Method and system for predicting SOH (state of health) of energy storage battery

Abstract

The invention discloses an energy storage battery SOH prediction method and system, and belongs to the technical field of energy storage battery management. The method comprises the following steps: acquiring operation data of a plurality of groups of single cells in a plurality of continuous time steps in real time, and preprocessing the operation data to generate a time-cell two-dimensional input matrix; inputting the time-cell two-dimensional input matrix into a pre-trained adaptive space-time fusion prediction model, outputting an SOH prediction value of each single cell, and evaluating the health state of the cell according to the SOH value; according to the method, the SOH prediction precision under the complex working condition is improved, and the SOH prediction requirements of the battery cell under different energy storage scenes are met.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the state of energy (SOH) of an energy storage battery, belonging to the field of energy storage battery management technology. Background Technology

[0002] Lithium-ion batteries have become a core component of energy storage systems due to their high energy density, with lithium iron phosphate batteries dominating the market due to their long cycle life and high safety. Energy storage battery packs typically consist of hundreds of cells connected in series and parallel. Under complex operating conditions such as peak-valley arbitrage and backup power, problems arise such as increased cell inconsistency, significant differences in aging rates, and large temperature gradient variations, making traditional cell SOH prediction methods difficult to apply directly.

[0003] The existing technology has three main limitations: First, it does not consider the spatiotemporal correlation between cells in the energy storage battery pack, and cannot capture key features such as the equalization current and voltage difference within the pack; second, the simple model has a gradient decay problem in handling long sequences and multi-node correlation analysis; and third, single PSO optimization is prone to getting trapped in local optima, making it difficult to support energy storage operation and maintenance decisions.

[0004] Therefore, developing a SOH prediction method that is adapted to the characteristics of energy storage batteries and has both high accuracy and strong interpretability is of great significance for improving the reliability and economy of energy storage systems. Summary of the Invention

[0005] The purpose of this invention is to overcome the problem of cross-cell spatiotemporal feature modeling of energy storage battery packs in the prior art, and to provide a method and system for predicting the state of health (SOH) of energy storage batteries, thereby improving the accuracy of SOH prediction under complex operating conditions and meeting the needs of cell SOH prediction in different energy storage scenarios.

[0006] To achieve the above objectives / to solve the above technical problems, the present invention is implemented using the following technical solution:

[0007] An adaptive spatiotemporal fusion SOH prediction method for energy storage batteries, the method comprising:

[0008] Real-time acquisition of multiple sets of individual battery cells at multiple consecutive time steps; preprocessing of the operating data to generate a time-cell two-dimensional input matrix.

[0009] The time-cell two-dimensional input matrix is ​​input into the pre-trained adaptive spatiotemporal fusion prediction model, and the SOH prediction value of each individual cell is output. The health status of the cell is evaluated based on the SOH value.

[0010] The adaptive spatiotemporal fusion prediction model includes:

[0011] Convolutional Neural Networks (CNNs) are used to convolve each battery cell along the time dimension to extract local temporal abrupt change features of a single battery cell.

[0012] The ConvTransformer layer is used to capture cross-cell dependencies and long-sequence patterns based on the local temporal mutation characteristics of the cell, and output spatiotemporal fusion features.

[0013] The stage-adaptive LSTM performs cell-level aggregation based on spatiotemporal fusion features, inputs the aggregated time series into the LSTM, and outputs the SOH prediction value of each individual cell.

[0014] The above technical solution significantly improves the accuracy of cell SOH value prediction through multi-module collaboration, optimizes inference time and improves inference efficiency by leveraging sparse attention and model quantization, and meets the real-time monitoring requirements of energy storage BMS; it is compatible with various types of energy storage batteries such as lithium iron phosphate and ternary lithium, effectively meeting the need for accurate SOH prediction in various energy storage scenarios, estimating the battery's operating status and health status based on the cell SOH, and ensuring the safe and stable operation of the energy storage system.

[0015] Optionally, the adaptive spatiotemporal fusion prediction model uses a PSO-SA hybrid optimization module combined with particle swarm optimization algorithm and simulated annealing algorithm to optimize the pre-set structural hyperparameters in the adaptive spatiotemporal fusion prediction model, thereby obtaining a trained adaptive spatiotemporal fusion prediction model.

[0016] The above technical solution effectively improves the accuracy of hyperparameter optimization and the hit rate of the global optimal solution through the PSO-SA hybrid optimization module, enhances the model's prediction accuracy and spatiotemporal feature capture capability, accelerates model training efficiency and convergence speed, and obtains a prediction model with better performance in a shorter time.

