DC power supply health state prediction model optimization method based on extreme learning machine

Abstract

The invention discloses a DC power supply health state prediction model optimization method based on an extreme learning machine, and relates to the technical field of power supply intelligent management. According to the method, an extreme learning machine basic model is improved, multi-layer ELM-AE is adopted to carry out greedy unsupervised learning on input data layer by layer, advanced features of an original data set are extracted, a mixed kernel function is adopted as a regression layer of a supervised learning stage, the complex nonlinear problem is solved, and the method has the advantages of being high in robustness and high in robustness. The adaptability of the prediction model to diversified characteristics under different battery working conditions is improved, the generalization ability and prediction precision of the model are improved, and the convergence speed and prediction precision of the model are remarkably improved by optimizing the model parameters through the improved black-wing calculation algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of power supply intelligent management, and in particular to a direct current power supply health state prediction model optimization method based on an extreme learning machine. Background Art

[0002] In the station DC power supply system, the application of new batteries is the key to ensure the stability and reliability of power supply. DC power supply systems often need to cope with the demand for long-term power supply, but their performance will gradually decline with the use time and charge and discharge cycles. This decline is manifested as a decline in the state of health (SOH). In the DC power supply system, the health state of the battery is directly related to the power supply capacity and reliability of the system. If the battery health state is not effectively managed, it may cause insufficient power supply to the system or even power outages at critical moments. In order to ensure the continuous and reliable operation of the DC power supply system, the health state of lithium batteries must be accurately predicted and managed. By monitoring the key performance parameters of lithium batteries, such as charge and discharge characteristics, discharge depth, energy density, thermal stability and cycle life, the health status of the battery can be understood in real time, and preventive maintenance or replacement of battery packs can be performed when necessary, thereby ensuring that the system can continue to operate stably under any operating conditions.

[0003] At present, the prediction technology of the health status of DC batteries such as lithium batteries and lead-acid batteries mainly relies on physical models and data-driven. Physical models generally rely on the built battery model, which is highly complex and has a high modeling threshold. In addition, the measured data usually contains a lot of noise, is easily disturbed by noise, and has poor adaptability. The data-driven method usually only relies on external observable data, does not need to obtain difficult-to-obtain internal parameters, can handle complex nonlinear relationships and is easy to expand, making it an efficient and flexible solution for battery health status prediction.

[0004] Extreme Learning Machine (ELM) is a widely used data-driven prediction model with a simple structure, fast learning speed and easy implementation. However, the traditional ELM also has some limitations: (1) The prediction accuracy of the ELM model depends on the number of hidden layer nodes. The selection of this parameter is often empirical and it is difficult to determine the optimal value; (2) The kernel function of ELM is relatively simple. Common kernel functions such as radial basis function (RBF) and linear kernel function cannot adapt well to the diverse characteristics of different battery working conditions, resulting in limited generalization ability and prediction accuracy of the model. Summary of the invention

[0005] The present invention provides a method for optimizing a DC power supply health status prediction model based on an extreme learning machine, so as to solve the problems that the existing prediction model has poor adaptability to the diverse working conditions of different batteries, and the prediction accuracy and generalization ability of the model are limited.

[0006] The present invention is achieved through the following technical solutions:

[0007] A first aspect of the present invention provides a method for optimizing a DC power supply health status prediction model based on an extreme learning machine, comprising:

[0008] Extracting capacity characteristics and aging characteristics of the DC power supply, and quantifying the capacity characteristics into power supply health status indicator values;

[0009] Constructing a prediction model based on an extreme learning machine, the prediction model comprising a multi-layer extreme learning machine autoencoder, a kernel function and an output layer; wherein the kernel function combines a local kernel function and a global kernel function;

[0010] The multi-layer extreme learning machine autoencoder is used to extract the aging features layer by layer to obtain hidden features;

[0011] The kernel function is used to extract the hidden features to obtain high-dimensional prediction features;

[0012] The output layer is used to predict and output the high-dimensional prediction features;

[0013] The prediction model is trained using the aging characteristics and the power health status indicator value.

[0014] The present invention combines a multi-layer extreme learning machine autoencoder and a hybrid kernel function to improve the extreme learning machine model. By extracting data features layer by layer through a deep structure, higher-dimensional data features can be extracted, thereby better capturing nonlinear components, and to a certain extent avoiding the influence of the number of hidden layer nodes. At the same time, by integrating two types of kernel functions, global kernel functions and local kernel functions, instead of a single kernel function, better adaptation to complex battery operating condition data is achieved, and the generalization ability and prediction accuracy of the model are improved.

[0015] Furthermore, the global kernel function adopts one of a linear kernel function, a polynomial kernel function, and a Sigmoid kernel, and the local kernel function adopts a Gaussian kernel function.

[0016] Furthermore, the kernel function is a linear combination of a Gaussian kernel function and a polynomial kernel function.

[0017] Furthermore, the method also includes: optimizing model hyperparameters using a genetic algorithm, a particle swarm optimization algorithm, a differential evolution algorithm or a black kite algorithm.

[0018] Furthermore, the optimization of model hyperparameters by using the Black Kite algorithm includes:

[0019] S1, establish the fitness function according to the mean square error between the model prediction value and the actual value, set the maximum number of iterations and the dimension of the optimization parameter;

[0020] S2, evenly distribute the position of each black-winged kite to obtain the initial position, and select the best individual as the leader in the initial population according to the fitness function;

[0021] S3, updating the position of each black-winged kite in the next iteration according to the black-winged kite attack behavior model;

[0022] S4, updating leaders based on the black-winged kite migration behavior model;

[0023] S5, repeat iterative steps S3 to S4 until the maximum number of iterations is reached, output the optimal position of the population, and obtain the optimal hyperparameters.

