Energy storage power station battery health state prediction method and system based on data driving

Abstract

The invention discloses an energy storage power station battery health state prediction method and system based on data driving, and relates to the technical field of battery health state prediction, and the method comprises the steps: obtaining a capacity data sequence of a lithium battery in an energy storage power station, and carrying out the variational mode decomposition of the capacity data sequence, and obtaining at least two component signals; dividing the at least two component signals into training data and test data according to a preset proportion, training by adopting the training data of the main degradation trend signal to obtain a first prediction model, and training by adopting the training data of the at least one high-frequency signal to obtain a second prediction model; and superposing a prediction result of the main degradation trend signal by the first prediction model and a prediction result of the high-frequency signal by the second prediction model to obtain a prediction value of the health state of the lithium battery. Through the intelligent decomposition strategy and the differential modeling method, the prediction precision of the lithium battery health state of the energy storage power station is remarkably improved, and particularly, excellent adaptability and stability are shown when complex working conditions such as capacity regeneration are processed.

Description

Technical Field

[0001] The present invention relates to the technical field of battery state of health prediction, and specifically to a method and system for predicting the state of health of energy storage power station batteries based on data-driven. Background Technique

[0002] As the core component of an energy storage power station, the accurate prediction of the state of health of lithium-ion batteries is of great significance for ensuring the safe operation of the energy storage system and extending its service life. At present, the methods for predicting the state of health of lithium batteries mainly include two categories: mechanism model-based and data-driven. The mechanism model method mainly constructs a battery performance degradation model based on electrochemical principles. Although it has good physical interpretability, it is difficult to establish an accurate mathematical model due to the complex internal reaction mechanism of the battery. The data-driven method directly learns the battery performance degradation law from the monitoring data through machine learning algorithms, and shows stronger adaptability in practical applications. In recent years, with the development of deep learning technology, prediction methods based on neural networks have been widely studied. For example, Chen et al. proposed a prediction method based on Transformer, Shi Yongsheng et al. adopted an estimation method that improves bidirectional GRU with attention, and Li et al. proposed a hybrid model based on CNN-LSTM, etc.

[0003] However, the existing data-driven methods have the following problems when dealing with the battery capacity change characteristics: First, the battery capacity data often contains multi-scale characteristics, such as long-term decay trends and short-term fluctuation characteristics. Most of the existing methods use a single model structure and it is difficult to effectively capture these different-scale change characteristics at the same time. Second, in the process of optimizing model parameters, the optimization strategy of a fixed parameter interval is usually adopted, without considering the influence of different types of signal characteristics on the model performance, resulting in low parameter optimization efficiency. Third, in the process of integrating prediction results, most of them use a simple weighted average method, lacking a recognition and processing mechanism for special phenomena such as capacity regeneration, which affects the accuracy and reliability of the prediction results. In particular, when the battery shows the phenomenon of capacity regeneration, the prediction results are prone to sudden changes that do not conform to physical laws, affecting the reliability of the prediction model in practical applications. Summary of the Invention

[0004] In view of the above existing problems, the present invention provides a method and system for predicting the state of health of energy storage power station batteries based on data-driven, which can solve the problems mentioned in the background technique.

[0005] To solve the above technical problems, the technical solution provided by the present invention is as follows: A data-driven method for predicting the health state of energy storage power station batteries, including: obtaining the capacity data sequence of lithium batteries in the energy storage power station, and performing variational mode decomposition on the capacity data sequence to obtain at least two component signals; the at least two component signals include a main degradation trend signal and at least one high-frequency signal; dividing the at least two component signals into training data and test data according to a preset ratio respectively, training a first prediction model using the training data of the main degradation trend signal, and training a second prediction model using the training data of the at least one high-frequency signal; superimposing the prediction result of the first prediction model for the main degradation trend signal and the prediction result of the second prediction model for the high-frequency signal to obtain the predicted value of the health state of the lithium battery.

[0006] As a preferred embodiment of the data-driven method for predicting the health state of energy storage power station batteries of the present invention, it includes: obtaining the capacity data sequence of lithium batteries in the energy storage power station, and performing variational mode decomposition on the capacity data sequence to obtain at least two component signals, including the following steps: obtaining the capacity data sequence of lithium batteries in the energy storage power station during the charging and discharging process based on the sampling time interval; calculating the central frequency distribution of the capacity data sequence, and determining the number K of decomposition of the variational mode decomposition as the number of central frequencies with the smallest frequency difference; using the number K of decomposition to perform variational mode decomposition on the capacity data sequence to obtain a main degradation trend signal with the lowest central frequency and K-1 high-frequency signals.

[0007] As a preferred embodiment of the data-driven method for predicting the health state of energy storage power station batteries of the present invention, wherein: the first prediction model is a Transformer neural network prediction model; the second prediction model is a gated recurrent unit GRU neural network prediction model; the hyperparameters of the Transformer neural network prediction model and the gated recurrent unit GRU neural network prediction model are optimized by a particle swarm optimization algorithm.

[0008] As a preferred solution of the data-driven energy storage power station battery health state prediction method described in the present invention, wherein: obtaining the prediction result of the first prediction model includes the following steps. Based on the amplitude and change frequency of the main degradation trend signal, a fitness function including the influence weight of model hyperparameters on the signal fitting accuracy is established. The fitness function is used to guide the particle swarm optimization algorithm to optimize the number of network layers, the number of neurons, the learning rate, the number of iterations, the batch size, and the number of heads of the Transformer neural network prediction model. The parameters of the Transformer neural network prediction model are trained based on the training data of the main degradation trend signal. The Transformer neural network prediction model includes an encoder and a linear layer. The trained Transformer neural network prediction model is tested and verified using the test data of the main degradation trend signal, and a capacity attenuation sequence corresponding to the main degradation trend signal is generated as the prediction result of the first prediction model.