[0017] Optionally, the runtime data includes individual-level features and system-level derived features;

[0018] The individual cell-level characteristics include the voltage, current, temperature, and cumulative cycle count of each cell at each time step;

[0019] The system-level derived features include the voltage range of the battery pack, the integral value of the equalization current, the module temperature gradient, and the dynamic change rate of the charge / discharge rate at the corresponding time step.

[0020] The above technical solution can effectively enrich the training dataset of battery cells by processing and extracting features from the basic and derived data of the battery cells, and improve the performance of the model by using multi-feature data to train the model.

[0021] Optionally, the step of preprocessing the operating data to generate a time-cell two-dimensional input matrix includes:

[0022] The individual-level features and system-level derived features are aligned with the cell number according to the time step and organized into a spatiotemporal feature tensor with time as the first dimension, cell as the second dimension, and feature vector as the third dimension, denoted as the time-cell two-dimensional input matrix.

[0023] The above technical solution improves the relationship between data and data utilization by converting the data into a two-dimensional matrix, thereby enhancing the model's data processing capabilities and the performance of the trained model.

[0024] Optionally, the preprocessing includes: using the isolated forest algorithm to identify and remove abnormal data in the running data.

[0025] The above technical solution reduces data redundancy and the impact of abnormal data on model training stability by removing outlier data, thereby increasing model training stability, accelerating model convergence, and reducing training time.

[0026] Optionally, the adaptive spatiotemporal fusion prediction model uses a multi-objective fitness function to optimize the model's hyperparameters, wherein the fitness function is: ; in, The root mean square error for SOH prediction. For model inference time, To preset the maximum allowed inference time, These are the weighting coefficients.

[0027] The above technical solution achieves synergistic optimization of prediction accuracy and inference efficiency through a multi-objective fitness function, improves the engineering practicality and embedded deployment adaptability of the model, enhances the goal orientation and convergence effectiveness of the optimization process, and obtains a training model with better performance.

[0028] Optionally, the convolutional neural network (CNN) adopts a 3-layer 1D-CNN to extract local mutation patterns of single-cell timing features.

[0029] The above technical solution utilizes a CNN model to perform preliminary feature extraction on the data, thereby obtaining local mutation features in the time-series data.

[0030] Optionally, the ConvTransformer layer employs a sparse self-attention mechanism.

[0031] Second aspect: A spatiotemporal fusion SOH prediction system for energy storage batteries, the system comprising:

[0032] The feature module is configured to collect operating data of multiple individual cells in real time at multiple consecutive time steps, and preprocess the operating data to generate a time-cell two-dimensional input matrix.

[0033] An adaptive spatiotemporal fusion model is configured to input a time-cell two-dimensional input matrix into a pre-trained adaptive spatiotemporal fusion prediction model, output the SOH prediction value of each individual cell, and evaluate the cell health status based on the SOH value.

[0034] The above technical solution involves processing multiple data points from the battery cell using a feature module to form a two-dimensional data matrix, increasing the utilization of the data by an adaptive spatiotemporal fusion module, and simultaneously mining the relationships between the data to improve the model's accuracy in estimating the battery cell's state of harmonics (SOH).

[0035] Optionally, the adaptive spatiotemporal fusion prediction model includes:

[0036] A convolutional neural network (CNN) is configured to convolve each cell along the time dimension to extract local temporal abrupt features of a single cell.

[0037] The ConvTransformer layer is configured to capture cross-cell dependencies and long-sequence patterns based on the local temporal mutation characteristics of the cell, and output spatiotemporal fusion features.

[0038] The stage-adaptive LSTM is configured to aggregate data at the cell level based on spatiotemporal fusion features. The aggregated time series is then input into the LSTM, and the SOH prediction value of each individual cell is output.

[0039] The above technical solution combines CNN, ConvTransformer layers, and stage-adaptive LSTM layers to fully utilize the multi-dimensional features of battery cell data, explore the relationship between battery cell SOH and data in various dimensions, and improve the model's predictive performance and robustness.