[0024] Furthermore, the black kite attack behavior model is:

[0025]

[0026] n = a*exp(b*exp(ct));

[0027] in, represents the position of the i-th black-winged kite in the j-th dimension and the t-th iteration step, represents the position of the i-th black-winged kite in the t+1-th iteration step of the j-th dimension, r is a random number between 0 and 1, p is a constant of 0.9; T is the total number of iterations, t is the number of iterations completed so far, a is the upper asymptote, which controls the initial value of the step size, b is the control parameter, which controls the displacement of t; c is the scaling factor, which controls the growth rate.

[0028] Furthermore, the black-winged kite migration behavior model is:

[0029]

[0030] in, represents the leading scorer of the j-th black-winged kite for the t-th iteration completed so far, F i represents the j-th dimension current position of any black-winged kite in the t-th iteration, F ri represents the fitness value of any black kite at a random position in the jth dimension in the tth iteration, C(0,1) represents the Cauchy mutation, m is the displacement factor, and t max is the maximum number of iterations, μ and σ are random numbers between 0 and 1.

[0031] Furthermore, S2 includes: using a good point set initialization algorithm to generate a good point set of initial hyperparameters, mapping the good point set to a black kite population search space, and obtaining an initial position of the population.

[0032] A second aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for optimizing a DC power supply health status prediction model based on an extreme learning machine as described in any one of the first aspects of the present invention is implemented.

[0033] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for optimizing a DC power supply health status prediction model based on an extreme learning machine as described in any one of the first aspects of the present invention is implemented.

[0034] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0035] The present invention improves the basic model of the extreme learning machine, extracts data features layer by layer through a deep structure, and extracts higher-dimensional data features, thereby better capturing nonlinear components and achieving more accurate predictions.

[0036] By integrating two types of kernel functions, global kernel function and local kernel function, instead of a single kernel function, better adaptation to complex battery operating condition data is achieved, and the generalization ability and prediction accuracy of the model are improved.

[0037] Through the improved black-winged kite algorithm (IBKA), the kernel function parameters and model weights in the prediction model are optimized, which improves the global search and local search balance ability of the BKA algorithm, avoids falling into the local optimal solution, and significantly improves the convergence speed and prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0039] Figure 1 This is a battery health status prediction architecture diagram of a station DC power supply system according to an embodiment of the present invention;

[0040] Figure 2 is a prediction network diagram based on a deep hybrid kernel extreme learning machine (DHKELM) according to an embodiment of the present invention;

[0041] Figure 3a is the battery capacity degradation curve of data set A;

[0042] Figure 3b is the battery capacity degradation curve of data set A;

[0043] Figure 4a is the charging curve of a DC battery;

[0044] Figure 4b is the discharge curve of a DC battery;

[0045] Figure 5a This is a comparison chart of the prediction results of different algorithms for B5 batteries;

[0046] Figure 5b This is a comparison chart of the prediction results of different algorithms for B18 batteries;

[0047] Figure 5c This is a comparison chart of the prediction results of different algorithms for CS35 batteries;

[0048] Figure 5d This is a comparison chart of the prediction results of different algorithms for CS36 batteries. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0050] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to or inherent to other steps or units of the device.

[0051] The terms used in the various embodiments of the application are only used to describe the purpose of specific embodiments and are not intended to limit the various embodiments of the application. As used herein, the singular form is intended to also include the plural form, unless the context clearly indicates otherwise. Unless otherwise limited, all terms used here (including technical terms and scientific terms) have the same meaning as the meanings commonly understood by ordinary technicians in the field of the various embodiments of the application. The terms (such as the terms defined in the dictionary generally used) will be interpreted as having the same meaning as the contextual meaning in the relevant technical field and will not be interpreted as having an idealized meaning or an overly formal meaning, unless clearly defined in the various embodiments of the application.

[0052] In data-driven models, extreme learning machine (ELM) is a widely used prediction algorithm. It has attracted widespread attention due to its advantages such as fast learning speed, simple structure and easy implementation. ELM is a single hidden layer feedforward neural network with good generalization ability and can quickly process large amounts of data. However, the traditional ELM has the following limitations: First, the kernel function selection of ELM is relatively simple. Common kernel functions such as radial basis function (RBF) and linear kernel function cannot adapt well to the diverse characteristics under different battery working conditions, resulting in limited generalization ability and prediction accuracy of the model; second, the prediction accuracy of the ELM model depends on the selection of the number of hidden layer nodes. This parameter selection is often empirical and it is difficult to determine the optimal value; in addition, the weight and bias initialization in ELM are random. Although this feature speeds up the training speed, it is also easy to cause insufficient model stability and result fluctuations.

[0053] In terms of optimization, traditional data-driven models often combine genetic algorithms (GA) and particle swarm optimization (PSO) to optimize parameters. These optimization algorithms have certain performance in global search capabilities, but they often have problems such as easy to fall into local optimality and slow convergence, especially in the optimization process of high-dimensional complex problems. This directly affects the accuracy and reliability of battery health status prediction.