[0009] As a preferred solution of the data-driven energy storage power station battery health state prediction method described in the present invention, wherein: obtaining the prediction result of the second prediction model includes the following steps. Based on the fluctuation period and change law of the high-frequency signal, a fitness function including the influence weight of model hyperparameters on the time series feature extraction ability is constructed. The fitness function is used to guide the particle swarm optimization algorithm to optimize the number of network layers, the number of neurons, the learning rate, the number of iterations, and the batch size of the gated recurrent unit (GRU) neural network prediction model. The parameters of the gated recurrent unit (GRU) neural network prediction model are trained based on the training data of the high-frequency signal. The gated recurrent unit (GRU) neural network prediction model includes an update gate and a reset gate. The trained gated recurrent unit (GRU) neural network prediction model is tested and verified using the test data of the high-frequency signal, and a capacity fluctuation sequence corresponding to the high-frequency signal is generated as the prediction result of the second prediction model.

[0010] As a preferred embodiment of the data-driven battery health state prediction method for an energy storage power station according to the present invention, the method includes: superimposing the prediction result of the first prediction model on the main degradation trend signal and the prediction result of the second prediction model on the high-frequency signal to obtain the predicted value of the lithium battery health state, which includes the following steps: calculating the root mean square error (RMSE) of the prediction result of the main degradation trend signal and the prediction result of the high-frequency signal at different time points, and establishing an adaptive weight function based on the RMSE; when the RMSE is greater than a preset threshold, increasing the weight coefficient of the prediction result of the high-frequency signal and decreasing the weight coefficient of the prediction result of the main degradation trend signal; if the preset threshold is triggered at N consecutive time points, the prediction results of the high-frequency signal at these N time points are corrected according to the following rules: when the fluctuation amplitude of the prediction result of the high-frequency signal exceeds the first preset capacity threshold and is less than the second preset capacity threshold, and the fluctuation duration is greater than the first preset duration and less than the second preset duration, it is determined as a capacity regeneration phenomenon; when the fluctuation amplitude of the prediction result of the high-frequency signal exceeds the second preset capacity threshold or the fluctuation duration exceeds the second preset duration, it is determined as an abnormal fluctuation; when the fluctuation amplitude of the prediction result of the high-frequency signal is less than the first preset capacity threshold or the fluctuation duration is less than the first preset duration, it is determined as a normal fluctuation; for the time points determined as the capacity regeneration phenomenon, keeping the prediction result of the high-frequency signal unchanged; for the time points determined as the abnormal fluctuation, correcting the prediction result of the high-frequency signal to the average value of the prediction results of the adjacent K normal fluctuation time points; for the time points determined as the normal fluctuation, keeping the prediction result of the high-frequency signal unchanged.

[0011] As a preferred embodiment of the data-driven battery health state prediction method for an energy storage power station according to the present invention, the method includes: superimposing the corrected prediction result of the high-frequency signal and the prediction result of the main degradation trend signal according to the adaptive weight function to obtain the predicted value of the lithium battery health state: constructing a dynamic weighting coefficient based on time series attenuation; the dynamic weighting coefficient is negatively correlated with the occurrence frequency of the capacity regeneration phenomenon in the corrected prediction result of the high-frequency signal and positively correlated with the monotonicity of the main degradation trend signal; normalizing the dynamic weighting coefficient and the adaptive weight function to obtain the final weighting coefficient; where, when the capacity regeneration phenomenon occurs, the final weighting coefficient adopts a linear decreasing function for smooth transition during the duration of the capacity regeneration phenomenon; using the final weighting coefficient to perform weighted superposition of the corrected prediction result of the high-frequency signal and the prediction result of the main degradation trend signal, and when the superposition result is greater than the predicted value of the lithium battery health state at the previous moment, correcting the predicted value of the lithium battery health state at this moment to the weighted average value of the predicted value at the previous moment and the superposition result.

[0012] To further solve the above technical problems, the present invention provides the following technical solutions: A data-driven battery health state prediction system for an energy storage power station, comprising: a data decomposition module, configured to obtain the capacity data sequence of the lithium battery in the energy storage power station, and perform variational mode decomposition on the capacity data sequence to obtain at least two component signals; a model training module, configured to divide the at least two component signals into training data and test data according to a preset ratio respectively, train a first prediction model using the training data of the main degradation trend signal, and train a second prediction model using the training data of the at least one high-frequency signal; a health state prediction module, configured to superimpose the prediction result of the first prediction model on the main degradation trend signal and the prediction result of the second prediction model on the high-frequency signal to obtain the predicted value of the lithium battery health state.

[0013] A computer device, comprising a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned data-driven battery health state prediction method for an energy storage power station are implemented.

[0014] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned data-driven battery health state prediction method for an energy storage power station are implemented.

