A method for predicting the remaining service life of an energy storage battery

Abstract

The application discloses a kind of energy storage battery remaining useful life prediction methods, comprising the following steps: obtaining capacity time series data and real remaining useful life label, obtain standard capacity time series data;Variational mode decomposition is carried out;Capacity degradation feature data is constructed;Build BP neural network model, and expand to obtain BP parameter vector;Improved philosophy proposition optimization algorithm is configured, and BP parameter candidate population is generated;Remaining useful life prediction is carried out to obtain candidate prediction error;Long-term decay error contribution value, short-period fluctuation error contribution value and local recovery error contribution value are obtained;Proposition state alternation factor is generated;Optimal BP parameter vector is obtained and written into BP neural network model, and energy storage battery remaining useful life prediction result is obtained.The application utilizes variational mode decomposition and improved philosophy proposition optimization BP to predict battery life, with the advantages of high precision and strong stability.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery life prediction technology, and in particular to a method for predicting the remaining lifespan of energy storage batteries. Background Technology

[0002] In recent years, the global transition to clean and low-carbon energy has been continuously advancing, and energy storage systems are playing an increasingly important role in renewable energy consumption and grid stability. As the core carrier of energy storage systems, the service life of energy storage batteries directly affects the system's reliability, economy, and safety. Since battery costs account for a large proportion of the total lifecycle cost of an energy storage system, inaccurate estimation of remaining battery life can easily lead to premature replacement or exceeding the service life, increasing maintenance costs and potentially causing safety hazards such as leakage and fire. Therefore, improving the accuracy of remaining battery life prediction is of great significance for the operation and maintenance management of energy storage systems.

[0003] Battery remaining life refers to the remaining time or number of charge-discharge cycles a battery can continue to operate normally from its current state until it can no longer meet performance requirements. It is an important basis for battery management systems to optimize life, provide safety warnings, and make replacement decisions. The capacity degradation process of energy storage batteries is affected by factors such as charge-discharge rate, ambient temperature and humidity, number of cycles, depth of charge and discharge, and internal electrochemical reactions, exhibiting significant nonlinearity and uncertainty. Furthermore, in capacity time series, the capacity degradation process usually simultaneously shows a long-term degradation trend, short-cycle fluctuations, and local capacity rebounds, causing the true capacity degradation and disruptive capacity changes to intertwine, increasing the difficulty of predicting remaining life.

[0004] Currently, methods for predicting the remaining lifespan of batteries mainly include mechanistic modeling, statistical analysis, and empirical formula fitting. Mechanistic modeling is complex and computationally intensive, making it difficult to adapt to complex operating conditions; statistical analysis relies on a large amount of complete aging data and has insufficient generalization ability; empirical formula fitting is limited by experimental conditions and cannot take into account multiple influencing factors. In addition, prediction methods based on BP neural networks are prone to misidentifying short-period fluctuations or local capacity rebounds as true capacity degradation patterns when processing capacity time series; even when variational mode decomposition is used to obtain multiple intrinsic mode functions, these intrinsic mode functions are usually used as ordinary input features without optimizing the BP neural network weights and thresholds according to the error contributions corresponding to long-term degradation, short-period fluctuations, and local rebounds. Therefore, there is an urgent need for a method for predicting the remaining lifespan of energy storage batteries that can distinguish different capacity degradation components and improve prediction stability. Summary of the Invention

[0005] One objective of this invention is to propose a method for predicting the remaining lifespan of energy storage batteries. This invention utilizes variational mode decomposition and improved philosophical propositions to optimize BP prediction of battery lifespan, which has the advantages of high accuracy and strong stability.

[0006] A method for predicting the remaining service life of an energy storage battery according to an embodiment of the present invention includes the following steps: Acquire the capacity time series data and the actual remaining service life tag of the energy storage battery during continuous charge and discharge cycles, preprocess the capacity time series data to obtain standard capacity time series data; Variational mode decomposition is performed on standard capacity time series data to obtain multiple intrinsic mode functions and the center frequency, bandwidth and energy percentage of each intrinsic mode function; Capacity degradation characteristic data are constructed based on the capacity change direction, capacity change amplitude, center frequency, bandwidth, and energy proportion of each intrinsic mode function; Build a BP neural network model, and expand the weights and thresholds of the BP neural network model into a BP parameter vector; An improved philosophical proposition optimization algorithm is configured using the BP parameter vector as the optimization object, and a candidate population of BP parameters is generated by using Henon chaotic mapping. The candidate weights and candidate thresholds corresponding to each candidate BP parameter vector are written into the BP neural network model. The remaining useful life is predicted based on the capacity degradation feature data, and the candidate prediction error is obtained based on the actual remaining useful life label. The contribution values ​​of long-term decay error, short-cycle fluctuation error, and local rebound error are obtained based on the candidate prediction error. The proposition state rotation factor is generated based on the contribution values ​​of long-term decay error, short-cycle fluctuation error, and local recovery error. By utilizing the candidate individual fitness evaluation strategy in the improved philosophical proposition optimization algorithm, the fitness values ​​of candidate individuals are determined. The candidate population of BP parameters is iteratively updated according to the proposition state rotation factor to obtain the optimal BP parameter vector. The optimal BP parameter vector is written into the BP neural network model, and the remaining service life prediction result of the energy storage battery is obtained by inputting capacity degradation feature data.

[0007] Optionally, the generation of the standard capacity time series data includes: Acquire the capacity time series data of the energy storage battery during continuous charge-discharge cycles; Abnormal capacity values ​​are removed, missing capacity values ​​are filled in, and normalization is performed on the capacity time series data to obtain standard capacity time series data.

[0008] Optionally, obtaining multiple eigenmode functions and the corresponding center frequency, bandwidth, and energy percentage of each eigenmode function includes: The number of modes for variational mode decomposition is determined based on the standard capacity time series data, and the standard capacity time series data is decomposed and initialized according to the number of modes; Frequency domain constraint decomposition is performed on the capacity change components in the standard capacity time series data, and different capacity change components are converged to their corresponding modal frequency regions. The capacity variation components in each modal frequency region are iteratively updated to obtain multiple intrinsic mode functions corresponding to the standard capacity time series data; The center frequency is determined based on the frequency concentration location of each intrinsic mode function, and the corresponding bandwidth is determined based on the frequency distribution range of each intrinsic mode function around the center frequency. The modal energy is calculated based on the capacity component values ​​contained in each intrinsic modal function, and the energy proportion of each intrinsic modal function is obtained based on the proportion of the modal energy of each intrinsic modal function in the total modal energy of all intrinsic modal functions.

[0009] Optionally, the construction of the capacity degradation feature data includes: Read the capacity component values ​​corresponding to each intrinsic mode function according to the charge-discharge cycle arrangement order of the standard capacity time series data, calculate the capacity difference between adjacent cycles for the same intrinsic mode function, and determine the downward direction, the upward direction and the capacity change amplitude; Based on the capacity change magnitude corresponding to the downward direction, the cumulative value of the decrease and the contribution value of the capacity decrease are obtained; The cumulative value of the recovery amplitude is obtained based on the capacity change amplitude corresponding to the recovery direction, and the recovery disturbance coefficient is determined according to the center frequency and bandwidth of the corresponding intrinsic mode function to obtain the local recovery contribution value of the corresponding intrinsic mode function; The amplitude fluctuation is determined based on the dispersion of the capacity change amplitude within the number of consecutive cycles, and the frequency fluctuation is determined based on the center frequency and bandwidth. The amplitude fluctuation and frequency fluctuation are weighted and combined to obtain the fluctuation intensity value of the corresponding intrinsic mode function. The capacity degradation feature vector of the corresponding intrinsic mode function is obtained by arranging the capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth and energy ratio in a fixed order for the same intrinsic mode function. The capacity degradation feature vectors are summarized according to the order of the multiple intrinsic mode functions to obtain the capacity degradation feature data.