[0040] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0041] By working together with multiple modules, the accuracy of cell SOH value prediction is greatly improved. With the help of sparse attention and model quantization, inference time is optimized and inference efficiency is improved to meet the real-time monitoring requirements of energy storage BMS. It is compatible with multiple types of energy storage batteries such as lithium iron phosphate and ternary lithium, and effectively meets the needs of accurate battery SOH prediction in various energy storage scenarios. Based on the cell SOH, the battery's working status and health status are estimated to ensure the safe and stable operation of the energy storage system. Attached Figure Description

[0042] Figure 1 The diagram shows the processing flow of the adaptive spatiotemporal fusion SOH prediction method for energy storage batteries according to the present invention. Figure 2 The diagram shows the prediction model structure of the adaptive spatiotemporal fusion SOH prediction method for energy storage batteries. Figure 3The figure shows the prediction results of the adaptive spatiotemporal fusion SOH prediction method for energy storage batteries. Detailed Implementation

[0043] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0044] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0045] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0046] Example 1, as Figure 1 As shown, a method for predicting the state of harm (SOH) of an energy storage battery is disclosed, the method comprising:

[0047] Real-time acquisition of multiple sets of individual battery cells at multiple consecutive time steps; preprocessing of the operating data to generate a time-cell two-dimensional input matrix.

[0048] The time-cell two-dimensional input matrix is ​​input into the pre-trained adaptive spatiotemporal fusion prediction model, and the SOH prediction value of each individual cell is output. The health status of the cell is evaluated based on the SOH value.

[0049] The adaptive spatiotemporal fusion prediction model includes:

[0050] Convolutional Neural Networks (CNNs) are used to convolve each battery cell along the time dimension to extract local temporal abrupt change features of a single battery cell.

[0051] The ConvTransformer layer is used to capture cross-cell dependencies and long-sequence patterns based on the local temporal mutation characteristics of the cell, and output spatiotemporal fusion features.

[0052] The stage-adaptive LSTM aggregates cell-level data based on spatiotemporal fusion features. The aggregated time series is then input into the LSTM, which outputs the predicted SOH value for each individual cell. SOH is an abbreviation for State of Health, representing the health status.

[0053] In this embodiment, the adaptive spatiotemporal fusion prediction model uses a PSO-SA hybrid optimization module combined with particle swarm optimization algorithm and simulated annealing algorithm to optimize the pre-set structural hyperparameters in the adaptive spatiotemporal fusion prediction model, thereby obtaining a trained adaptive spatiotemporal fusion prediction model.

[0054] The above technical solution effectively improves the accuracy of hyperparameter optimization and the hit rate of the global optimal solution through the PSO-SA hybrid optimization module, enhances the model's prediction accuracy and spatiotemporal feature capture capability, accelerates model training efficiency and convergence speed, and obtains a prediction model with better performance in a shorter time.

[0055] In this embodiment, the operational data includes both individual-level features and system-level derived features.

[0056] The individual cell-level characteristics include the voltage, current, temperature, and cumulative cycle count of each cell at each time step;

[0057] The system-level derived features include the voltage range of the battery pack, the integral value of the equalization current, the module temperature gradient, and the dynamic change rate of the charge / discharge rate at the corresponding time step.

[0058] The above technical solution can effectively enrich the training dataset of battery cells by processing and extracting features from the basic and derived data of the battery cells, and improve the performance of the model by using multi-feature data to train the model.

[0059] In this embodiment, the training of the adaptive spatiotemporal fusion prediction model includes the following:

[0060] First, the Isolation Forest algorithm is used to identify and remove outlier data, reducing their impact on the final result. Second, multi-cell data from different locations are synchronized based on timestamps to construct a two-dimensional "time-cell" feature matrix. This forms the final dataset, which is then split into: 70% for the dataset itself, 10% for the validation set, and 20% for the test set.