[0054] In view of the above technical deficiencies, the present invention proposes a method for optimizing a DC power supply health status prediction model based on an extreme learning machine, which is used to predict the health status of DC batteries in DC power supply systems for substations. In view of the problem that the traditional ELM model kernel function is single and cannot adapt well to the diversified features under different battery working conditions, the present invention adopts a multi-layer extreme learning machine autoencoder structure to realize abstract feature extraction, and then uses the extracted abstract features to train a hybrid kernel extreme learning machine to realize classification. By introducing different types of kernel functions, better adaptation to complex battery working condition data is achieved, and the generalization ability and prediction accuracy of the model are improved. In view of the problem that the model optimization algorithm is prone to fall into local optimality and slow convergence speed, the present invention adopts the black-winged kite algorithm, and by improving the black-winged kite algorithm, its search efficiency in high-dimensional optimization problems is enhanced. The deep hybrid kernel extreme learning machine achieves better adaptation to complex lithium battery working condition data by introducing different types of kernel functions, improves the generalization ability and prediction accuracy of the model, and at the same time, the improved black-winged kite algorithm can effectively optimize the hidden layer parameters, weights and biases of the ELM, so that the model has significant improvements in efficiency and accuracy.

[0055] In a first aspect, an embodiment of the present invention provides a method for optimizing a DC power supply health status prediction model based on an extreme learning machine, see Figure 1 The station DC power system battery health status prediction architecture diagram shown includes the following steps:

[0056] Step 1: Extract the capacity characteristics and aging characteristics of the DC power supply, and quantify the capacity characteristics into power supply health status indicator values.

[0057] Aging characteristics, also known as health factors, refer to various characteristics that reflect the battery status, such as battery voltage, current, temperature, etc. The battery management system monitors the battery operation data, extracts time series features, and obtains diverse aging characteristics for model training. In general, capacity can directly represent the health status of the battery, but in practice, battery capacity is often not easy to measure. Therefore, by analyzing the historical health factor data of the battery and establishing a battery health model, the battery health status can be predicted, providing a basis for battery maintenance and replacement.

[0058] Battery health status (SOH) refers to the ratio of a battery's performance parameters to its nominal parameters after a period of use. According to IEEE standards, when the capacity of a fully charged battery is less than 80% of the rated capacity, it is generally considered that the battery health status is poor and needs to be replaced. SOH is one of the important indicators for evaluating the degree of battery aging and remaining service life, so the capacitance characteristics are quantified as SOH indicators for prediction.

[0059] Step 2: Construct a prediction model DHKELM based on extreme learning machine, which includes a multi-layer extreme learning machine autoencoder, a kernel function and an output layer; wherein the kernel function combines the local kernel function and the global kernel function.

[0060] During model training or prediction, the aging features are extracted layer by layer through a multi-layer extreme learning machine autoencoder to obtain hidden features; the hidden features are extracted through a kernel function that combines the local kernel function and the global kernel function to obtain high-dimensional prediction features; finally, the high-dimensional prediction features are predicted and output through the output layer.

[0061] Step 3: Train the prediction model based on extreme learning machine through aging characteristics and power health status indicator values.

[0062] Before training, the acquired battery data is divided into a training set and a test set. The constructed prediction model is trained with the training set, and data prediction is performed with the test set. Error analysis is performed based on the prediction results and model parameters are adjusted.

[0063] The prediction network based on deep hybrid kernel extreme learning machine (DHKELM) constructed by the present invention is shown in Figure 2As shown in Figure 1, it constructs an extreme learning machine model with a hybrid kernel function based on the extreme learning machine-autoencoder (ELM-AE). ELM generates a local optimal solution by determining the number of hidden layer nodes, regardless of weights and biases. This method eliminates the need to set many parameters to train the network, which is required in traditional neural networks. Therefore, ELM has a faster convergence speed and higher learning efficiency than other methods.

[0064] The output model of ELM is shown in equation (1).

[0065] f(x)=βH (1)

[0066] Where H is the output matrix of the hidden layer nodes and β is the weight matrix between the hidden layer and the output layer, which is calculated using the following equation:

[0067]

[0068] In the formula, H + is the Mooree-Penrose generalized inverse matrix of the hidden layer node output matrix H, T is the actual value, H T is the transpose of the matrix H, C is the penalty parameter, which is a positive constant, and I is the identity matrix.

[0069] Therefore, the detailed expression of ELM can be obtained as follows:

[0070] f(x)=HH T (HH T +I / C) -1 T (3)

[0071] However, improper selection of the number of hidden layer nodes in ELM can lead to algorithm instability. To address this problem, the kernel function K(·) is incorporated into ELM, resulting in a kernel extreme learning machine (KELM) with enhanced stability and improved performance.

[0072] In KELM, the kernel matrix f KELM HH instead of ELM T Matrix, kernel function matrix is ​​K(x i ,x j ) is represented by, which is defined as in formula (4).

[0073]

[0074] In the formula, x i 、x j represents the data input sample, h(x i )、h(x j ) means converting the input sample into a feature vector in a high-dimensional feature space through a mapping function h.

[0075] The modeling function of KELM can be written as formula (5).

[0076]

[0077] In the formula, N represents the number of samples, x represents the input data sample vector, and x i represents the data input sample, C is the penalty parameter, f KELM is the kernel matrix and T is the actual value.

[0078] Traditional kernel functions include two types: global kernel functions and local kernel functions. The global kernel function emphasizes the influence of the entire data set on the kernel function, and has strong generalization ability, but limited learning ability. The local kernel function emphasizes the influence of data points close to the critical point on the kernel function, and has strong learning ability, but weak generalization ability. In order to improve the learning ability and generalization of the model, and to adapt to the diversified characteristics under different battery working conditions, the present invention adopts a kernel function that combines the two types of global kernel functions and local kernel functions to replace the single kernel function in KELM, and obtains a hybrid kernel extreme learning machine (HKELM), which can be fused by linear combination. Common linear kernels, polynomial kernels, and Sigmoid kernels belong to global kernel functions, and Gaussian kernel functions belong to local kernel functions.