[0010] Optionally, the generation of the BP parameter vector includes: Based on the capacity degradation feature data, including capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy proportion, the number of input layer nodes of the BP neural network model is determined, and the capacity degradation feature data is input into the input layer of the BP neural network model in the order of capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy proportion. The predicted remaining lifespan of the energy storage battery is determined as the output object of the BP neural network model, and the number of outputs corresponding to the output object is determined as the number of output layer nodes of the BP neural network model. The number of hidden layer nodes in the BP neural network model is determined based on the number of input layer nodes and the number of output layer nodes, and the BP neural network model is built according to the structure of the input layer, hidden layer and output layer connected in sequence. The connection parameters between the input layer nodes and the hidden layer nodes are determined as the input layer to hidden layer weights, the connection parameters between the hidden layer nodes and the output layer nodes are determined as the hidden layer to output layer weights, the bias parameters corresponding to the hidden layer nodes are determined as the hidden layer thresholds, and the bias parameters corresponding to the output layer nodes are determined as the output layer thresholds. The BP parameter vector is obtained by vectorizing the weights from the input layer to the hidden layer, the weights from the hidden layer to the output layer, the hidden layer threshold, and the output layer threshold in a fixed order.

[0011] Optionally, the generation of the BP parameter candidate population includes: The BP parameter vector is determined as the optimization object of the improved philosophical proposition optimization algorithm. The candidate individual dimension is determined according to the number of parameters in the BP parameter vector, and the mapping range of each dimension of the candidate individual is determined according to the value range of the BP parameter vector. In the improved philosophical proposition optimization algorithm, the initial value of the two-dimensional chaotic recursion of the Henon chaotic initialization strategy is set according to the candidate individual dimension, and the Henon chaotic mapping recursion is executed according to the initial value of the two-dimensional chaotic recursion to generate a chaotic sequence consistent with the candidate individual dimension. By utilizing the Henon chaos initialization strategy in the improved philosophical proposition optimization algorithm, the chaotic values ​​in the chaotic sequence are mapped to intervals according to the mapping range of each dimension of the candidate individuals, so as to obtain the candidate BP parameter vector with the same dimension as the BP parameter vector. Repeatedly execute the Henon chaotic mapping recursion and interval mapping to obtain multiple candidate BP parameter vectors, and form a BP parameter candidate population from these multiple candidate BP parameter vectors.

[0012] Optionally, the generation of the candidate prediction error includes: According to the fixed arrangement order of the BP parameter vectors, the parameter values ​​in each candidate BP parameter vector are read and written into the weight position and threshold position of the BP neural network model respectively, so as to obtain the BP neural network model after writing the parameters of each candidate BP parameter vector. After the capacity degradation feature data is written into the BP neural network model, forward propagation is performed according to the connection order of the input layer, hidden layer and output layer of the BP neural network model to obtain the candidate remaining lifetime prediction results corresponding to each candidate BP parameter vector. Read the actual remaining useful life label corresponding to the capacity degradation feature data, and calculate the difference between each candidate remaining useful life prediction result and the actual remaining useful life label to obtain the error value corresponding to each candidate BP parameter vector. According to the order of the candidate BP parameter vectors in the BP parameter candidate population, the error values ​​corresponding to each candidate BP parameter vector are arranged, and the arranged error values ​​are determined as the candidate prediction errors corresponding to each candidate BP parameter vector.

[0013] Optionally, the generation of the local recovery error contribution value includes: Read the candidate prediction error, capacity decline contribution value, fluctuation intensity value, and local recovery contribution value corresponding to each candidate BP parameter vector; The contribution values ​​of capacity decrease, fluctuation intensity, and local recovery are summed to obtain the total modal contribution value. The proportions of the contribution values ​​of capacity decrease, fluctuation intensity, and local recovery in the total modal contribution value are calculated to obtain the proportion values ​​of capacity decrease, fluctuation intensity, and local recovery. The candidate prediction error is allocated according to the proportion of capacity decline to obtain the error component corresponding to the capacity decline contribution value, and the error component is determined as the long-term decay error contribution value. The candidate prediction errors are allocated according to the proportion of fluctuation intensity to obtain the error components corresponding to the fluctuation intensity values, and the error components are determined as the contribution values ​​of short-cycle fluctuation errors. The candidate prediction error is allocated according to the proportion of local recovery, and the error component corresponding to the local recovery contribution value is obtained. The error component is then determined as the local recovery error contribution value.

[0014] Optionally, the generation of the proposition state rotation factor includes: Under the modal error-driven proposition state rotation strategy in the improved philosophical proposition optimization algorithm, the long-term decay error contribution value, short-period fluctuation error contribution value, and local recovery error contribution value are read and summed to obtain the total modal error contribution value. The ratio of the long-term decay error contribution value to the total modal error contribution value is determined as the JTB state weight, the ratio of the short-period fluctuation error contribution value to the total modal error contribution value is determined as the UTB state weight, and the ratio of the local recovery error contribution value to the total modal error contribution value is determined as the PFB state weight. The JTB state weight is determined as the participation ratio of the JTB state in the propositional state rotation factor, the UTB state weight is determined as the participation ratio of the UTB state in the propositional state rotation factor, and the PFB state weight is determined as the participation ratio of the PFB state in the propositional state rotation factor. The participation ratios of JTB, UTB, and PFB states are arranged in a fixed order to obtain the proposition state rotation factor.

[0015] Optionally, the generation of the predicted remaining lifespan of the energy storage battery includes: Under the candidate individual fitness evaluation strategy in the improved philosophical proposition optimization algorithm, the candidate prediction error, long-term decay error contribution value, short-period fluctuation error contribution value and local recovery error contribution value are read to obtain the candidate individual fitness value corresponding to each candidate BP parameter vector; The candidate BP parameter vectors in the candidate population are sorted according to the fitness values ​​of the candidate individuals, and the candidate BP parameter vector with the smallest fitness value is determined as the current optimal BP parameter vector. The participation ratios of JTB, UTB, and PFB states in the BP parameter candidate population are determined based on the proposition state rotation factor, and the BP parameter candidate population is divided into JTB update candidate group, UTB update candidate group, and PFB update candidate group according to the participation ratio. Approach update, perturbation update, and rise suppression update are performed on the JTB update candidate group, UTB update candidate group, and PFB update candidate group respectively, and the updated BP parameter candidate population is obtained by summarizing. Based on the updated BP parameter candidate population, the fitness values ​​of candidate individuals are re-determined, and the determination of the current optimal BP parameter vector and the update of the BP parameter candidate population are repeated until the iteration termination condition is met, and the optimal BP parameter vector is determined. The weights and thresholds corresponding to the optimal BP parameter vector are used as the optimal weights and thresholds of the BP neural network model. The capacity degradation feature data is then input into the BP neural network model with the optimal weights and thresholds to obtain the prediction results of the remaining service life of the energy storage battery.

