The corresponding trainable linear mapping matrix, Electrochemical parameters Feature mapping; Step S4.3: Perform weighted fusion: ； in, Electrochemical parameters attention weights of features Characteristics of a healthy state.

6. The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to claim 5, characterized in that, The specific structure of the cross-parameter feature fusion layer is as follows: The input dimension is 7×128; Linear dimension upscaling is performed by linear projection layer and layer normalization, increasing the dimension from 128 to 256, with an output dimension of 7×256. The interaction dependencies between electrochemical parameters are captured by a cross-parameter attention module; the cross-parameter attention module has 4 attention heads and a single head dimension of 64. Feature reconstruction is performed using a residual structure and a feedforward network; the feature dimensions of the residual structure and the feedforward network are 256, 512, and 256, respectively, and GELU activation is used. Attention weights are calculated using adaptive weighted fusion. Attention weights are calculated using Softmax among various electrochemical parameters, and fusion features are obtained by weighted summation. Subsequently, multi-scale one-dimensional convolution parallel convolution and feature concatenation are performed, with an output dimension of 1×384; in the multi-scale one-dimensional convolution, the kernel sizes are 1, 3, and 5 respectively; the stride is 1; and the number of channels is 128; After multi-scale residual fusion and linear mapping layer, the dimension is reduced from 384 to 256. After GELU activation and layer normalization, the output dimension is 1×256. Cross-timescale feature reconstruction is achieved through a Transformer encoder. The Transformer encoder has a feature dimension of 256, 4 attention heads, and 2 layers. Finally, the dimension is reduced from 256 to 128 through global average pooling and linear transformation. After layer normalization, a health status feature with a dimension of 128 is obtained.

7. The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to claim 6, characterized in that, The specific method for step S5 is as follows: The fused health status features are input into the regression prediction head, and the output is the predicted value, including the capacity at the next time step, the future multi-step capacity degradation sequence, and the remaining lifetime estimate: ； In the formula, This indicates the regression prediction head processing. This represents the model's final predicted output; The specific formula is as follows: ； In the formula, This represents the predicted value at the next moment. This is the linear mapping matrix of the regression prediction head; This is a bias term.

8. The battery capacity prediction method based on multi-source asynchronous sensing and time-hybrid modeling according to claim 7, characterized in that, The main loss function constructed in step S6 is as follows: ； in, Represents the main loss function. This represents the mean squared error loss function. This represents the trend-constrained loss function. For the weighting factor; Trend constraint loss function as follows: ； Where N represents the length of the time series used to calculate the trend constraint; and These represent the predicted capacity value and the actual capacity value at time step [1]. The value of ; and These represent the predicted capacity value and the actual capacity value at time step [1]. The value of .

Images