[0079] The Gaussian kernel function is a typical local kernel function, as shown in the following formula:

[0080]

[0081] Where σ is used to control the radial range.

[0082] As a typical global kernel function, the polynomial kernel function can be expressed as:

[0083] K poly (x i ,x j )=((x i ,x j )+q) p (7)

[0084] Where q and p are function parameters.

[0085] In one embodiment of the present invention, in the fused kernel function, the global kernel function adopts one of a linear kernel function, a polynomial kernel function, and a Sigmoid kernel, and the local kernel function adopts a Gaussian kernel function.

[0086] Preferably, a linear combination of a Gaussian kernel function and a polynomial kernel function is used as the fusion kernel function, as shown in formula (8).

[0087] K(x i ,x j )=vK RBF +(1-v)K poly ,v∈[0,1] (8)

[0088] Among them, v represents the weight coefficient, v∈[0,1].

[0089] Autoencoder (AE) is an unsupervised neural network model commonly used in deep learning. The core concept of this model is to use input data for supervision, guide the neural network to establish a mapping relationship, and obtain a reconstructed output that is close to the input. However, AE still uses the error back propagation method during training, which leads to long training time and easy to fall into local optimality. Therefore, the researchers integrated the AE concept into ELM and proposed the extreme learning machine-autoencoder (ELM-AE), which constructs a mapping relationship between the input data of ELM through AE to obtain the corresponding reconstructed output.

[0090] The difference between ELM-AE and traditional AE is that the algorithm retains the characteristics of the ELM algorithm to randomly initialize the input weights and biases and make them orthogonal. By orthogonalizing the weights and biases, the input data can be transformed into a space with different dimensions. This transformation allows the creation of feature representations for various functions. Compared with AE, ELM-AE has stronger mapping and generalization capabilities.

[0091] In 2006, Hinton et al. proposed a method to construct a multi-layer neural network, first using unsupervised learning to train the parameters of each layer, and then using supervised learning to fine-tune the network. On this basis, Tang et al. proposed the concept of deep extreme learning machine (DELM) in 2016. This model uses ELM-AE as the basic unit of DELM and ELM as the final regression layer. Unlike traditional deep learning algorithms, ELM-AE and the final DELM regression layer use the least squares method, and only one step of inverse calculation is required to obtain the updated weights, thus eliminating the need for backward fine-tuning. This means that compared with traditional multi-layer neural networks, the training time and computing resource requirements of this algorithm will be significantly reduced.

[0092] Drawing on the above advantages, the hybrid kernel extreme learning machine (HKELM) model is used to replace the original regression layer in the deep extreme learning machine (DELM) model, and the optimization model proposed by the present invention, the deep hybrid kernel extreme learning machine (DHKELM), is obtained. The algorithm flow of the model includes two stages: an unsupervised learning stage and a supervised learning stage. In the unsupervised learning stage, ELM-AE is used as the base unit to perform greedy unsupervised pre-training on the input data layer by layer to learn the high-level features of the original data set. ELM-AE can reduce the reconstruction error. In this process, the low layer mainly extracts simple features and provides them to the high layer, while the high layer extracts more abstract and complex features. At the same time, the output weight matrix of the ELM-AE model obtained by the least squares method is saved. Then, HKELM is used as the regression layer in the supervised learning stage to solve complex nonlinear problems and improve prediction accuracy. HKELM is initialized using the multi-layer sparse features and output layer weights obtained in the unsupervised learning stage, and the output layer weights are combined with the label information.

[0093] The structure of the DHKELM network is as follows Figure 1 As shown in Figure 1, the network consists of m+1 hidden layers, where the first m hidden layers are extreme learning machine autoencoder (ELM-AE) units, and the last hidden layer is HKELM, which uses a fused kernel function to replace the regression layer in DELM. The state matrix of the hidden layer can be calculated using formula (9).

[0094]

[0095] In the formula, h t Represents the state matrix of the tth layer.

[0096] The weight of the i+1th hidden layer is given by formula (10).

[0097]

[0098] In the formula, H i represents the output of the i-th layer.

[0099] Through extraction, features are continuously passed from the previous layer to the next layer, and the output H of the previous layer i It is also used as the input of the next layer, as shown in formula (11).

[0100]

[0101] For the last hidden layer, HKELM uses the features extracted from the mth hidden layer for regression prediction. The weights and biases of the hybrid kernel mapping layer can be obtained by minimizing the regularized cost function through least squares estimation, as shown in formula (12).

[0102] Y=g(ain H m +b in ) (12)

[0103] Among them, g(·) represents the activation function, a in is the weight of the hybrid kernel mapping layer, b in is the bias of the hybrid kernel mapping layer, H m is the output of the mth hidden layer.

[0104] However, the performance of DHKELM still depends on the selection of its parameters, especially the kernel function parameters and model weights. Therefore, further, the present invention uses genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm or black kite optimization algorithm to optimize the model hyperparameters.

[0105] Genetic algorithms, particle swarm optimization algorithms, differential evolution algorithms, etc. are commonly used optimization algorithms. These optimization algorithms usually have certain performance in global search capabilities, but when faced with the optimization of high-dimensional complex problems, they are prone to fall into local optimality and slow convergence. Based on the improvement of the extreme learning machine prediction model by the present invention, the application effect of the above algorithms has been improved to a certain extent, but there are still certain limitations, which affect the accuracy and reliability of battery health status prediction.