[0016] The beneficial effects of this invention are: This invention uses variational mode decomposition to decompose the time series data of energy storage battery capacity, obtaining multiple intrinsic mode functions and their corresponding center frequencies, bandwidths and energy percentages. Furthermore, it extracts the contribution values ​​of capacity decline, local rebound and fluctuation intensity values ​​to construct capacity degradation characteristic data, thereby reducing the impact of the mixture of long-term decay trends, short-period fluctuations and local capacity rebounds in the original capacity time series on the prediction of remaining service life. Meanwhile, this invention utilizes an improved philosophical proposition optimization algorithm to optimize the weights and thresholds of the BP neural network model. It improves the dispersion of the BP parameter candidate population through a Henon chaotic initialization strategy and generates JTB, UTB, and PFB state weights based on the long-term decay error contribution, short-cycle fluctuation error contribution, and local recovery error contribution, thereby generating a proposition state rotation factor to control the differentiated iterative update of the BP parameter candidate population. This enables the BP neural network model to reduce the misleading influence of short-cycle fluctuations and local capacity recovery on prediction results, improving the accuracy and stability of remaining battery life prediction. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a method for predicting the remaining service life of an energy storage battery proposed in this invention. Figure 2 This is a schematic diagram illustrating the process of extracting capacity decline contribution value, local recovery contribution value, and fluctuation intensity value based on intrinsic mode function and constructing capacity degradation characteristic data in the energy storage battery remaining service life prediction method proposed in this invention. Figure 3 This is a flowchart illustrating the optimization of the BP parameter vector using an improved philosophical proposition optimization algorithm in a method for predicting the remaining lifespan of an energy storage battery proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A method for predicting the remaining lifespan of an energy storage battery includes the following steps: Acquire the capacity time series data and corresponding actual remaining service life tags of the energy storage battery during continuous charge and discharge cycles, preprocess the capacity time series data to obtain standard capacity time series data; Variational mode decomposition is performed on standard capacity time series data to obtain multiple intrinsic mode functions and the center frequency, bandwidth and energy percentage of each intrinsic mode function; Based on the capacity change direction, capacity change amplitude, center frequency, bandwidth, and energy percentage of each intrinsic mode function between adjacent cycles, the capacity decline contribution value, local recovery contribution value, and fluctuation intensity value are extracted respectively. Capacity degradation characteristic data are then constructed from the capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy percentage. A BP neural network model is constructed, with capacity degradation feature data as input and the predicted remaining lifespan of the energy storage battery as output. The weights and thresholds of the BP neural network model are expanded into a BP parameter vector in a fixed order. An improved philosophical proposition optimization algorithm is configured using BP parameter vectors as the optimization object. The improved philosophical proposition optimization algorithm includes Henon chaotic initialization strategy, modality error-driven proposition state rotation strategy and candidate individual fitness evaluation strategy. Henon chaotic mapping is used to generate chaotic sequence, and the chaotic sequence is mapped to the value range of BP parameter vector to obtain a BP parameter candidate population composed of multiple candidate BP parameter vectors. Each candidate BP parameter vector corresponds to a set of candidate weights and candidate thresholds of the BP neural network model. The candidate weights and candidate thresholds corresponding to each candidate BP parameter vector are written into the BP neural network model. The remaining useful life is predicted based on the capacity degradation feature data to obtain the candidate remaining useful life prediction results. The candidate prediction error is obtained based on the difference between the candidate remaining useful life prediction results and the actual remaining useful life label. The long-term decay error contribution value is obtained based on the correspondence between candidate prediction error and capacity decline contribution value; the short-cycle fluctuation error contribution value is obtained based on the correspondence between candidate prediction error and fluctuation intensity value; and the local rebound error contribution value is obtained based on the correspondence between candidate prediction error and local rebound contribution value. By utilizing the modal error-driven proposition state rotation strategy in the improved philosophical proposition optimization algorithm, JTB state weights are generated based on the long-term decay error contribution value, UTB state weights are generated based on the short-period fluctuation error contribution value, and PFB state weights are generated based on the local recovery error contribution value. The proposition state rotation factor is then generated based on the JTB state weights, UTB state weights, and PFB state weights. By utilizing the candidate individual fitness evaluation strategy in the improved philosophical proposition optimization algorithm, the fitness value of candidate individuals is determined based on the contribution values ​​of candidate prediction error, long-term decay error, short-cycle fluctuation error, and local recovery error. The candidate population of BP parameters is iteratively updated by controlling the JTB state, UTB state, and PFB state according to the proposition state rotation factor until the iteration termination condition is met, resulting in the optimal BP parameter vector. The weights and thresholds corresponding to the optimal BP parameter vector are then used as the optimal weights and thresholds of the BP neural network model. Capacity degradation feature data is input into the BP neural network model with the optimal weights and thresholds to obtain the predicted remaining lifespan of the energy storage battery.

[0020] In this embodiment, the generation of standard capacity time series data includes: Acquire the capacity time series data of the energy storage battery during continuous charge-discharge cycles; Abnormal capacity values ​​are removed, missing capacity values ​​are filled in, and normalization is performed on the capacity time series data to obtain standard capacity time series data.

[0021] In this embodiment, obtaining multiple intrinsic mode functions and the corresponding center frequency, bandwidth, and energy percentage of each intrinsic mode function includes: The number of modes for variational mode decomposition is determined based on the standard capacity time series data, and the standard capacity time series data is decomposed and initialized according to the number of modes; The optimization problem of variational mode decomposition can be expressed as the following constrained model:

[0022] In the formula, Let F(t) represent the Kth intrinsic mode function, and let F(t) represent the standard capacity time series data. This represents the bandwidth-constrained expression of the Kth intrinsic mode function after processing with its time partial derivative. This represents the modal representation of the Kth intrinsic mode function after frequency domain demodulation. This represents the frequency domain demodulation operator corresponding to the Kth eigenmode function. For imaginary numbers, Indicates time Find the partial derivative; The number of modes is determined based on the complexity of the frequency distribution of capacity variation components in the standard capacity time series data. The number of modes is used to limit the number of intrinsic mode functions output by the subsequent variational mode decomposition. During decomposition initialization, each mode to be decomposed is assigned a corresponding initial frequency position so that different capacity variation components converge in the subsequent frequency domain constrained decomposition. Frequency domain constraint decomposition is performed on the capacity change components in the standard capacity time series data, and different capacity change components are converged to their corresponding modal frequency regions. Frequency domain constraint decomposition is used to allow the capacity change components in standard capacity time series data to enter different modal frequency regions according to the degree of frequency concentration. The low-frequency modal frequency region corresponds to the long-term capacity decay component, the high-frequency modal frequency region corresponds to the short-period fluctuation component, and the intermediate frequency region is used to carry the local capacity rebound disturbance component. The capacity variation components in each modal frequency region are iteratively updated to obtain multiple intrinsic mode functions corresponding to the standard capacity time series data; The specific modal update formulas include:

[0023] In the formula, This represents the frequency domain estimation result of the Kth eigenmode function at the (n+1)th iteration. This is the inverse Fourier transform. This indicates that standard capacity time series data is in frequency Frequency domain representation at that location, This is the bandwidth constraint coefficient. For time step, Indicates frequency as The estimated value of the k-th component. For the Laplace operator; The iterative update includes adjusting the bandwidth constraint of the capacity variation components in the current modal frequency region based on the center frequency corresponding to the current modal frequency region in each iteration, and updating the frequency domain expression corresponding to the capacity variation components. After the capacity variation components in each modal frequency region are updated, they are reconstructed together with the capacity variation components in other modal frequency regions. The difference between the reconstruction result and the standard capacity time series data is used as the iteration error. When the iteration error meets the convergence condition, the capacity variation components corresponding to each modal frequency region are converted into time domain capacity components to obtain multiple intrinsic mode functions. The center frequency is determined based on the frequency concentration location of each intrinsic mode function, and the corresponding bandwidth is determined based on the frequency distribution range of each intrinsic mode function around the center frequency. The center frequency is determined by the location of the frequency energy concentration of the intrinsic mode function, and the bandwidth is determined by the frequency extension range of the intrinsic mode function around the center frequency. The center frequency and bandwidth are used to distinguish the capacity change type corresponding to the intrinsic mode function and participate in the calculation of the local recovery contribution value and the fluctuation intensity value. The modal energy is calculated based on the capacity component values ​​contained in each intrinsic modal function, and the energy proportion of each intrinsic modal function is obtained based on the proportion of the modal energy of each intrinsic modal function in the total modal energy of all intrinsic modal functions.