[0061] In this invention, three networks—1D-CNN, ConvTransformer, and stage-adaptive LSTM—are combined. The CNN first performs feature learning on the temporal data, primarily extracting local features of a single battery cell in the time dimension. Subsequently, the ConvTransformer is used to perform spatiotemporal fusion of cross-cell data to obtain the dependencies between cells and the contextual relationships of long-term sequences. A dual-dimensional attention mechanism of "cell-time" is employed: the input tensor is transformed from "time × cell × feature" to "attention dimension × cell × time," while attention is calculated on both the "cell dimension" (spatial correlation) and the "time dimension" (temporal correlation). Cell-dimensional attention: calculates the... Energy-saving cells and the first Dependence weight of energy-saving cells (e.g., the impact of cells with high balancing current on the surrounding area), time-dimensional attention: calculating the first Time and the Feature correlation at different times. Multidimensional computation effectively improves the accuracy of SOH estimation. Finally, an adaptive LSTM with gating stage is used to extract features from long-term data dependencies, fully capturing the potential information of long-term data, and dynamically adjusting the forget gate weights through an aging stage recognizer (based on segmented data of capacity decay rate). LSTM, as a classic model for solving the gradient vanishing problem in recurrent neural networks, achieves selective memory and forgetting of long-term information through a "gating mechanism" and "cell state." Its standard structure includes three gate units: a forget gate, an input gate, and an output gate. The limitation of traditional LSTM is that the gating parameters are fixed after training, which cannot adapt to the temporal differences in the aging patterns of energy storage batteries at different stages (e.g., early aging (SOH>90%) with gradual capacity decay, and late aging (SOH<70%) with accelerated decay), leading to a decrease in long-term prediction accuracy. Therefore, the entire battery lifecycle is divided into three stages: early stage (SOH 90%-100%), middle stage (SOH 70%-90%), and late stage (SOH < 70%). A stage state variable is introduced into the forgetting gate, allowing the gating weights to adjust with the aging stage. Furthermore, a multi-objective fitness function is constructed, balancing prediction accuracy and computational efficiency, as shown in the following formula: ; in These are the weighting coefficients. For model inference time, By incorporating simulated annealing, getting trapped in local optima can be avoided. Furthermore, parameters such as the learning rate of CNN, ConvTransformer, and stage-adaptive LSTM networks are optimized.

[0062] The SOH data of the battery cell was estimated using the trained network structure, and the results are shown in the figure below. Figure 3As shown in the figure, the results demonstrate that this method can accurately predict the SOH value of the battery cell, with a small error compared to the reference value.

[0063] This embodiment proposes an adaptive spatiotemporal fusion SOH prediction method for energy storage batteries. It uses multiple modules to predict the SOH of the battery cells, thereby ensuring the operational stability of the energy storage batteries and the safety of the system operation.

[0064] Example 2 discloses a SOH prediction system for energy storage batteries, the system comprising:

[0065] The feature module is configured to collect operating data of multiple individual cells in real time at multiple consecutive time steps, and preprocess the operating data to generate a time-cell two-dimensional input matrix.

[0066] An adaptive spatiotemporal fusion model is configured to input a time-cell two-dimensional input matrix into a pre-trained adaptive spatiotemporal fusion prediction model, output the SOH prediction value of each individual cell, and evaluate the cell health status based on the SOH value.

[0067] The feature module collects basic parameters such as voltage, current, temperature, and cycle count of individual battery cells, as well as features such as battery pack voltage range, equalization current integral, module temperature gradient, and dynamic change rate of charge / discharge rate, generating a two-dimensional "time-cell" input matrix. The adaptive spatiotemporal fusion prediction model constructs a three-level model: CNN-ConvTransformer-Stage Adaptive LSTM. CNN extracts local temporal mutation features of individual cells, ConvTransformer layers capture cross-cell dependencies and long-sequence patterns, and the stage adaptive LSTM dynamically adjusts the forget gate weights through a recognizer. An additional "cell-time" dual-dimensional attention layer and depthwise separable convolutions are introduced to effectively reduce computational complexity. The PSO-SA hybrid optimization module combines particle swarm optimization and simulated annealing to optimize parameters such as the number of CNN convolutional kernels and the number of ConvTransformer attention heads.

[0068] This invention uses multiple modules to jointly optimize and predict the SOH value of the battery cell, and uses the SOH value to understand the working status and health status of the battery cell, thereby ensuring the safe and stable operation of the energy storage battery compartment.

[0069] The feature module collects basic parameters such as voltage, current, temperature, and cycle count of individual battery cells, as well as features such as battery pack voltage range, equalization current integral, module temperature gradient, and dynamic change rate of charge / discharge rate, generating a two-dimensional "time-cell" input matrix. An isolated forest algorithm is used to identify and remove abnormal data from the sensors based on the collected cell voltage and current parameters. Simultaneously, multi-cell data is synchronized based on timestamps to construct a two-dimensional "time-cell" feature matrix.