[0015] The beneficial effects of the present invention: Through systematic data decomposition and a novel model training strategy, the present invention realizes high-precision prediction of the health state of lithium batteries in energy storage power stations. First, based on the intelligent decomposition strategy of the VMD algorithm, the optimal decomposition number is determined through the analysis of the central frequency distribution, effectively separating the capacity data sequence into two types of characteristic signals: the main degradation trend and high-frequency fluctuations; secondly, different parameter optimization and model training methods are designed for different characteristic signals. Among them, the fitness function based on signal characteristics guides the PSO algorithm for hyperparameter optimization, avoiding the problem of blind search in the fixed parameter interval of traditional methods. At the same time, a simplified Transformer structure and a GRU network are used to process the main trend and high-frequency signals respectively, giving full play to the advantages of different models; finally, a prediction result superposition mechanism based on double-threshold discrimination and dynamic weighting is proposed. Through the adaptive weight function and the linear decreasing smooth transition strategy, it not only ensures the sensitive capture of transient characteristics such as capacity regeneration, but also ensures the continuity and rationality of the health state prediction, effectively avoiding the problem of sudden change in the predicted value. The present invention shows stronger adaptability and stability when dealing with complex working conditions containing capacity regeneration phenomena, so it has obvious advantages over the prior art in terms of prediction accuracy and generalization performance. Description of the Drawings

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0017] Figure 1 It is a schematic diagram of the overall process of the method for predicting the battery health state of an energy storage power station based on data driving proposed by the present invention; Figure 2 It is a diagram of the VMD decomposition result in the method for predicting the battery health state of an energy storage power station based on data driving proposed by the present invention; Figure 3 It is a schematic diagram of the overall structure in the system for predicting the battery health state of an energy storage power station based on data driving proposed by the present invention.

[0018] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0020] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, providing a method for predicting the battery health state of an energy storage power station based on data driving.

[0021] This application can effectively solve the above-mentioned problems. Next, multiple embodiments will be combined to elaborate in detail how to implement the method for predicting the battery health state of an energy storage power station based on data driving.

[0022] As Figure 1 shown, the method for predicting the battery health state of an energy storage power station based on data driving includes the following steps: S1: Obtain the capacity data sequence of the lithium battery in the energy storage power station, and perform variational mode decomposition on the capacity data sequence to obtain at least two component signals.

[0023] In an optional embodiment, the at least two component signals include a main degradation trend signal and at least one high-frequency signal.

[0024] S1.1: Obtain the capacity data sequence of lithium batteries in the energy storage power station during the charging and discharging processes at intervals based on the sampling time.

[0025] Specifically, during the operation of the energy storage power station, it is necessary to monitor the capacity change of lithium batteries in real time. Data collection of the batteries is carried out based on a fixed sampling period to obtain the capacity data sequence reflecting the charging and discharging processes of the batteries. The selection of the sampling time interval needs to ensure that the change characteristics of the battery capacity can be captured, especially the capacity regeneration phenomenon. These data sequences contain the main trend changes and local fluctuation information during the battery aging process, providing a data basis for subsequent decomposition.

[0026] S1.2: Calculate the central frequency distribution of the capacity data sequence, and determine the number of decomposition K of the variational mode decomposition as the number of central frequencies with the smallest frequency difference.

[0027] Specifically, determining the number of modes K of the variational mode decomposition (VMD) is a key step. Taking battery B5 as an example, when K = 6, the central frequency after decomposition has a very small difference from that when K = 5, indicating that the variational mode decomposition (VMD) over-decomposes the signal; when K = 4, since the central frequency difference between IMF4 and IMF5 is relatively large, the signal decomposition is incomplete. By observing the central frequency distribution of IMF under different K values (as shown in Table 1), select the K value with the smallest central frequency difference between adjacent IMFs, which can not only avoid over-decomposition but also ensure sufficient decomposition, so as to obtain the optimal number of decompositions.

[0028] Table 1 Central frequencies corresponding to different K values

[0029] S1.3: Use the number of decompositions K to perform variational mode decomposition on the capacity data sequence to obtain a main degradation trend signal with the lowest central frequency and K - 1 high-frequency signals.

[0030] Specifically, the variational mode decomposition (VMD) algorithm is used to decompose the capacity data sequence. This algorithm is based on the variational problem framework and the ADMM algorithm. Among the K intrinsic mode functions (IMFs) obtained by decomposition, the IMF with the lowest central frequency corresponds to the main degradation trend signal, reflecting the overall trend of battery capacity attenuation; the remaining K - 1 IMFs are high-frequency signals, containing local fluctuation information such as capacity regeneration. As Figure 2 , the frequency characteristics of each IMF after decomposition can be clearly seen. Among them, IMF5 shows obvious low-frequency characteristics, corresponding to the main attenuation trend of the battery capacity.

[0031] Preferably, in step S1, through the systematic collection and scientific decomposition of the lithium battery capacity data of the energy storage power station, the effective separation of the capacity degradation characteristics is achieved. First, based on a fixed sampling period, a battery capacity data sequence is obtained to ensure the continuity and integrity of the data; secondly, by analyzing the central frequency distribution characteristics under different numbers of modes, the decomposition number K of the VMD algorithm is optimized to avoid over-decomposition and under-decomposition of the signal; finally, based on the determined decomposition number K, the VMD algorithm is used to decompose the capacity data sequence into a low-frequency main degradation trend signal reflecting the overall attenuation trend and high-frequency signals characterizing local fluctuation characteristics. This decomposition method fully considers the multi-scale characteristics of the battery capacity change, laying a data foundation for the targeted training of different types of neural networks in the follow-up, and is the key preprocessing step to achieve accurate prediction of the battery health state.

[0032] S2: Divide at least two component signals into training data and test data according to a preset ratio respectively. Use the training data of the main degradation trend signal to train a first prediction model, and use the training data of at least one high-frequency signal to train a second prediction model.

[0033] Among them, the first prediction model is a Transformer neural network prediction model, and the second prediction model is a gated recurrent unit GRU neural network prediction model. The hyperparameters of the Transformer neural network prediction model and the gated recurrent unit GRU neural network prediction model are optimized by the particle swarm optimization algorithm.

[0034] S2.1: Divide at least two component signals into training data and test data according to a preset ratio respectively.