[0024] In this embodiment, the construction of capacity degradation feature data includes: According to the charging and discharging cycle arrangement order corresponding to the standard capacity time series data, the capacity component values ​​corresponding to each intrinsic mode function are read. The capacity difference between adjacent cycles of the same intrinsic mode function is calculated. The capacity change direction corresponding to the capacity difference being less than zero is determined as the decreasing direction, and the capacity change direction corresponding to the capacity difference being greater than zero is determined as the rising direction. The absolute value of the capacity difference is determined as the capacity change amplitude. The capacity component values ​​corresponding to each intrinsic mode function are arranged in the order of charge and discharge cycles corresponding to the standard capacity time series data. The capacity difference between adjacent capacity component values ​​is used to determine the direction and magnitude of capacity change of the intrinsic mode function at adjacent cycle positions. The direction of capacity change is used to distinguish between decreasing and rising changes. For the same intrinsic mode function, the capacity change amplitude corresponding to the descent direction is accumulated in order of the number of cycles to obtain the cumulative value of the descent amplitude. The cumulative value of the descent amplitude is then multiplied by the energy proportion of the corresponding intrinsic mode function to obtain the capacity descent contribution value of the corresponding intrinsic mode function. For the same intrinsic mode function, the capacity change amplitude corresponding to the recovery direction is accumulated according to the charging and discharging cycle arrangement order corresponding to the standard capacity time series data to obtain the cumulative recovery amplitude value. The recovery disturbance coefficient is determined according to the center frequency and bandwidth of the corresponding intrinsic mode function. The cumulative recovery amplitude value is multiplied by the recovery disturbance coefficient to obtain the local recovery contribution value of the corresponding intrinsic mode function. The rise disturbance coefficient is determined by the center frequency and bandwidth, and is used to characterize the disturbance property of the local rise change in the corresponding intrinsic mode function. When the center frequency is more concentrated and the bandwidth is narrower, the local rise contribution value is more biased towards the stable rise component; when the bandwidth expansion range increases, the local rise contribution value is more biased towards the disturbed rise component. For the same intrinsic mode function, the amplitude fluctuation is determined based on the dispersion of the capacity change amplitude within the number of consecutive cycles, and the frequency fluctuation is determined based on the center frequency and bandwidth. The amplitude fluctuation and frequency fluctuation are weighted and combined to obtain the fluctuation intensity value of the corresponding intrinsic mode function. The amplitude fluctuation is determined by the degree of dispersion of the capacity change amplitude in the same intrinsic mode function, and the frequency fluctuation is determined by the center frequency and bandwidth. The amplitude fluctuation and frequency fluctuation are weighted and combined to simultaneously characterize the impact of capacity numerical fluctuation and frequency distribution fluctuation on the prediction of remaining useful life. The capacity degradation feature vector of the corresponding intrinsic mode function is obtained by arranging the capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth and energy ratio in a fixed order for the same intrinsic mode function. The capacity degradation feature vectors are summarized according to the order of the multiple intrinsic mode functions to obtain the capacity degradation feature data.

[0025] In this embodiment, the generation of the BP parameter vector includes: Based on the capacity degradation feature data, including capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy proportion, the number of input layer nodes of the BP neural network model is determined, and the capacity degradation feature data is input into the input layer of the BP neural network model in the order of capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy proportion. The input features in the capacity degradation feature data are arranged in a fixed order of capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth and energy proportion. The number of input layer nodes of the BP neural network model is consistent with the feature dimension when the capacity degradation feature data is input, so that each input layer node corresponds to a capacity degradation feature. The predicted remaining lifespan of the energy storage battery is determined as the output object of the BP neural network model, and the number of outputs corresponding to the output object is determined as the number of output layer nodes of the BP neural network model. The number of hidden layer nodes in the BP neural network model is determined based on the number of input layer nodes and the number of output layer nodes, and the BP neural network model is built according to the structure of the input layer, hidden layer and output layer connected in sequence. The connection parameters between the input layer nodes and the hidden layer nodes are determined as the input layer to hidden layer weights, the connection parameters between the hidden layer nodes and the output layer nodes are determined as the hidden layer to output layer weights, the bias parameters corresponding to the hidden layer nodes are determined as the hidden layer thresholds, and the bias parameters corresponding to the output layer nodes are determined as the output layer thresholds. Following a fixed arrangement of input layer to hidden layer weights, hidden layer to output layer weights, hidden layer thresholds, and output layer thresholds, the input layer to hidden layer weights, hidden layer to output layer weights, hidden layer thresholds, and output layer thresholds are vectorized to obtain the BP parameter vector. The BP parameter vector is formed by arranging the weights from the input layer to the hidden layer, the weights from the hidden layer to the output layer, the hidden layer threshold, and the output layer threshold in sequence. This fixed arrangement order is used to ensure that the parameter values ​​in the candidate BP parameter vector can be accurately written back to the corresponding weight and threshold positions of the BP neural network model.

[0026] In this embodiment, the generation of the BP parameter candidate population includes: The BP parameter vector is determined as the optimization object of the improved philosophical proposition optimization algorithm. The candidate individual dimension is determined according to the number of parameters in the BP parameter vector, and the mapping range of each dimension of the candidate individual is determined according to the value range of the BP parameter vector. The improved philosophical proposition optimization algorithm includes a Henon chaotic initialization strategy, a modal error-driven proposition state rotation strategy, and a candidate individual fitness evaluation strategy. The Henon chaotic initialization strategy is used to generate the candidate population of BP parameters, the modal error-driven proposition state rotation strategy is used to generate the proposition state rotation factor, and the candidate individual fitness evaluation strategy is used to determine the fitness value of the candidate individual and update the candidate population of BP parameters according to the proposition state rotation factor. Henon chaotic mapping initializes the population:

[0027]

[0028]