[0106] The Black Kite Algorithm is an emerging swarm intelligence optimization algorithm that mimics the group migration behavior of black kite birds to solve high-dimensional complex optimization problems. Black kites are small birds with a blue-gray upper body and a white lower body. Their notable characteristics include migration and predatory behavior. They feed on small mammals, reptiles, birds, and insects and have strong hovering capabilities. Its optimization process is inspired by its hunting skills and migration habits and includes the following steps.

[0107] 1. Population initialization

[0108] As with most optimization algorithms, the position of each black kite is uniformly distributed in the population initialization, and the initial position is expressed as:

[0109] X i =BK lb +rand(BK ub -BK lb ) (13)

[0110] Among them, i is an integer between 1 and pop, BK lb and BK ub are the lower and upper bounds of the j-th dimension of the black-winged kite, respectively, and rand is a randomly selected value between [0, 1]. The black-winged kite dimension represents the parameter dimension of the optimized DHKELM model, that is, the number of dimensions is the same as the number of optimized parameters.

[0111] During the initialization process, BKA selects the individual with the best fitness value as the initial leader X in the initial group. L , which is considered the optimal position for black kites. Below is the initial leader X L Mathematical representation of , taking the minimum value as an example.

[0112] f best =min(f(X i )) (14)

[0113] X L =X(find(f best = = f(X i ))) (15)

[0114] Among them, f best represents the optimal position of the population, f(X i ) represents individual X i The fitness of .

[0115] 2. Aggressive Behavior

[0116] The black-winged kite is a predator of small mammals and insects in the grassland. During flight, it adjusts the angle of its wings and tail according to the wind speed, quietly hovers to observe its prey, and then quickly dives to attack. This strategy includes different attack behaviors for global exploration and search. There are two states, one is hovering in the air waiting to attack, and the other is hovering in the air looking for prey. Equations (16) and (17) are the mathematical models of the black-winged kite's attack behavior.

[0117]

[0118] In the formula, and denote the position of the ith black-winged kite in the tth and t+1th iteration steps of the jth dimension, r is a random number between 0 and 1, and p is a constant of 0.9; t max is the maximum number of iterations, and t is the number of iterations completed so far.

[0119] 3. Migration Behavior

[0120] Bird migration is a complex behavior affected by environmental factors such as climate and food supply. Bird migration is to adapt to seasonal changes, and many birds migrate from north to south in winter to obtain better living conditions and resources. Migration is usually led by leaders, and their navigation skills are crucial to the success of the team. We propose a hypothesis based on bird migration: if the fitness value of the current population is less than the fitness value of the random population, the leader will give up leadership and join the migrating population, indicating that it is not suitable to lead the population forward. On the contrary, if the fitness value of the current population is greater than the fitness value of the random population, it will guide the population to its destination. This strategy can dynamically select excellent leaders to ensure the success of migration.

[0121] Below is a mathematical model of the migratory behavior of black-winged kites:

[0122]

[0123] m=2×sin(r+π / 2) (19)

[0124] In the formula, represents the leading scorer of the j-th black-winged kite for the t-th iteration completed so far, F i represents the j-th dimension current position obtained by any black-winged kite in the t-th iteration; F ri represents the fitness value of any black-winged kite at a random position in the jth dimension in the tth iteration. C(0,1) represents the Cauchy mutation, which is defined as: The one-dimensional Cauchy distribution is a continuous probability distribution with two parameters. The probability density function of the one-dimensional Cauchy distribution is:

[0125]

[0126] Among them, when δ=1, μ=0, its probability density function becomes the standard form. The exact formula is as follows:

[0127]

[0128] The Black Kite Algorithm has certain global search capabilities and local development capabilities. However, although the Black Kite Algorithm has shown good optimization capabilities in many applications, it still has some significant shortcomings and deficiencies. First, when initializing the population, the Black Kite Algorithm uses a random method to generate individuals in the population, which leads to uneven distribution of solutions in the search space. It is easy for the population to be concentrated in a specific area, while there is a lack of individual coverage in other areas. This uneven distribution not only reduces the global search ability of the algorithm, but may also lead to insufficient population diversity, making the algorithm more likely to fall into the local optimal solution. Secondly, the original algorithm uses an exponential decay function when controlling the individual search step size. Although this method can gradually reduce the step size as the number of iterations increases, thereby improving the local search ability, the exponential decay is too fast, which will cause the step size to decrease rapidly, especially in the middle and late stages. The algorithm will lose the ability to explore a wider area in advance, resulting in premature convergence to a suboptimal solution. Finally, the migration behavior of the Black Kite Algorithm simulates the group flight of kite swarms, but due to the fixed migration step size, lack of randomness and flexibility, individuals are prone to repeated searches in local areas, and it is difficult to effectively jump out of the local optimal trap, which further affects the global search efficiency of the algorithm. In view of these shortcomings, the present invention adopts three improvement strategies: 1. In the initialization stage, the good point set initialization method is introduced to replace the original random initialization; 2. The attack behavior is improved by using the Gompertz growth model; 3. The migration behavior is improved by using the Levy flight strategy. Through these improvements, the Black Kite algorithm has been significantly improved in terms of the ability to balance global search and local development, jump out of local optimality and convergence speed, providing a more effective tool for solving the complex optimization problem of the health status of lithium batteries.

[0129] (1) Initialization of good point set

[0130] Since the Black Kite Algorithm uses a randomly initialized population, the generated population is unevenly distributed in the entire solution space. It is very concentrated in some areas and very dispersed in some areas, resulting in low utilization of the entire search space and weak population diversity. The initial population generated by the good point set initialization algorithm is more evenly distributed in the search space, avoiding the aggregation phenomenon that may occur during random initialization, and can cover the search space more comprehensively, thereby improving the global search ability of the algorithm, accelerating the convergence speed of the algorithm, and reducing the uncertainty caused by randomness. Therefore, the good point set initialization is chosen to improve the Black Kite Algorithm, and its principle and structure are as follows.