[0035] Specifically, divide each component signal obtained by VMD decomposition. Use the first 60% of the samples in the data set as training data, and the last 40% of the samples as test data. This preset ratio division method of 6:4 is a commonly used data set division method in the field of machine learning, which can ensure the sufficiency of training data while reserving enough test samples for model performance evaluation. Taking battery B5 as an example, perform this ratio division on the main degradation trend signal IMF5 and each high-frequency signal IMF1-IMF4 obtained by its VMD decomposition. Such a division method ensures that each component signal has an independent training set and test set, avoiding data leakage, and at the same time maintaining the continuity and integrity of the signal characteristics. The training data is used for parameter learning and optimization of the model, and the test data is used to evaluate the generalization performance and prediction accuracy of the model. This data division strategy provides a standardized data foundation for the training and verification of different neural network models in the follow-up.

[0036] S2.2: Use the training data of the main degradation trend signal to train the first prediction model, where the first prediction model is a Transformer neural network prediction model, and the hyperparameters of the Transformer neural network prediction model are optimized by the particle swarm optimization algorithm.

[0037] First, based on the amplitude and change frequency of the main degradation trend signal, establish a fitness function that includes the influence weights of model hyperparameters on signal fitting accuracy. Use the fitness function to guide the particle swarm optimization algorithm to optimize the number of network layers, the number of neurons, the learning rate, the number of iterations, the batch size, and the number of heads of the Transformer neural network prediction model.

[0038] Then, train the parameters of the Transformer neural network prediction model based on the training data of the main degradation trend signal; among them, the Transformer neural network prediction model includes an encoder and a linear layer.

[0039] Finally, use the test data of the main degradation trend signal to test and verify the trained Transformer neural network prediction model, and generate the capacity attenuation sequence corresponding to the main degradation trend signal as the prediction result of the first prediction model.

[0040] Preferably, in this embodiment, the training process of the Transformer model for the main degradation trend signal includes three key steps. First, optimize the PSO algorithm by constructing a new fitness function. The traditional PSO only searches for parameters within a fixed interval (such as the number of network layers [1, 10], the learning rate [0.00001, 0.1], etc.), while the improved fitness function fully considers the amplitude and change frequency characteristics of the main degradation trend signal, assigns different influence weights to different hyperparameters, and makes the parameter optimization process more targeted. Second, based on the Transformer network structure, use a simplified encoder-linear layer structure to replace the encoder-decoder structure of the standard Transformer. This simplification not only retains the advantage of the multi-head self-attention mechanism to capture the internal correlation of data but also avoids the cumulative error that may occur in the decoder structure on small-scale data sets. Finally, through test data verification, output the capacity attenuation sequence, which reflects the overall degradation trend of the battery capacity over time and provides the main basis for subsequent SOH prediction.

[0041] S2.3: Use the training data of at least one high-frequency signal to train the second prediction model, where the second prediction model is a gated recurrent unit (GRU) neural network prediction model, and the hyperparameters of the gated recurrent unit (GRU) neural network prediction model are optimized by the particle swarm optimization algorithm.

[0042] First, based on the fluctuation period and variation law of the high-frequency signal, a fitness function is constructed that includes the influence weights of model hyperparameters on the ability to extract time series features. The fitness function is used to guide the particle swarm optimization algorithm to optimize the number of network layers, the number of neurons, the learning rate, the number of iterations, and the batch size of the gated recurrent unit (GRU) neural network prediction model.

[0043] Then, based on the training data of the high-frequency signal, parameter training is performed on the gated recurrent unit (GRU) neural network prediction model, and the gated recurrent unit (GRU) neural network prediction model includes an update gate and a reset gate.

[0044] Finally, the trained gated recurrent unit (GRU) neural network prediction model is tested and verified using the test data of the high-frequency signal, and a capacity fluctuation sequence corresponding to the high-frequency signal is generated as the prediction result of the second prediction model.

[0045] It should be noted that the training of the GRU model for high-frequency signals in this embodiment also includes three core links. First, parameter optimization is performed based on time series features. By analyzing the fluctuation period and variation law of high-frequency signals, a new fitness function is constructed. This fitness function particularly emphasizes the model's ability to extract time series features, which is essentially different from traditional parameter optimization methods. The GRU network processes sequence information through an update gate and a reset gate, so the fitness function focuses on the influence of hyperparameters on the performance of these two gating mechanisms. Second, the advantages of the GRU network structure are utilized. Compared with the LSTM, the GRU structure is simpler and performs better on small data sets, which exactly meets the modeling requirements of local fluctuation features such as capacity regeneration contained in high-frequency signals. Finally, through test data verification, a capacity fluctuation sequence is output, which accurately captures the local fluctuation features of the battery capacity, especially the capacity regeneration phenomenon, providing important supplementary information for the final SOH prediction.

[0046] Preferably, although the data partitioning strategy in S2.1 is a conventional operation, it lays a standardized foundation for subsequent differential modeling. For the main degradation trend signal, the fitness function based on signal characteristics proposed in S2.2 makes the parameter optimization process more targeted, avoiding the problem of blind search in a fixed parameter interval in the traditional PSO algorithm and improving the parameter optimization efficiency. At the same time, the simplified Transformer structure is adopted, which not only retains the advantages of the multi-head self-attention mechanism but also avoids the cumulative error problem on small-scale datasets. For high-frequency signals, the parameter optimization method based on temporal characteristics designed in S2.3 particularly emphasizes the model's ability to extract fluctuation characteristics, enabling the GRU network to more accurately capture local fluctuation phenomena such as capacity regeneration, which is difficult to achieve by traditional single models. The capacity decay sequence and capacity fluctuation sequence output by the two models respectively reflect different dimensional characteristics of battery performance degradation, providing a more comprehensive information basis for subsequent SOH prediction. Overall, through differential parameter optimization strategies and model training methods, the present invention realizes multi-dimensional accurate modeling of battery capacity change characteristics, and has obvious advantages in prediction accuracy and generalization performance compared with the single model or simple combination scheme in the prior art, especially showing stronger adaptability when dealing with complex working conditions containing capacity regeneration phenomena.