[0029] In the formula, in the formula, Represents the initial population. represents the value of the i-th individual in the j-th dimension, i∈1,2…n; j∈1,2,…,d; n is the population size, d is the dimension of the problem; ub and lb are the upper and lower bounds of the problem, respectively, and rand represents a random number between 0 and 1; The improvement of the philosophical proposition optimization algorithm is mainly reflected in two aspects: Henon chaotic initialization strategy and modality error-driven proposition state rotation strategy. In the initialization stage, the BP parameter vector is used as the optimization object. The input layer to hidden layer weights, hidden layer to output layer weights, hidden layer thresholds and output layer thresholds of the BP neural network model are uniformly expanded into BP parameter vectors. Henon chaotic mapping is used to generate chaotic sequences, which are then mapped to the value range of the BP parameter vectors to form a BP parameter candidate population. This makes the candidate weights and candidate thresholds more dispersed in the search space, reducing the risk of concentrated search starting points and getting trapped in local optima due to random initialization. During the iterative update phase, instead of evaluating candidate BP parameter vectors solely based on overall candidate prediction errors, the candidate prediction errors are allocated as long-term decay error contribution values, short-cycle fluctuation error contribution values, and local recovery error contribution values ​​based on capacity decline contribution values, fluctuation intensity values, and local recovery error contribution values. Based on these, JTB state weights, UTB state weights, and PFB state weights are generated, forming a propositional state rotation factor. The JTB state guides the candidate BP parameter vectors toward the current optimal BP parameter vector, the UTB state performs perturbation search, and the PFB state performs recovery suppression update. This allows the BP parameter optimization process to perform differentiated searches for actual capacity decay, short-cycle capacity fluctuations, and local capacity recovery errors, improving the stability and accuracy of remaining lifespan prediction for energy storage batteries. The range of values ​​for the BP parameter vector is determined by the boundary of the weights and thresholds in the BP neural network model. The candidate individual dimension is consistent with the number of parameters in the BP parameter vector. Each candidate individual dimension corresponds to a weight parameter or threshold parameter in the BP parameter vector, so that the candidate BP parameter vector generated by the Henon chaotic mapping can be directly used as the candidate parameters of the BP neural network model. In the improved philosophical proposition optimization algorithm, the two-dimensional chaotic recursion initial value of the Henon chaotic initialization strategy is set according to the candidate individual dimension. The two-dimensional chaotic recursion initial value is used to start the Henon chaotic mapping recursion process. The chaotic values ​​obtained by recursion are truncated or combined into a chaotic sequence according to the candidate individual dimension. The Henon chaotic mapping recursion is executed according to the two-dimensional chaotic recursion initial value to generate a chaotic sequence consistent with the candidate individual dimension. By utilizing the Henon chaos initialization strategy in the improved philosophical proposition optimization algorithm, the chaotic values ​​in the chaotic sequence are mapped to intervals according to the mapping range of each dimension of the candidate individuals, so as to obtain the candidate BP parameter vector with the same dimension as the BP parameter vector. During interval mapping, each chaotic value in the chaotic sequence is mapped to the mapping range of the corresponding dimension of the candidate individual. Each dimension value after mapping corresponds to a parameter position in the BP parameter vector, resulting in a candidate BP parameter vector with the same dimension as the BP parameter vector. Repeatedly execute Henon chaotic mapping recursion and interval mapping to obtain multiple candidate BP parameter vectors. These multiple candidate BP parameter vectors form a BP parameter candidate population. According to the fixed arrangement order of the BP parameter vectors, the parameter values ​​in each candidate BP parameter vector are mapped to a set of candidate weights and candidate thresholds for the BP neural network model.

[0030] In this embodiment, the generation of candidate prediction error includes: According to the fixed arrangement order of the BP parameter vectors, the parameter values ​​in each candidate BP parameter vector are read and written into the weight position and threshold position of the BP neural network model respectively, so as to obtain the BP neural network model after writing the parameters of each candidate BP parameter vector. When writing parameters, the parameter values ​​in the candidate BP parameter vector are written sequentially according to the fixed arrangement order of the BP parameter vector. The input layer to hidden layer weights, the hidden layer to output layer weights, the hidden layer threshold, and the output layer threshold are written in turn. The BP neural network model after writing is used to make candidate predictions for the same capacity degradation feature data. After the capacity degradation feature data is written into the BP neural network model, forward propagation is performed according to the connection order of the input layer, hidden layer and output layer of the BP neural network model to obtain the candidate remaining lifetime prediction results corresponding to each candidate BP parameter vector. Read the actual remaining useful life label corresponding to the capacity degradation feature data, and calculate the difference between each candidate remaining useful life prediction result and the actual remaining useful life label to obtain the error value corresponding to each candidate BP parameter vector. According to the order of the candidate BP parameter vectors in the BP parameter candidate population, the error values ​​corresponding to each candidate BP parameter vector are arranged, and the arranged error values ​​are determined as the candidate prediction errors corresponding to each candidate BP parameter vector. The candidate prediction errors are stored according to the order of the candidate BP parameter vectors in the BP parameter candidate population. This correspondence is used to further allocate the prediction error of each candidate BP parameter vector into long-term decay error contribution value, short-period fluctuation error contribution value, and local recovery error contribution value.

[0031] In this embodiment, the generation of the local recovery error contribution value includes: Read the candidate prediction error, capacity decline contribution value, fluctuation intensity value, and local recovery contribution value corresponding to each candidate BP parameter vector; The contribution values ​​of capacity decrease, fluctuation intensity, and local recovery are summed to obtain the total modal contribution value. The proportions of the contribution values ​​of capacity decrease, fluctuation intensity, and local recovery in the total modal contribution value are calculated to obtain the proportion values ​​of capacity decrease, fluctuation intensity, and local recovery. The total modal contribution is obtained by summing the capacity decline contribution, fluctuation intensity contribution, and local recovery contribution, and is used to distribute the three types of modal contributions on the same scale. The proportion of capacity decline, fluctuation intensity, and local recovery respectively represent the contribution ratio of the three types of degradation components in the current capacity degradation feature data. The candidate prediction error is allocated according to the proportion of capacity decline to obtain the error component corresponding to the capacity decline contribution value, and the error component is determined as the long-term decay error contribution value. Allocating candidate prediction errors according to the proportion of capacity decline means taking the candidate prediction errors as the total amount of errors to be allocated, taking the proportion of capacity decline as the long-term decay error allocation coefficient, and multiplying the candidate prediction errors and the proportion of capacity decline to obtain the error components corresponding to the contribution value of capacity decline. The candidate prediction errors are allocated according to the proportion of fluctuation intensity to obtain the error components corresponding to the fluctuation intensity values, and the error components are determined as the contribution values ​​of short-cycle fluctuation errors. Allocating candidate prediction errors according to the proportion of fluctuation intensity means taking the candidate prediction errors as the total amount of errors to be allocated, taking the proportion of fluctuation intensity as the allocation coefficient for short-cycle fluctuation errors, and multiplying the candidate prediction errors and the proportion of fluctuation intensity to obtain the error components corresponding to the fluctuation intensity values. The candidate prediction error is allocated according to the proportion of local recovery, and the error component corresponding to the local recovery contribution value is obtained. The error component is then determined as the local recovery error contribution value. Allocating candidate prediction errors according to the proportion of local recovery means taking the candidate prediction error as the total error to be allocated, taking the proportion of local recovery as the local recovery error allocation coefficient, and multiplying the candidate prediction error and the proportion of local recovery to obtain the error component corresponding to the contribution value of local recovery.

[0032] In this embodiment, the generation of the proposition state rotation factor includes: Under the modal error-driven proposition state rotation strategy in the improved philosophical proposition optimization algorithm, the long-term decay error contribution value, short-period fluctuation error contribution value, and local recovery error contribution value are read, and the long-term decay error contribution value, short-period fluctuation error contribution value, and local recovery error contribution value are summed to obtain the total modal error contribution value, which is used to normalize the three types of error contributions. The ratio of the long-term decay error contribution value to the total modal error contribution value is determined as the JTB state weight, the ratio of the short-period fluctuation error contribution value to the total modal error contribution value is determined as the UTB state weight, and the ratio of the local recovery error contribution value to the total modal error contribution value is determined as the PFB state weight. The JTB state weight is determined as the participation ratio of the JTB state in the propositional state rotation factor, the UTB state weight is determined as the participation ratio of the UTB state in the propositional state rotation factor, and the PFB state weight is determined as the participation ratio of the PFB state in the propositional state rotation factor. The participation ratios of JTB, UTB, and PFB states are arranged in a fixed order to obtain the proposition state rotation factor.