[0070] The adaptive spatiotemporal fusion prediction model constructs a three-level model: CNN-ConvTransformer-Stage Adaptive LSTM. The CNN extracts local temporal mutation features of a single battery cell, the ConvTransformer layer captures cross-cell dependencies and long-term sequence patterns, and the stage adaptive LSTM dynamically adjusts the forget gate weights through a recognizer. An additional "cell-time" dual-dimensional attention layer and depthwise separable convolutions are introduced to effectively reduce computational complexity. The CNN uses a 3-layer 1D-CNN with a kernel size of 4×1 and a stride of 2 to extract local mutation patterns (such as voltage data inflection points) of single-cell temporal features. The ConvTransformer layer employs a sparse self-attention mechanism (8 attention heads, 0.6 sparsity) to capture inter-cell dependencies and long-term sequence patterns. A gated LSTM module is designed to dynamically adjust the forget gate weights through an aging stage recognizer (based on segmented capacity decay rates).

[0071] The PSO-SA hybrid optimization module combines particle swarm optimization and simulated annealing to optimize parameters such as the number of CNN convolutional kernels and the number of attention heads in ConvTransformer.

[0072] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the state of harm (SOH) of an energy storage battery, characterized in that, The method includes: Real-time acquisition of multiple sets of individual battery cells at multiple consecutive time steps; preprocessing of the operating data to generate a time-cell two-dimensional input matrix. The time-cell two-dimensional input matrix is ​​input into the pre-trained adaptive spatiotemporal fusion prediction model, and the SOH prediction value of each individual cell is output. The health status of the cell is evaluated based on the SOH value. The adaptive spatiotemporal fusion prediction model includes: Convolutional Neural Networks (CNNs) are used to convolve each battery cell along the time dimension to extract local temporal abrupt change features of a single battery cell. The ConvTransformer layer is used to capture cross-cell dependencies and long-sequence patterns based on the local temporal mutation characteristics of the cell, and output spatiotemporal fusion features. The stage-adaptive LSTM performs cell-level aggregation based on spatiotemporal fusion features, inputs the aggregated time series into the LSTM, and outputs the SOH prediction value of each individual cell.

2. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The adaptive spatiotemporal fusion prediction model uses a PSO-SA hybrid optimization module combined with particle swarm optimization algorithm and simulated annealing algorithm to optimize the pre-set structural hyperparameters in the adaptive spatiotemporal fusion prediction model, thus obtaining a trained adaptive spatiotemporal fusion prediction model.

3. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The operational data includes individual-level features and system-level derived features; The individual cell-level characteristics include the voltage, current, temperature, and cumulative cycle count of each cell at each time step; The system-level derived features include the voltage range of the battery pack, the integral value of the equalization current, the module temperature gradient, and the dynamic change rate of the charge / discharge rate at the corresponding time step.

4. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The process of preprocessing the operating data to generate a time-cell two-dimensional input matrix includes: The individual-level features and system-level derived features are aligned with the cell number according to the time step and organized into a spatiotemporal feature tensor with time as the first dimension, cell as the second dimension, and feature vector as the third dimension, denoted as the time-cell two-dimensional input matrix.

5. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The preprocessing includes: using the isolated forest algorithm to identify and remove abnormal data in the running data.

6. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The adaptive spatiotemporal fusion prediction model uses a multi-objective fitness function to optimize the model's hyperparameters. The fitness function is: ； in, The root mean square error for SOH prediction. For model inference time, To preset the maximum allowed inference time, These are the weighting coefficients.

7. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The convolutional neural network (CNN) uses a 3-layer 1D-CNN to extract local mutation patterns of single-cell timing features.

8. The method for predicting the state of harm (SOH) of an energy storage battery according to claim 1, characterized in that, The ConvTransformer layer employs a sparse self-attention mechanism.

9. A SOH prediction system for energy storage batteries, characterized in that, The system includes: The feature module is configured to collect operating data of multiple individual cells in real time at multiple consecutive time steps, and preprocess the operating data to generate a time-cell two-dimensional input matrix. An adaptive spatiotemporal fusion model is configured to input a time-cell two-dimensional input matrix into a pre-trained adaptive spatiotemporal fusion prediction model, output the SOH prediction value of each individual cell, and evaluate the cell health status based on the SOH value.

10. The energy storage battery SOH prediction system according to claim 9, characterized in that, The adaptive spatiotemporal fusion prediction model includes: A convolutional neural network (CNN) is configured to convolve each cell along the time dimension to extract local temporal abrupt features of a single cell. The ConvTransformer layer is configured to capture cross-cell dependencies and long-sequence patterns based on the local temporal mutation characteristics of the cell, and output spatiotemporal fusion features. The stage-adaptive LSTM is configured to aggregate data at the cell level based on spatiotemporal fusion features. The aggregated time series is then input into the LSTM, and the SOH prediction value of each individual cell is output.

Images