[0047] S3: Superimpose the prediction result of the first prediction model for the main degradation trend signal and the prediction result of the second prediction model for the high-frequency signal to obtain the predicted value of the lithium battery health state.

[0048] S3.1: Calculate the root mean square error RMSE of the prediction result of the main degradation trend signal and the prediction result of the high-frequency signal at different time points, and establish an adaptive weight function based on the root mean square error RMSE. When the root mean square error RMSE is greater than a preset threshold, increase the weight coefficient of the prediction result of the high-frequency signal and decrease the weight coefficient of the prediction result of the main degradation trend signal.

[0049] It should be noted that in this embodiment, the adaptive weight function adopts a piecewise continuous function structure based on the root mean square error (RMSE), specifically as follows: when the RMSE is less than the first preset threshold, the weight coefficient of the main degradation trend signal adopts a fixed value of 0.8, and the weight coefficient of the high-frequency signal adopts a fixed value of 0.2; when the RMSE is greater than the first preset threshold and less than the second preset threshold, the weight coefficient of the main degradation trend signal decreases linearly with the RMSE, specifically 0.8 - (RMSE - the first preset threshold) × 0.4 / (the second preset threshold - the first preset threshold), and the weight coefficient of the high-frequency signal increases linearly accordingly to ensure that the sum of the weights is 1; when the RMSE is greater than the second preset threshold, the weight coefficient of the main degradation trend signal adopts a fixed value of 0.4, and the weight coefficient of the high-frequency signal adopts a fixed value of 0.6; wherein, the first preset threshold and the second preset threshold are determined according to the prediction error distribution of the main degradation trend signal and the high-frequency signal in the training dataset. The first preset threshold takes the mean value of the RMSE of the training dataset, and the second preset threshold takes the mean value of the RMSE of the training dataset plus twice the standard deviation. Through the piecewise design of the RMSE threshold, this adaptive weight function can, while maintaining the dominant role of the main degradation trend signal, increase the weight of the high-frequency signal in a timely manner when the prediction error is large, so as to achieve a rapid response and accurate characterization of the battery capacity fluctuation.

[0050] S3.2: If the preset threshold is triggered within N consecutive time points, then the prediction results of the high-frequency signals at these N time points are corrected according to the following rules: When the fluctuation amplitude of the high-frequency signal prediction result exceeds the first preset capacity threshold and is less than the second preset capacity threshold, and the fluctuation duration is greater than the first preset duration and less than the second preset duration, it is determined as a capacity regeneration phenomenon; When the fluctuation amplitude of the high-frequency signal prediction result exceeds the second preset capacity threshold or the fluctuation duration exceeds the second preset duration, it is determined as an abnormal fluctuation; When the fluctuation amplitude of the high-frequency signal prediction result is less than the first preset capacity threshold or the fluctuation duration is less than the first preset duration, it is determined as a normal fluctuation; For the time points determined as capacity regeneration phenomena, keep their high-frequency signal prediction results unchanged; For the time points determined as abnormal fluctuations, correct their high-frequency signal prediction results to the average value of the prediction results of the adjacent K normal fluctuation time points; For the time points determined as normal fluctuations, keep their high-frequency signal prediction results unchanged.

[0051] S3.3: Superimpose the corrected high-frequency signal prediction results and the prediction results of the main degradation trend signal according to the adaptive weight function to obtain the predicted value of the lithium battery health state.

[0052] First, construct a dynamic weighting coefficient based on time series attenuation; the dynamic weighting coefficient is negatively correlated with the occurrence frequency of the capacity regeneration phenomenon in the predicted result of the corrected high-frequency signal, and positively correlated with the monotonicity of the main degradation trend signal; Then, normalize the dynamic weighting coefficient and the adaptive weight function to obtain the final weighting coefficient; among them, when the capacity regeneration phenomenon occurs, the final weighting coefficient adopts a linear decreasing function for smooth transition during the duration of the capacity regeneration phenomenon; Finally, use the final weighting coefficient to perform weighted superposition on the predicted result of the corrected high-frequency signal and the predicted result of the main degradation trend signal. When the superposition result is greater than the predicted value of the health state of the lithium battery at the previous moment, the predicted value of the health state of the lithium battery at this moment is corrected to the weighted average of the predicted value at the previous moment and the superposition result.

[0053] It should be noted that under the dual constraints of the dynamic weighting coefficient based on time series attenuation and the adaptive weight function, the weighted superposition process is designed according to the physical characteristics of battery capacity decay. Among them, the calculation of the dynamic weighting coefficient uses the exponential decay function exp(-λ×n / N), where λ is the decay coefficient, n is the number of times of the capacity regeneration phenomenon observed currently, and N is the average occurrence number of the capacity regeneration phenomenon in historical data; the smooth transition of the final weighting coefficient uses the linear decreasing function of t / T, where t is the current moment within the duration of the capacity regeneration phenomenon, and T is the total duration of the capacity regeneration phenomenon; the correction of the health state predicted value uses a fixed weight ratio of 0.7 and 0.3, that is, when the superposition result exceeds the predicted value at the previous moment, the predicted value at the current moment is set to 0.7 times the predicted value at the previous moment plus 0.3 times the current superposition result. This ratio is determined based on the physical constraint that the health state of the battery will not be significantly improved in a short time; through the above weighted superposition mechanism, it not only ensures the sensitivity of the predicted result to transient characteristics such as capacity regeneration, but also ensures the continuity and rationality of the health state prediction, effectively avoiding the problem of sudden change of the predicted value that is prone to occur in traditional simple superposition methods.