[0033] In this embodiment, the generation of the remaining lifespan prediction result of the energy storage battery includes: Under the candidate individual fitness evaluation strategy in the improved philosophical proposition optimization algorithm, the contribution values ​​of candidate prediction error, long-term decay error, short-period fluctuation error, and local recovery error are read, and the candidate prediction error, long-term decay error, short-period fluctuation error, and local recovery error are weighted and combined to obtain the candidate individual fitness value corresponding to each candidate BP parameter vector. The candidate BP parameter vectors in the candidate population are sorted according to the fitness values ​​of the candidate individuals, and the candidate BP parameter vector with the smallest fitness value is determined as the current optimal BP parameter vector. The participation ratios of JTB, UTB, and PFB states in the BP parameter candidate population are determined based on the proposition state rotation factor, and the BP parameter candidate population is divided into JTB update candidate group, UTB update candidate group, and PFB update candidate group according to the participation ratio. Perform a convergent update on the candidate BP parameter vectors in the JTB update candidate group to obtain the JTB update parameter vector; perform a perturbation update on the candidate BP parameter vectors in the UTB update candidate group to obtain the UTB update parameter vector; and perform a rise suppression update on the candidate BP parameter vectors in the PFB update candidate group to obtain the PFB update parameter vector. Approach update refers to using the current optimal BP parameter vector as the approach target, calculating the parameter difference between the candidate BP parameter vectors in the JTB update candidate group and the current optimal BP parameter vector, scaling the parameter difference according to the participation ratio of the JTB state in the proposition state rotation factor, and superimposing the scaled parameter difference onto the corresponding candidate BP parameter vector to obtain the JTB update parameter vector. Perturbation update refers to generating a perturbation amplitude based on the current position of the candidate BP parameter vector in the UTB update candidate group, according to the participation ratio of the UTB state in the proposition state rotation factor, and adding perturbation displacement to each parameter dimension of the candidate BP parameter vector to obtain the UTB update parameter vector. The rise suppression update refers to determining the rise suppression coefficient based on the contribution value of the local rise error, and scaling the update direction of the candidate BP parameter vector in the PFB update candidate group according to the participation ratio of the PFB state in the proposition state rotation factor, so that the parameter update direction that leads to the increase of the contribution value of the local rise error is weakened, and the PFB update parameter vector is obtained. The updated parameter vectors of JTB, UTB, and PFB are summarized to obtain the updated candidate population of BP parameters. Based on the updated BP parameter candidate population, the fitness values ​​of candidate individuals are re-determined, and the determination of the current optimal BP parameter vector and the update of the BP parameter candidate population are repeated until the iteration termination condition is met. The current optimal BP parameter vector that meets the iteration termination condition is determined as the optimal BP parameter vector. The weights and thresholds corresponding to the optimal BP parameter vector are used as the optimal weights and thresholds of the BP neural network model. The capacity degradation feature data is then input into the BP neural network model with the optimal weights and thresholds to obtain the prediction results of the remaining service life of the energy storage battery.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to the battery operation and maintenance management scenario of an energy storage power station. This energy storage power station has long been responsible for smoothing new energy output, peak shaving and valley filling, and backup support. The energy storage batteries in the station continuously undergo charge and discharge cycles. During daily management, the operation and maintenance personnel found that the capacity curves of some batteries did not decline smoothly and continuously, but instead exhibited long-term degradation, short-cycle fluctuations, and local capacity rebounds simultaneously. When predicting the remaining service life directly based on the original capacity curves, the BP neural network model is prone to misinterpreting short-cycle fluctuations as actual capacity changes, or misinterpreting local capacity rebounds as battery recovery, leading to deviations in the prediction results and affecting subsequent maintenance, group rotation, and replacement arrangements.

[0035] In this scenario, the battery management system first acquires the capacity time-series data of the energy storage battery during continuous charge-discharge cycles, along with the corresponding true remaining lifespan tags. The capacity time-series data primarily comes from capacity records after the end of charge-discharge cycles. The true remaining lifespan tags are labeled based on the remaining operating state when the battery capacity approaches the end-of-life capacity threshold. Subsequently, the capacity time-series data undergoes abnormal capacity value removal, missing capacity value imputation, and normalization to obtain standard capacity time-series data. This processing unifies the data scale across different batteries, and reduces the impact of abnormal sampling by sensors, communication interruptions, and missing records on subsequent predictions.

[0036] Next, variational mode decomposition is performed on the standard capacity time series data to obtain multiple intrinsic mode functions (IMFs) and their corresponding center frequencies, bandwidths, and energy proportions. Based on the direction and magnitude of capacity changes between adjacent capacity components of each IMF, capacity decline contribution values, local recovery contribution values, and fluctuation intensity values ​​are extracted. These values ​​are then combined in a fixed order to construct capacity degradation feature data. In this way, the real capacity decay, short-period disturbances, and local capacity recoveries mixed in the original capacity time series are transformed into feature data that can be input into the model, allowing subsequent models to no longer rely solely on changes in a single capacity curve for judgment.

[0037] Then, a BP neural network model is constructed, using capacity degradation feature data as input and the predicted remaining lifespan of the energy storage battery as output. The weights and thresholds in the BP neural network model are expanded into BP parameter vectors and used as optimization objects for the improved philosophical proposition optimization algorithm. This algorithm first generates a candidate population of BP parameters through a Henon chaotic initialization strategy, making the candidate weights and thresholds more dispersed in the search space, reducing the parameter concentration problem caused by random initialization, and also reducing the risk of the BP neural network model getting stuck in local optimization during early training.

[0038] During training, each candidate BP parameter vector is written into the BP neural network model. The model outputs the candidate remaining useful life prediction results based on the capacity degradation feature data and calculates the difference with the actual remaining useful life label to obtain the candidate prediction error. Then, based on the contribution value of capacity decline, fluctuation intensity, and local recovery, the candidate prediction error is allocated as the contribution value of long-term decay error, short-cycle fluctuation error, and local recovery error. The improved philosophical proposition optimization algorithm then generates JTB state weights, UTB state weights, and PFB state weights based on the three types of error contribution values, and forms a proposition state rotation factor, so that subsequent parameter updates can be adjusted according to different error sources.

[0039] During iterative updates, the JTB state guides the candidate BP parameter vectors toward the current optimal BP parameter vector, enhancing the model's learning of the true capacity decay trend; the UTB state performs perturbation updates to correct local prediction biases caused by short-cycle capacity fluctuations; the PFB state performs recovery suppression updates to reduce the skewed impact of local capacity recovery on lifetime prediction results; when the iteration termination condition is met, the optimal BP parameter vector is output, and its corresponding weights and thresholds are used as the optimal weights and thresholds of the BP neural network model, ultimately obtaining the remaining lifetime prediction result of the energy storage battery.