[0131] Let V d is a unit cube in d-dimensional space, if r∈V d , the set of numbers that form formula (22) can be called the good point set.

[0132]

[0133] Among them, the deviation satisfy C(r,ε)n -1+ε is a constant related only to r and ε (ε is any positive number), r is the optimal point, It means taking the decimal part, n is the number of good points, generally r={2cos(2π / p),1≤k≤d}, p is the smallest prime number satisfying d≤(p-3) / 2.

[0134] After the good point set is generated using formula (22), it is mapped to the search space to generate the initial population. The mapping formula is as follows: i =P n (k)*(BK ub -BK lb )+BK lb , B.K. lb , BK ub Respectively represent the lower and upper bounds of the black-winged kite.

[0135] (2) Improving aggressive behavior using the Gompertz growth model

[0136] The function of parameter n in formula (17) is to adjust the position update step of the individual in the search space. It gradually decreases with the increase of the number of iterations t, thereby controlling the exploration behavior of the algorithm. Specifically, n is a time-dependent coefficient, which reduces the search range through an exponential decay function, so that the algorithm has a stronger global search capability in the early stage, and pays more attention to local search and fine adjustment in the later stage. However, the exponential decay function used in the original formula may cause the decay rate to be very fast. As the number of iterations t increases, n decreases rapidly. This rapid decay will cause the algorithm to have a larger step size when performing global search in the early stage, and be able to explore a wider area, but the step size will shrink sharply in the middle and late stages, which may cause the algorithm to fall into the local optimal solution too early. Therefore, we use the Gompertz growth model to replace the parameter n in formula (17), so that the step size decays more gently and balances the global and local search capabilities. This improvement allows the algorithm to transition to the local search stage more smoothly, thereby improving the overall performance. The Gompertz growth model is as follows:

[0137]

[0138] n=a*exp(b*exp(ct)) (24)

[0139] In the formula, represents the position of the i-th black-winged kite in the j-th dimension and the t-th iteration step, represents the position of the i-th black-winged kite in the t+1-th iteration step of the j-th dimension, r is a random number between 0 and 1, p is a constant of 0.9; T is the total number of iterations, t is the number of iterations completed so far, a is the upper asymptote, which controls the initial value of the step size; b is the control parameter, which controls the displacement of t; c is the scaling factor, which controls the growth rate.

[0140] (3) Levy flight strategy improves migration behavior

[0141] The Levy flight strategy is a distribution with infinite variance, which contains many short-distance jumps and occasional large-step jumps. This allows for detailed exploration in local search while retaining the ability to jump out of the local optimal solution, improving the efficiency of global search. Through random jumps of different step lengths, Levy flight can effectively jump in a wide search space, avoid the algorithm from falling into the local optimum, and ensure a good balance between exploration and development. Therefore, the Levy flight strategy is introduced into the migration behavior of the black kite algorithm. After introducing this improvement, the modification of formula (18) is as follows:

[0142]

[0143] In the formula, represents the leading scorer of the j-th black-winged kite for the t-th iteration completed so far, F i represents the j-th dimension current position of any black-winged kite in the t-th iteration, F ri represents the fitness value of any black kite at a random position in the jth dimension in the tth iteration, C(0,1) represents the Cauchy mutation, m is the displacement factor, and t max The maximum number of iterations, μ and σ are random numbers between 0 and 1.

[0144] In view of the shortcomings of traditional ELM in dealing with nonlinear complex problems, the present invention proposes to adopt a hybrid kernel method, combined with an autoencoder, to form a deep hybrid kernel extreme learning machine (DHKELM) to enhance the nonlinear fitting ability of the model. However, since the performance of DHKELM still depends on the selection of its parameters, the present invention further proposes to use an improved black-winged kite algorithm (IBKA) to optimize the key parameters of DHKELM. It enhances the global exploration ability and local development ability of the algorithm by introducing a strategy of balancing global search and local search and a mechanism of jumping out of the local optimal solution, thereby avoiding falling into the local optimal solution. Compared with the traditional black-winged kite algorithm, IBKA has further improved the convergence speed and optimization accuracy, which can enable the DHKELM model to obtain better parameter values. The main steps are as follows:

[0145] S1, establish the fitness function according to the mean square error between the model prediction value and the actual value, set the maximum number of iterations and the dimension of the optimization parameter; the fitness function is shown in formula (28):

[0146]

[0147] In the formula, is the predicted value of SOH; y i is the actual SOH value.

[0148] S2, evenly distribute the position of each black-winged kite to obtain the initial position, and select the best individual as the leader in the initial population according to the fitness function;

[0149] S3, updating the position of each black-winged kite in the next iteration according to the black-winged kite attack behavior model;

[0150] S4, updating leaders based on the black-winged kite migration behavior model;

[0151] S5, repeat iterative steps S3 to S4 until the maximum number of iterations is reached, output the optimal position of the population, and obtain the optimal hyperparameters.

[0152] In some embodiments, the above step S3 uses the black kite attack behavior model optimized by the above method.

[0153] In some embodiments, the above step S4 uses the black kite migration behavior model optimized by the above method.

[0154] In some implementations, in step S2 above, a good point set initialization algorithm is used instead of a random initialization algorithm. First, a good point set of initial hyperparameters is generated using the good point set initialization algorithm, and the good point set is mapped to the black kite population search space to obtain the initial position of the population.