[0054] In summary, through systematic data decomposition and personalized model training strategies, the present invention achieves high-precision prediction of the health state of lithium batteries in energy storage power stations. First, based on the intelligent decomposition strategy of the VMD algorithm, the optimal decomposition number is determined through the analysis of the central frequency distribution, effectively separating the capacity data sequence into two types of characteristic signals: the main degradation trend and high-frequency fluctuations. Secondly, different parameter optimization and model training methods are designed for different characteristic signals. Among them, the fitness function based on signal characteristics guides the PSO algorithm to optimize hyperparameters, avoiding the problem of blind search in the fixed parameter interval of traditional methods. At the same time, the simplified Transformer structure and GRU network are used to process the main trend and high-frequency signals respectively, giving full play to the advantages of different models. Finally, a prediction result superposition mechanism based on double-threshold discrimination and dynamic weighting is proposed. Through the adaptive weight function and linear decreasing smooth transition strategy, it not only ensures the sensitive capture of transient characteristics such as capacity regeneration, but also ensures the continuity and rationality of the health state prediction, effectively avoiding the problem of sudden change in the predicted value. The present invention shows stronger adaptability and stability when dealing with complex working conditions containing capacity regeneration phenomena, thus having obvious advantages in prediction accuracy and generalization performance compared with the prior art.

[0055] Example 2, referring to Figure 3 , which is an embodiment of the present invention, provides a data-driven health state prediction system for energy storage power station batteries. The system mainly consists of a data decomposition module, a model training module, and a health state prediction module.

[0056] Among them, the data decomposition module obtains the capacity data sequence of the lithium battery in the energy storage power station and performs variational mode decomposition on the capacity data sequence to obtain at least two component signals.

[0057] The model training module divides at least two component signals into training data and test data according to a preset ratio respectively, and trains a first prediction model with the training data of the main degradation trend signal and trains a second prediction model with the training data of at least one high-frequency signal.

[0058] The health state prediction module superimposes the prediction result of the first prediction model for the main degradation trend signal and the prediction result of the second prediction model for the high-frequency signal to obtain the predicted value of the health state of the lithium battery.

[0059] Such as Figure 3As shown, the data decomposition module of the data-driven energy storage power station battery state of health prediction system is responsible for obtaining the capacity data sequence of lithium batteries in the energy storage power station and performing variational mode decomposition on it to obtain at least two component signals. This module first obtains the capacity data sequence of lithium batteries during charge and discharge from the energy storage power station based on a preset sampling time interval, then calculates the central frequency distribution of the obtained data sequence, and determines the optimal number of variational mode decomposition K by finding the number of central frequencies with the smallest frequency difference. Finally, the capacity data sequence is decomposed by variational mode decomposition using the determined number of decompositions K, and a main degradation trend signal with the lowest central frequency and K - 1 high-frequency signals are output, providing a data basis for subsequent model training.

[0060] The model training module receives the signals output by the data decomposition module and divides these component signals into training data and test data according to a preset ratio (usually 6:4). For the main degradation trend signal, this module uses a Transformer neural network structure to construct a first prediction model and optimizes the hyperparameters of the network through a particle swarm optimization algorithm, including the number of network layers, the number of neurons, the learning rate, the number of iterations, the batch size, and the number of heads, etc. For the high-frequency signals, this module uses a gated recurrent unit (GRU) neural network structure to construct a second prediction model and also optimizes the network hyperparameters through a particle swarm optimization algorithm. The two prediction models are respectively trained with the training data of the corresponding signals and verified with the test data for the generalization ability of the models to ensure that the models can accurately capture the long-term decay trend and short-term fluctuation characteristics of the battery capacity.

[0061] The state of health prediction module is responsible for fusing the prediction results of the first prediction model for the main degradation trend signal and the prediction results of the second prediction model for the high-frequency signals to obtain the final predicted value of the lithium battery state of health. This module first calculates the root mean square error (RMSE) of the two signal prediction results at different time points and constructs an adaptive weight function based on these errors. When the RMSE exceeds a preset threshold, the system increases the weight of the prediction result of the high-frequency signal to better capture the capacity fluctuation. At the same time, this module uses a double-threshold discrimination mechanism to identify the capacity regeneration phenomenon, abnormal fluctuations, and normal fluctuations in the high-frequency signals and corrects the signals accordingly. Finally, this module constructs a dynamic weighting coefficient based on time-series attenuation, superimposes the corrected prediction result of the high-frequency signal and the prediction result of the main degradation trend signal after weighting, and makes necessary adjustments to the superimposed result according to the physical laws of the battery state of health to ensure the rationality and continuity of the prediction output.

[0062] The data-driven battery state of health prediction system for energy storage power stations realizes high-precision prediction of the state of health of lithium batteries in energy storage power stations through the intelligent decomposition strategy of the data decomposition module, the differential modeling method of the model training module, and the adaptive fusion mechanism of the state of health prediction module. The data flow between each module is clear, and they work together to form a complete closed-loop system. In practical applications, the data-driven battery state of health prediction system for energy storage power stations can be deployed in the battery management system (BMS) of energy storage power stations, receive battery operation data in real time and output the state of health prediction results, provide decision-making support for the operation and maintenance of power stations, and provide effective technical guarantees for extending the service life of batteries in energy storage power stations and improving system safety.