[0040] By applying this invention to the operation and maintenance management of energy storage power stations, it can maintain stable prediction results when capacity curves exhibit short-term fluctuations, avoid overly optimistic lifespan assessments when capacity shows localized rebounds, and improve the ability to identify the actual degradation process when long-term capacity degradation trends are obvious. Maintenance personnel can classify and track batteries based on prediction results, placing batteries with continuously increasing degradation trends under close observation, while continuing to monitor batteries with significant short-term fluctuations but stable long-term degradation trends. Batteries experiencing localized rebound interference can be comprehensively assessed in conjunction with subsequent capacity changes. Therefore, this invention can reduce the misleading influence of capacity fluctuations and localized rebounds on BP neural network prediction results, improve the accuracy and stability of remaining lifespan predictions for energy storage batteries, and provide a more reliable basis for energy storage battery inspection, maintenance, and replacement scheduling.

[0041] Table 1. Comparison of Overall Performance of Predicted Remaining Lifespan of Energy Storage Batteries

[0042] As shown in Table 1, the traditional BP neural network prediction method has a mean absolute error of 42.6 iterations, a root mean square error of 56.8 iterations, a mean absolute percentage error of 8.73%, a prediction fluctuation range of 31.4 iterations, 9 misclassifications due to local capacity rebounds, and a prediction time of 0.21 seconds. Although this method is computationally fast, it directly performs nonlinear mapping on the capacity degradation data, making it susceptible to short-period fluctuations and interference from local capacity rebounds. Therefore, both the error and the number of misclassifications are at a high level.

[0043] The VMD-BP prediction method improves the representation of input data through variational mode decomposition, reducing the mean absolute error to 31.8 iterations, the root mean square error to 42.5 iterations, the mean absolute percentage error to 6.41%, the fluctuation range of the prediction result to 24.7 iterations, the number of local rebound misclassifications to 6, and the time taken for a single prediction to 0.28 seconds. These results demonstrate that VMD can mitigate the impact of the non-stationarity of the original capacity time series on the BP neural network. However, this method does not further incorporate the error sources corresponding to different intrinsic mode functions into the parameter optimization process.

[0044] The VMD-Particle Swarm Optimization (PSO) BP prediction method has a mean absolute error of 22.7 iterations, a root mean square error of 30.4 iterations, a mean absolute percentage error of 4.76%, a prediction fluctuation range of 17.9 iterations, 4 instances of local rebound misjudgments, and a single prediction time of 0.53 seconds. This method improves both the input data and the BP parameter optimization process, thus outperforming traditional BP neural network prediction methods and VMD-BP prediction methods. However, its fitness evaluation still primarily relies on the overall prediction error, making it difficult to distinguish whether the error originates from long-term decay, short-period fluctuations, or local rebounds.

[0045] The method of this invention has a mean absolute error of 15.3 iterations, a root mean square error of 21.2 iterations, a mean absolute percentage error of 3.18%, a prediction fluctuation range of 10.6 iterations, one local rebound misjudgment, and a single prediction time of 0.61 seconds. Compared with the VMD-Particle Swarm Optimization BP prediction method, the method of this invention reduces the mean absolute error by 7.4 iterations, the root mean square error by 9.2 iterations, the mean absolute percentage error by 1.58 percentage points, the prediction fluctuation range by 7.3 iterations, and the number of local rebound misjudgments by 3.

[0046] The reason this invention achieves the aforementioned effects is that it further transforms the intrinsic mode function into a capacity decline contribution value, a local rebound contribution value, and a fluctuation intensity value, and allocates candidate prediction errors as long-term decay error contribution values, short-period fluctuation error contribution values, and local rebound error contribution values, enabling the BP parameter optimization process to identify error sources. The Henon chaotic initialization strategy improves the dispersion of the BP parameter candidate population, while the mode error-driven propositional state rotation strategy performs approach update, perturbation update, and rebound suppression update through JTB state, UTB state, and PFB state, respectively, thus reducing the misleading effect of local rebounds and short-period fluctuations on the prediction results. Although the single prediction time of this invention is 0.61 seconds, which is higher than other methods, it still meets the requirements for energy storage battery operation and maintenance prediction.

[0047] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining service life of an energy storage battery, characterized in that, Includes the following steps: Acquire the capacity time series data and the actual remaining service life tag of the energy storage battery during continuous charge and discharge cycles, preprocess the capacity time series data to obtain standard capacity time series data; Variational mode decomposition is performed on standard capacity time series data to obtain multiple intrinsic mode functions and the center frequency, bandwidth and energy percentage of each intrinsic mode function; Capacity degradation characteristic data are constructed based on the capacity change direction, capacity change amplitude, center frequency, bandwidth, and energy proportion of each intrinsic mode function; Build a BP neural network model, and expand the weights and thresholds of the BP neural network model into a BP parameter vector; An improved philosophical proposition optimization algorithm is configured using the BP parameter vector as the optimization object, and a candidate population of BP parameters is generated by using Henon chaotic mapping. The candidate weights and candidate thresholds corresponding to each candidate BP parameter vector are written into the BP neural network model. The remaining useful life is predicted based on the capacity degradation feature data, and the candidate prediction error is obtained based on the actual remaining useful life label. The contribution values ​​of long-term decay error, short-cycle fluctuation error, and local rebound error are obtained based on the candidate prediction error. The proposition state rotation factor is generated based on the contribution values ​​of long-term decay error, short-cycle fluctuation error, and local recovery error. By utilizing the candidate individual fitness evaluation strategy in the improved philosophical proposition optimization algorithm, the fitness values ​​of candidate individuals are determined. The candidate population of BP parameters is iteratively updated according to the proposition state rotation factor to obtain the optimal BP parameter vector. The optimal BP parameter vector is written into the BP neural network model, and the remaining service life prediction result of the energy storage battery is obtained by inputting capacity degradation feature data.

2. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the standard capacity time series data includes: Acquire the capacity time series data of the energy storage battery during continuous charge-discharge cycles; Abnormal capacity values ​​are removed, missing capacity values ​​are filled in, and normalization is performed on the capacity time series data to obtain standard capacity time series data.

3. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The process of obtaining multiple eigenmode functions and the corresponding center frequency, bandwidth, and energy percentage for each eigenmode function includes: The number of modes for variational mode decomposition is determined based on the standard capacity time series data, and the standard capacity time series data is decomposed and initialized according to the number of modes; Frequency domain constraint decomposition is performed on the capacity change components in the standard capacity time series data, and different capacity change components are converged to their corresponding modal frequency regions. The capacity variation components in each modal frequency region are iteratively updated to obtain multiple intrinsic mode functions corresponding to the standard capacity time series data; The center frequency is determined based on the frequency concentration location of each intrinsic mode function, and the corresponding bandwidth is determined based on the frequency distribution range of each intrinsic mode function around the center frequency. The modal energy is calculated based on the capacity component values ​​contained in each intrinsic modal function, and the energy proportion of each intrinsic modal function is obtained based on the proportion of the modal energy of each intrinsic modal function in the total modal energy of all intrinsic modal functions.

4. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The construction of the capacity degradation feature data includes: Read the capacity component values ​​corresponding to each intrinsic mode function according to the charge-discharge cycle arrangement order of the standard capacity time series data, calculate the capacity difference between adjacent cycles for the same intrinsic mode function, and determine the downward direction, the upward direction and the capacity change amplitude; Based on the capacity change magnitude corresponding to the downward direction, the cumulative value of the decrease and the contribution value of the capacity decrease are obtained; The cumulative value of the recovery amplitude is obtained based on the capacity change amplitude corresponding to the recovery direction, and the recovery disturbance coefficient is determined according to the center frequency and bandwidth of the corresponding intrinsic mode function to obtain the local recovery contribution value of the corresponding intrinsic mode function; The amplitude fluctuation is determined based on the dispersion of the capacity change amplitude within the number of consecutive cycles, and the frequency fluctuation is determined based on the center frequency and bandwidth. The amplitude fluctuation and frequency fluctuation are weighted and combined to obtain the fluctuation intensity value of the corresponding intrinsic mode function. The capacity degradation feature vector of the corresponding intrinsic mode function is obtained by arranging the capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth and energy ratio in a fixed order for the same intrinsic mode function. The capacity degradation feature vectors are summarized according to the order of the multiple intrinsic mode functions to obtain the capacity degradation feature data.

5. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the BP parameter vector includes: Based on the capacity degradation feature data, including capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy proportion, the number of input layer nodes of the BP neural network model is determined, and the capacity degradation feature data is input into the input layer of the BP neural network model in the order of capacity decline contribution value, local recovery contribution value, fluctuation intensity value, center frequency, bandwidth, and energy proportion. The predicted remaining lifespan of the energy storage battery is determined as the output object of the BP neural network model, and the number of outputs corresponding to the output object is determined as the number of output layer nodes of the BP neural network model. The number of hidden layer nodes in the BP neural network model is determined based on the number of input layer nodes and the number of output layer nodes, and the BP neural network model is built according to the structure of the input layer, hidden layer and output layer connected in sequence. The connection parameters between the input layer nodes and the hidden layer nodes are determined as the input layer to hidden layer weights, the connection parameters between the hidden layer nodes and the output layer nodes are determined as the hidden layer to output layer weights, the bias parameters corresponding to the hidden layer nodes are determined as the hidden layer thresholds, and the bias parameters corresponding to the output layer nodes are determined as the output layer thresholds. The BP parameter vector is obtained by vectorizing the weights from the input layer to the hidden layer, the weights from the hidden layer to the output layer, the hidden layer threshold, and the output layer threshold in a fixed order.

6. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the BP parameter candidate population includes: The BP parameter vector is determined as the optimization object of the improved philosophical proposition optimization algorithm. The candidate individual dimension is determined according to the number of parameters in the BP parameter vector, and the mapping range of each dimension of the candidate individual is determined according to the value range of the BP parameter vector. In the improved philosophical proposition optimization algorithm, the initial value of the two-dimensional chaotic recursion of the Henon chaotic initialization strategy is set according to the candidate individual dimension, and the Henon chaotic mapping recursion is executed according to the initial value of the two-dimensional chaotic recursion to generate a chaotic sequence consistent with the candidate individual dimension. By utilizing the Henon chaos initialization strategy in the improved philosophical proposition optimization algorithm, the chaotic values ​​in the chaotic sequence are mapped to intervals according to the mapping range of each dimension of the candidate individuals, so as to obtain the candidate BP parameter vector with the same dimension as the BP parameter vector. Repeatedly execute the Henon chaotic mapping recursion and interval mapping to obtain multiple candidate BP parameter vectors, and form a BP parameter candidate population from these multiple candidate BP parameter vectors.

7. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the candidate prediction error includes: According to the fixed arrangement order of the BP parameter vectors, the parameter values ​​in each candidate BP parameter vector are read and written into the weight position and threshold position of the BP neural network model respectively, so as to obtain the BP neural network model after writing the parameters of each candidate BP parameter vector. After the capacity degradation feature data is written into the BP neural network model, forward propagation is performed according to the connection order of the input layer, hidden layer and output layer of the BP neural network model to obtain the candidate remaining lifetime prediction results corresponding to each candidate BP parameter vector. Read the actual remaining useful life label corresponding to the capacity degradation feature data, and calculate the difference between each candidate remaining useful life prediction result and the actual remaining useful life label to obtain the error value corresponding to each candidate BP parameter vector. According to the order of the candidate BP parameter vectors in the BP parameter candidate population, the error values ​​corresponding to each candidate BP parameter vector are arranged, and the arranged error values ​​are determined as the candidate prediction errors corresponding to each candidate BP parameter vector.

8. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the local recovery error contribution value includes: Read the candidate prediction error, capacity decline contribution value, fluctuation intensity value, and local recovery contribution value corresponding to each candidate BP parameter vector; The contribution values ​​of capacity decrease, fluctuation intensity, and local recovery are summed to obtain the total modal contribution value. The proportions of the contribution values ​​of capacity decrease, fluctuation intensity, and local recovery in the total modal contribution value are calculated to obtain the proportion values ​​of capacity decrease, fluctuation intensity, and local recovery. The candidate prediction error is allocated according to the proportion of capacity decline to obtain the error component corresponding to the capacity decline contribution value, and the error component is determined as the long-term decay error contribution value. The candidate prediction errors are allocated according to the proportion of fluctuation intensity to obtain the error components corresponding to the fluctuation intensity values, and the error components are determined as the contribution values ​​of short-cycle fluctuation errors. The candidate prediction error is allocated according to the proportion of local recovery, and the error component corresponding to the local recovery contribution value is obtained. The error component is then determined as the local recovery error contribution value.

9. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the proposition state rotation factor includes: Under the modal error-driven proposition state rotation strategy in the improved philosophical proposition optimization algorithm, the long-term decay error contribution value, short-period fluctuation error contribution value, and local recovery error contribution value are read and summed to obtain the total modal error contribution value. The ratio of the long-term decay error contribution value to the total modal error contribution value is determined as the JTB state weight, the ratio of the short-period fluctuation error contribution value to the total modal error contribution value is determined as the UTB state weight, and the ratio of the local recovery error contribution value to the total modal error contribution value is determined as the PFB state weight. The JTB state weight is determined as the participation ratio of the JTB state in the propositional state rotation factor, the UTB state weight is determined as the participation ratio of the UTB state in the propositional state rotation factor, and the PFB state weight is determined as the participation ratio of the PFB state in the propositional state rotation factor. The participation ratios of JTB, UTB, and PFB states are arranged in a fixed order to obtain the proposition state rotation factor.

10. The method for predicting the remaining service life of an energy storage battery according to claim 1, characterized in that, The generation of the predicted remaining lifespan of the energy storage battery includes: Under the candidate individual fitness evaluation strategy in the improved philosophical proposition optimization algorithm, the candidate prediction error, long-term decay error contribution value, short-period fluctuation error contribution value and local recovery error contribution value are read to obtain the candidate individual fitness value corresponding to each candidate BP parameter vector; The candidate BP parameter vectors in the candidate population are sorted according to the fitness values ​​of the candidate individuals, and the candidate BP parameter vector with the smallest fitness value is determined as the current optimal BP parameter vector. The participation ratios of JTB, UTB, and PFB states in the BP parameter candidate population are determined based on the proposition state rotation factor, and the BP parameter candidate population is divided into JTB update candidate group, UTB update candidate group, and PFB update candidate group according to the participation ratio. Approach update, perturbation update, and rise suppression update are performed on the JTB update candidate group, UTB update candidate group, and PFB update candidate group respectively, and the updated BP parameter candidate population is obtained by summarizing. Based on the updated BP parameter candidate population, the fitness values ​​of candidate individuals are re-determined, and the determination of the current optimal BP parameter vector and the update of the BP parameter candidate population are repeated until the iteration termination condition is met, and the optimal BP parameter vector is determined. The weights and thresholds corresponding to the optimal BP parameter vector are used as the optimal weights and thresholds of the BP neural network model. The capacity degradation feature data is then input into the BP neural network model with the optimal weights and thresholds to obtain the prediction results of the remaining service life of the energy storage battery.

Images