[0155] Finally, the obtained model hyperparameters are used to train the DHKELM model.

[0156] The embodiment of the present invention also verifies the trained model through experimental data. In order to verify the implementation effect of the lithium battery SOH prediction model based on IBKA-DHKELM proposed in the present invention, the present invention selects two different aging data sets, which are used for ELM, DHKELM, BKA-DHKELM and IBKA-DHKELM models respectively, to verify the superiority of the method of the present invention.

[0157] (1) Experimental Dataset and Feature Extraction

[0158] The present invention uses two different data sets, data set A is a lithium battery aging data set from NASA PCoE laboratory, and data set B is a lithium ion battery data set from the Center for Advanced Life Cycle Engineering (CALCE) of the University of Maryland.

[0159] like Figure 3a , 3b The figures show the battery capacity degradation curves of data set A and data set B, respectively. There are two battery data in data set A, namely B5 and B18, both of which use 18650 lithium-ion batteries with a nominal capacity of 2Ah for charge and discharge aging tests. Batteries B5 and B7 were tested at room temperature (24°C), and both were obtained by CC-CV cycle tests on the batteries. First, the battery was charged with a constant current of 1.5A. When the battery voltage reached 4.2V, the battery was charged in constant voltage mode until the charging current dropped below 20mA. The discharge test was performed with a constant current of 2A until the voltages of B5 and B7 dropped to 2.7V and 2.2V, respectively.

[0160] The two batteries in data set B are CS35 and CS36. The positive electrode material of the battery is lithium cobalt oxide (LiCoO2), and the nominal capacity is 1.1Ah. At the same room temperature (24°C), constant current charging is performed at a rate of 0.5C until the battery voltage reaches 4.2V. Then, charging is continued in constant voltage mode until the charging current drops to 50mA. The discharge test is performed in 1C constant current output mode until the voltage drops to 2.7V.

[0161] In the experiment, considering the correlation between health factor and SOH and the difficulty of extracting health factor, we finally extracted constant voltage rise charging time, constant voltage drop discharge time, constant current charging time and constant voltage discharge time from the charge and discharge data as health factors. The process of extracting health factor is as follows:

[0162] like Figure 4a , 4b The following are the charging and discharging curves of the DC battery. Figure 4a It can be seen that the charging time required for the second cycle is the longest, while the charging time required for the 166th cycle is the shortest. As the number of charge and discharge cycles increases, the charging time in the CC stage gradually shortens. It can be seen that the charging voltage in the CC stage implies the characteristics of battery performance degradation and can indirectly characterize the battery health status information. Here, the equal voltage rise charging time with a starting voltage of 3.8V and an end voltage of 4.2V is selected as the health factor.

[0163] During the discharge period of equal voltage drop, Figure 4b It can be seen that as the number of charge and discharge cycles increases, the discharge time in the CC stage gradually shortens. Therefore, the equal voltage drop discharge time with a starting voltage of 4V and an end voltage of 3.2V is selected as the health factor.

[0164] (2) Evaluation criteria for prediction models

[0165] The present invention uses the following indicators to evaluate the prediction performance of the proposed model.

[0166] 1) Mean Absolute Percentage Error (MAPE):

[0167]

[0168] 2) Root Mean Square Error (RMSE):

[0169]

[0170] In the formula, To predict SOH; i is the actual SOH.

[0171] Among them, the smaller the values ​​of MAPE and RMSE, the more accurate the prediction results.

[0172] (3) Experimental results and analysis

[0173] In order to verify the effectiveness and superiority of the IBKA-DHKELM model proposed in the present invention for predicting SOH of lithium-ion batteries, 40% of the data of the four groups of batteries were selected as training sets, and compared with the three prediction models of ELM, DHKELM and BKA-DHKELM for analysis. Figure 5a-5d This is a comparison chart of prediction effects under different algorithms. Figure 5a This is the comparison result of B5 battery. Figure 5b This is the comparison result of B18 battery. Figure 5c This is the comparison result of CS35 battery. Figure 5d The comparison results for CS36 batteries are shown in Table 1.

[0174] Table 1 Error comparison of different algorithms for four battery groups

[0175]

[0176]

[0177] As can be seen from Figure 5, the prediction effects of different models are significantly different. Figure 5aFor example, it can be seen from the figure that the ELM model has the worst effect, especially in the later stage of prediction, the error gradually increases, and the prediction curve is irregular, which proves that it does not fully extract the characteristics of the data. The DHKELM model has been significantly improved compared to the ELM model, reducing the error, but still cannot achieve the ideal prediction accuracy. The BKA-DHKELM model is optimized by the Black Kite Algorithm (BKA), and the prediction effect is further improved, but it still has certain limitations. The IBKA-DHKELM model of the present invention performs best. It optimizes the hyperparameters through the improved Black Kite Algorithm (IBKA), significantly improves the prediction accuracy, and almost completely coincides with the true value.