[0063] Embodiment 3 is an embodiment of the present invention. The difference from the previous embodiment is that if the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0064] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0065] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other suitable processing as necessary, and then stored in a computer memory.

[0066] It should be understood that the various parts of the present invention can be implemented using hardware, software, firmware, or combinations thereof. In the above embodiments, the multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0067] Application example. To verify the beneficial effects of the present invention, a data-driven method for predicting the state of health of energy storage power station batteries is scientifically demonstrated through economic benefit calculation and simulation experiments.

[0068] To verify the effectiveness of the present invention, 10 groups of ternary lithium battery packs in a certain energy storage power station were selected as the experimental objects. The rated capacity of the battery packs is 100 Ah, and the rated voltage is 3.7 V. The experiment was carried out in a constant temperature environment (25±2 °C), and a Neware BTS-5V300A battery test system was used to collect data. The experimental process is as follows: First, the sampling period was set to 1 hour, and continuous monitoring was carried out for 400 days to obtain the charge and discharge capacity data sequence of the battery packs. By analyzing the central frequency distribution at different K values (K = 3 to 7), it was found that when K = 5, the frequency difference between adjacent IMFs was the smallest and the distribution was the most uniform. Therefore, the number of VMD decompositions was determined to be 5. For the obtained main degradation trend signal and 4 high-frequency signals, the training set and the test set were divided according to a ratio of 6:4. For the main degradation trend signal, a PSO fitness function based on signal amplitude and frequency characteristics was constructed, where the amplitude influence weight was 0.6 and the frequency influence weight was 0.4. After optimization, the optimal parameter configuration of the Transformer model was obtained: the number of network layers was 4, the number of hidden layer neurons was 256, the learning rate was 0.001, the number of iterations was 1000, the batch size was 64, and the number of attention heads was 8. For the high-frequency signals, a PSO fitness function based on the fluctuation period was designed, and the GRU model parameters were optimized: the number of network layers was 3, the number of hidden layer neurons was 128, the learning rate was 0.0005, the number of iterations was 800, and the batch size was 32. In the prediction result superposition stage, the first RMSE threshold was set to 0.015, the second threshold was set to 0.035, the first preset capacity threshold for capacity fluctuation discrimination was set to 0.02, the second preset capacity threshold was set to 0.05, the first preset duration was set to 24 hours, the second preset duration was set to 72 hours. The decay factor λ in the dynamic weighting coefficient was set to 0.8, and the number of adjacent normal points K was taken as 5.

[0069] Table 1 Experimental result record

[0070] Through the analysis of the experimental data provided in Table 1, the following conclusions can be drawn: First, in terms of prediction accuracy, the average RMSE of the prediction of the main degradation trend by the present invention is 0.0129, which is significantly better than the traditional single-model method (the typical RMSE is above 0.02). This benefits from the PSO fitness function optimization strategy based on signal characteristics, enabling the Transformer model to more accurately capture the long-term trend of capacity degradation. Second, in terms of high-frequency feature prediction, although the RMSE is slightly higher than that of the main trend prediction (average 0.0188), through the time-series feature extraction ability of the GRU network and combined with the dual-threshold discrimination mechanism, the high-precision identification of the capacity regeneration phenomenon is achieved, and the average identification accuracy reaches 94.98%, far exceeding the 85% level of the traditional method. It is particularly worth noting that, thanks to the dynamic weighting and smooth transition strategy, the final prediction error of the health state is controlled within 2%, and there are no significant numerical mutations during the prediction process, reflecting the superiority of the present invention in dealing with complex working conditions. In terms of computational efficiency, the average time consumption for a single prediction is 43.5 ms, meeting the requirements of real-time monitoring. By comparing the test results of different battery packs, it is found that this method has good stability and reliability, and the dispersion of each performance index is small, indicating that this method has wide applicability. Compared with the prior art, the present invention has achieved significant improvements in prediction accuracy, capacity regeneration identification ability, and computational efficiency, providing a more effective solution for the accurate prediction of the health state of energy storage power station batteries.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A data-driven method for predicting the state of health of energy storage power station batteries, characterized in that, Including: Obtain the capacity data sequence of the lithium battery in the energy storage power station, and perform variational mode decomposition on the capacity data sequence to obtain at least two component signals; The at least two component signals include a main degradation trend signal and at least one high-frequency signal; Divide the at least two component signals into training data and test data according to a preset ratio respectively. Train a first prediction model using the training data of the main degradation trend signal, and train a second prediction model using the training data of the at least one high-frequency signal; Superimpose the prediction result of the first prediction model for the main degradation trend signal and the prediction result of the second prediction model for the high-frequency signal to obtain the predicted value of the health state of the lithium battery.

2. The data-driven battery state of health prediction method for an energy storage power station according to claim 1, wherein: Obtain the capacity data sequence of the lithium battery in the energy storage power station, and perform variational mode decomposition on the capacity data sequence to obtain at least two component signals, including the following steps: Based on the sampling time interval, obtain the capacity data sequence of the lithium battery in the energy storage power station during the charge and discharge process; Calculate the central frequency distribution of the capacity data sequence, and determine the number of central frequencies with the smallest frequency difference as the decomposition number K of the variational mode decomposition; Use the decomposition number K to perform variational mode decomposition on the capacity data sequence to obtain a main degradation trend signal with the lowest central frequency and K - 1 high-frequency signals.

3. The data-driven method for predicting the battery health state of an energy storage power station according to claim 1, characterized in that: The first prediction model is a Transformer neural network prediction model; The second prediction model is a gated recurrent unit GRU neural network prediction model; The hyperparameters of the Transformer neural network prediction model and the gated recurrent unit GRU neural network prediction model are optimized by a particle swarm optimization algorithm.