[0178] As can be seen from Table 1, the prediction errors of different models on four groups of batteries are significantly different. Compared with the ELM and DHKELM models, the BKA-DHKELM and IBKA-DHKELM models have significantly improved in both MAPE and RMSE indicators, especially the IBKA-DHKELM model, which shows higher prediction accuracy. The following analysis is based on two different batteries, B5 and CS35, as examples: In the prediction error of B5 battery, the RMSE of the ELM model is 0.0235, and the prediction error is large. The RMSE of the DHKELM model is 0.0125, which is 46.8% higher than the ELM model. The RMSE of the BKA-DHKELM model is 0.0042, which further improves the prediction accuracy by 66.4%. The RMSE of the IBKA-DHKELM model is only 0.0015, which is 64.3%, 88.0% and 93.6% higher than the BKA-DHKELM, DHKELM and ELM models, respectively. In the prediction error of CS35 battery, the RMSE of the ELM model is 0.0279, which is a large error. The RMSE of the DHKELM model is 0.0170, and the prediction accuracy is improved by 39.1%. The RMSE of the BKA-DHKELM model is 0.0101, and the prediction accuracy is improved by 40.6% compared with DHKELM. The RMSE of the IBKA-DHKELM model is 0.0062, and the prediction accuracy is improved by 38.6%, 63.5% and 77.8% compared with the BKA-DHKELM, DHKELM and ELM models, respectively.

[0179] Comprehensive analysis shows that the MAPE values ​​of the IBKA-DHKELM model are all smaller than those of other models. Compared with the ELM, DHKELM and BKA-DHKELM models, the prediction accuracy of the IBKA-DHKELM on all battery packs has been significantly improved, showing extremely high robustness and accuracy. This shows that the IBKA-DHKELM model has a significant advantage in the SOH prediction of lithium batteries by optimizing the model hyperparameters through the improved Black Kite algorithm.

[0180] The core algorithm and model of this invention are not limited to the SOH prediction of lithium-ion batteries, but can also be extended to other similar products or systems, such as for state estimation of other types of batteries (such as lead-acid batteries, nickel-metal hydride batteries, etc.), and can also be used for battery management in electric vehicles, grid energy storage systems and renewable energy systems.

[0181] In a second aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to any of the above-mentioned embodiments of the present invention is implemented.

[0182] The memory is a computer-readable storage medium that can be used to store software programs, computer executable programs and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory.

[0183] The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0184] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a DC power supply health state prediction model optimization method based on an extreme learning machine according to any embodiment of the present invention. The storage medium may be a ROM / RAM, a magnetic disk, an optical disk, or the like.

[0185] In a fourth aspect, the present invention provides a computer program product. When the computer program product runs on a computer, it enables the computer to execute the DC power supply health status prediction model optimization method based on the extreme learning machine of any of the above-mentioned embodiments of the present invention.

[0186] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A DC power supply health status prediction model optimization method based on extreme learning machine, characterized in that: include: Extracting capacity characteristics and aging characteristics of the DC power supply, and quantifying the capacity characteristics into power supply health status indicator values; Constructing a prediction model based on an extreme learning machine, the prediction model comprising a multi-layer extreme learning machine autoencoder, a kernel function and an output layer; wherein the kernel function combines a local kernel function and a global kernel function; The multi-layer extreme learning machine autoencoder is used to extract the aging features layer by layer to obtain hidden features; The kernel function is used to extract the hidden features to obtain high-dimensional prediction features; The output layer is used to predict and output the high-dimensional prediction features; The prediction model is trained using the aging characteristics and the power health status indicator value.

2. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to claim 1, characterized in that: The global kernel function adopts one of a linear kernel function, a polynomial kernel function, and a Sigmoid kernel, and the local kernel function adopts a Gaussian kernel function.

3. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to claim 2, characterized in that: The kernel function is a linear combination of a Gaussian kernel function and a polynomial kernel function.

4. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to claim 1, characterized in that: The method also includes: optimizing model hyperparameters using a genetic algorithm, a particle swarm optimization algorithm, a differential evolution algorithm or a black kite algorithm.

5. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to claim 4, characterized in that: The method of optimizing model hyperparameters by using the Black Kite algorithm includes: S1, establish the fitness function according to the mean square error between the model prediction value and the actual value, set the maximum number of iterations and the dimension of the optimization parameter; S2, evenly distribute the position of each black-winged kite to obtain the initial position, and select the best individual as the leader in the initial population according to the fitness function; S3, updating the position of each black-winged kite in the next iteration according to the black-winged kite attack behavior model; S4, updating leaders based on the black-winged kite migration behavior model; S5, repeat iterative steps S3 to S4 until the maximum number of iterations is reached, output the optimal position of the population, and obtain the optimal hyperparameters.

6. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to claim 5, characterized in that: The black-winged kite attack behavior model is: n = a*exp(b*exp(ct)); in, represents the position of the i-th black-winged kite in the j-th dimension and the t-th iteration step, represents the position of the i-th black-winged kite in the t+1-th iteration step of the j-th dimension, r is a random number between 0 and 1, p is a constant of 0.9; T is the total number of iterations, t is the number of iterations completed so far, a is the upper asymptote, which controls the initial value of the step size, b is the control parameter, which controls the displacement of t; c is the scaling factor, which controls the growth rate.

7. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to claim 5, characterized in that: The black-winged kite migration behavior model is: Among them, L t j represents the leading scorer of the j-th black-winged kite for the t-th iteration completed so far, F i represents the j-th dimension current position of any black-winged kite in the t-th iteration, F ri represents the fitness value of any black kite at a random position in the jth dimension in the tth iteration, C(0,1) represents the Cauchy mutation, m is the displacement factor, and t max is the maximum number of iterations, μ and σ are random numbers between 0 and 1.

8. The method for optimizing a DC power supply health status prediction model based on an extreme learning machine according to any one of claims 5 to 7, characterized in that: S2 includes: using a good point set initialization algorithm to generate a good point set of initial hyperparameters, mapping the good point set to a black kite population search space, and obtaining an initial position of the population.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the DC power supply health status prediction model optimization method based on the extreme learning machine described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for optimizing a DC power supply health status prediction model based on an extreme learning machine as described in any one of claims 1 to 8 is implemented.

Images