4. The data-driven battery state of health prediction method for an energy storage power station according to claim 3, wherein: Obtaining the prediction result of the first prediction model includes the following steps: Based on the amplitude and change frequency of the main degradation trend signal, establish a fitness function including the influence weight of model hyperparameters on the signal fitting accuracy. Use the fitness function to guide the particle swarm optimization algorithm to optimize the number of network layers, the number of neurons, the learning rate, the number of iterations, the batch size, and the number of heads of the Transformer neural network prediction model; Perform parameter training on the Transformer neural network prediction model based on the training data of the main degradation trend signal; wherein, the Transformer neural network prediction model includes an encoder and a linear layer; Use the test data of the main degradation trend signal to test and verify the trained Transformer neural network prediction model, and generate the capacity attenuation sequence corresponding to the main degradation trend signal as the prediction result of the first prediction model.

5. The data-driven method for predicting the battery health state of an energy storage power station according to claim 1, wherein: Obtaining the prediction result of the second prediction model includes the following steps: Based on the fluctuation period and change law of the high-frequency signal, construct a fitness function including the influence weight of model hyperparameters on the time series feature extraction ability. Use the fitness function to guide the particle swarm optimization algorithm to optimize the number of network layers, the number of neurons, the learning rate, the number of iterations, and the batch size of the gated recurrent unit GRU neural network prediction model; Train the parameters of the gated recurrent unit (GRU) neural network prediction model based on the training data of the high-frequency signal. The gated recurrent unit (GRU) neural network prediction model includes an update gate and a reset gate. Use the test data of the high-frequency signal to test and verify the trained gated recurrent unit (GRU) neural network prediction model, and generate the capacity fluctuation sequence corresponding to the high-frequency signal as the prediction result of the second prediction model.

6. The data-driven battery state of health prediction method for an energy storage power station according to claim 1, wherein: Superimpose the prediction result of the first prediction model on the main degradation trend signal and the prediction result of the second prediction model on the high-frequency signal to obtain the predicted value of the lithium battery health state, including the following steps: Calculate the root mean square error (RMSE) of the prediction results of the main degradation trend signal and the high-frequency signal at different time points, and establish an adaptive weight function based on the root mean square error (RMSE). When the root mean square error (RMSE) is greater than the preset threshold, increase the weight coefficient of the prediction result of the high-frequency signal and decrease the weight coefficient of the prediction result of the main degradation trend signal. If the preset threshold is triggered at N consecutive time points, correct the prediction results of the high-frequency signal at these N time points according to the following rules: When the fluctuation amplitude of the high-frequency signal prediction result exceeds the first preset capacity threshold and is less than the second preset capacity threshold, and the fluctuation duration is greater than the first preset duration and less than the second preset duration, it is determined as a capacity regeneration phenomenon. When the fluctuation amplitude of the high-frequency signal prediction result exceeds the second preset capacity threshold or the fluctuation duration exceeds the second preset duration, it is determined as an abnormal fluctuation. When the fluctuation amplitude of the high-frequency signal prediction result is less than the first preset capacity threshold or the fluctuation duration is less than the first preset duration, it is determined as a normal fluctuation. For the time points determined as the capacity regeneration phenomenon, keep the prediction result of the high-frequency signal unchanged. For the time points determined as abnormal fluctuations, correct the prediction result of the high-frequency signal to the average value of the prediction results of the adjacent K normal fluctuation time points. For the time points determined as normal fluctuations, keep the prediction result of the high-frequency signal unchanged.

7. The method for predicting the battery health state of an energy storage power station based on data driving according to claim 6, characterized in that: Superimpose the corrected prediction result of the high-frequency signal and the prediction result of the main degradation trend signal according to the adaptive weight function to obtain the predicted value of the lithium battery health state: Construct a dynamic weighting coefficient based on time series attenuation; the dynamic weighting coefficient is negatively correlated with the occurrence frequency of the capacity regeneration phenomenon in the corrected prediction result of the high-frequency signal and positively correlated with the monotonicity of the main degradation trend signal. Normalize the dynamic weighting coefficient and the adaptive weight function to obtain the final weighting coefficient; where, when the capacity regeneration phenomenon occurs, the final weighting coefficient uses a linear decreasing function for smooth transition during the duration of the capacity regeneration phenomenon. The predicted result of the corrected high-frequency signal is weighted and superimposed with the predicted result of the main degradation trend signal by using the final weighting coefficient. When the superimposed result is greater than the predicted value of the health state of the lithium battery at the previous moment, the predicted value of the health state of the lithium battery at this moment is corrected to the weighted average of the predicted value at the previous moment and the superimposed result.

8. A system adopting the data-driven energy storage power station battery health state prediction method as described in any one of claims 1 to 7, characterized in that, Including: A data decomposition module, configured to obtain a capacity data sequence of the lithium battery in the energy storage power station, and perform variational mode decomposition on the capacity data sequence to obtain at least two component signals; A model training module, configured to divide the at least two component signals into training data and test data respectively according to a preset ratio, train a first prediction model by using the training data of the main degradation trend signal, and train a second prediction model by using the training data of the at least one high-frequency signal; A health state prediction module, configured to superimpose the predicted result of the main degradation trend signal by the first prediction model and the predicted result of the high-frequency signal by the second prediction model to obtain the predicted value of the health state of the lithium battery.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the data-driven method for predicting the health state of the energy storage power station battery according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the data-driven method for predicting the health state of the energy storage power station battery according to any one of claims 1 to 7 are implemented.

Images